Documentation · API reference

Models & catalogue

Read-only endpoints describe everything the marketplace serves: the models, the endpoints behind each model, the providers, the families, the published benchmarks, the per-endpoint measurement series, the live status board, and public usage statistics. Only GET /models/user requires a key, to read the presented account's own preferences; every other route here serves without one, and unauthenticated requests are rate-limited per client address (see Errors & rate limits). Every listed price is the provider's declared price in USD per stated unit, as an exact decimal string: tokens, characters, images, audio seconds, video seconds and megapixels are different meters. Every measured figure — uptime, the two timings, throughput, tool-call success — is measured by Heimdallm from its own traffic and is null until its window holds enough samples; a value a provider did not declare is absent rather than guessed. The timings are reported separately and never merged: ttft_* is time to first token, which exists only for streamed requests, and response_* is the whole request through to the last byte. Throughput is reported over the last day alone, and only over streamed generations long enough to have an output rate — a buffered response has no generation phase at all, and a two-token answer measures scheduler jitter rather than the model.

GET /models

GET https://api-staging.heimdallm.com/api/v1/models

Every model with at least one live endpoint, newest first, narrowed by every filter present. Without limit the whole filtered list comes back in one response. Each model's summary uses one live endpoint, selected by the lowest prompt-token price a request is charged, then completion-token price, then endpoint id; unknown token prices sort last. Charged, not listed: an endpoint declaring a discount is compared at the price a customer pays, so the summary quotes the endpoint that request costs least on. This selects a representative endpoint for media, not the lowest price per image or video second. Filters do not change that summary selection. Compare the endpoints listing for prices in the unit you need. Query parameters:

Parameter	Type	Meaning
`family`	string	Only models of one family, by id.
`q`	string	Case-insensitive substring match against the model id or its display name.
`supported_parameters`	string	Comma-separated parameter names. A model matches when every named one is declared by at least one of its live endpoints — not necessarily the same one.
`input_modalities`, `output_modalities`	string	Comma-separated modalities the model must declare, all of them.
`context`	integer	Minimum context length in tokens, against the model's own declared window.
`min_price`, `max_price`	string	Prompt price bounds, USD per million tokens as an exact decimal string. At least one live endpoint's price must fall inside the bound given.
`min_output_price`, `max_output_price`	string	The same bounds on completion price.
`providers`	string	Comma-separated provider slugs. At least one live endpoint must belong to one of them.
`model_authors`	string	Comma-separated authors, matched against the model's own declared author.
`zdr`	boolean	`true` keeps only models with at least one endpoint the ZDR fold keeps. `false` filters nothing.
`region`	string	Comma-separated country codes or datacenter regions. At least one live endpoint must declare a matching datacenter.
`min_age_days`, `max_age_days`	integer	Age bounds in days since the model's declared release date.
`min_tool_success_rate`, `max_tool_success_rate`	number	Bounds, 0-100, on measured `tool_call_success_1d`. At least one live endpoint must clear its own sample gate and meet the bound; an endpoint whose figure is `null` matches neither one.
`sort`	string	`newest` (default), `oldest`, `price-low-to-high`, `price-high-to-low` or `context-high-to-low`.
`limit`, `offset`	integer	Pagination, applied after every filter above. `/models/count` is always the unfiltered count.
`category`, `arch`, `distillable`, `min_intelligence_index`, `max_intelligence_index`, `min_coding_index`, `max_coding_index`, `min_agentic_index`, `max_agentic_index`	string / number	Accepted, never rejected — but this catalogue computes none of them. Naming one answers `"data": []` with `unsupported_filters` listing which ones the request sent, whatever every other filter present would otherwise have matched.

Every filter combines with AND semantics — a model must satisfy all of them to appear. A filter this catalogue does not compute is never dropped from that AND: it cannot be evaluated, so it cannot be satisfied, and the whole request answers empty rather than narrowed by the filters that could run. GET /models?context=8192&category=roleplay answers { "data": [], "unsupported_filters": ["category"] }, not the models context alone would have kept. unsupported_filters is a top-level array beside data, present only when the request named one of the filters above this catalogue does not compute.

{
  "data": [
    {
      "id": "deepseek/deepseek-v4-flash-0731-abliterated",
      "canonical_slug": "deepseek/deepseek-v4-flash-0731-abliterated",
      "hugging_face_id": "cebeuq/DeepSeek-V4-Flash-0731-abliterated",
      "name": "DeepSeek: DeepSeek V4 Flash 0731 (abliterated)",
      "created": 1785611388,
      "expiration_date": null,
      "description": "An abliterated rebuild of DeepSeek V4 Flash 0731...",
      "context_length": 131072,
      "architecture": {
        "modality": "text->text",
        "input_modalities": ["text"],
        "output_modalities": ["text"],
        "tokenizer": "DeepSeek",
        "instruct_type": null
      },
      "hugging_face": {
        "id": "cebeuq/DeepSeek-V4-Flash-0731-abliterated",
        "read_at": 1789030800,
        "availability": "public",
        "languages": ["en", "zh"],
        "weights_revision": "0e9e39f249a16976918f6564b8830bc894c89659",
        "weights_updated_at": 1785697788,
        "downloads": 128473,
        "likes": 941
      },
      "pricing": {
        "prompt": "0.0000001",
        "completion": "0.0000004",
        "input_cache_read": "0.000000037",
        "overrides": [{ "min_prompt_tokens": 200000, "prompt": "0.0000002" }],
        "discount": 0
      },
      "top_provider": {
        "context_length": 131072,
        "max_completion_tokens": 32768,
        "is_moderated": false
      },
      "per_request_limits": null,
      "supported_parameters": ["max_tokens", "response_format", "seed", "stop", "temperature", "tools"],
      "default_parameters": { "temperature": 1, "top_p": 1 },
      "family": {
        "id": "deepseek/deepseek-v4-flash-0731",
        "name": "DeepSeek: DeepSeek V4 Flash 0731"
      },
      "author": "deepseek",
      "tags": ["abliterated", "moe"],
      "endpoint_count": 1,
      "links": {
        "details": "/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated/endpoints"
      },
      "supported_voices": null,
      "supported_dimensions": null,
      "supported_languages": null
    }
  ]
}

Field	Type	Notes
`id`	string	`author/slug` — what you pass as `model` in a request.
`canonical_slug`	string	Equal to `id`.
`hugging_face_id`	string \| null	The weights repository on Hugging Face, exact casing; `null` when the weights are not published.
`hugging_face`	object	What the repository `hugging_face_id` names states about the weights, and the moment that statement was read: `id` and `read_at`, plus `availability`, `languages`, `papers`, `datasets`, `weights_revision`, `weights_updated_at`, `downloads`, `likes` and `evaluations` where the repository states them. Present only where a reading of the repository exists. (Heimdallm)
`name`	string	Display name, conventionally `Author: Model Name`.
`created`	integer	Release date, Unix seconds. The list is sorted on it, newest first.
`expiration_date`	string \| null	The date after which the model may be removed, `YYYY-MM-DD` in UTC, from the `deprecation_date` its provider document declares. `null` is a model with no announced end.
`description`	string \| null	Provider-written description.
`context_length`	integer	The model's context window in tokens.
`architecture`	object	`modality` (e.g. `text->text`, `text+image->text`), and `input_modalities` / `output_modalities` over the declared modality vocabulary. `tokenizer` is the tokenizer family the summary endpoint declares, `null` where it declares none. `instruct_type` is always `null`: no provider declaration carries a prompt-template name.
`pricing`	object	The summary endpoint's prices as a request is charged them, USD per unit as decimal strings, each key naming the unit it meters: `prompt` and `completion` per token, `input_cache_read` per cached prompt token, `input_cache_write` per cache-write token, `input_cache_write_1h` per token written to a cache that holds it for an hour where the endpoint prices that lifetime apart from its default one, `internal_reasoning` per reasoning token, `request` per request, `image` per image supplied in the prompt, `image_output` per output image, `image_output_token` per output image token, `image_megapixel` per megapixel of image output, `video_second` per second of video output, `mesh_output` per delivered mesh, `audio_second` per second of audio, `audio_output` per audio token generated in a reply, `character` per character synthesized. `overrides` beside them carries the prices that apply only under a stated condition — see conditional prices. A key is present only where the endpoint declared that price; an undeclared unit is unknown, never `0`. A rate can carry more than 24 decimals where a discount multiplied a declaration's own precision; a charge is that rate times the units consumed, floored once at the ledger's 24 decimals. A price per audio input token is never published, because the marketplace holds no price in that unit. `discount` beside the units is the fraction the quoted endpoint took off its own list to reach these figures, so a list price is `price / (1 - discount)`; it is always present, `0` where nothing is taken off and negative where the endpoint charges above its list. The endpoint quoted is the one a request costs least on, and its context window, tokenizer and controls above come from the same endpoint as these prices.
`top_provider`	object	The summary endpoint's `context_length` and `max_completion_tokens`; `is_moderated` is always `false`.
`per_request_limits`	null	Always `null`.
`supported_parameters`	string[]	Union of the request parameters declared by the model's live endpoints, sorted.
`reasoning`	object	Optional declared controls from the same endpoint as the model summary price. Every member is optional: `supported_efforts` lists an explicit domain strongest first; `mandatory: false` declares an off control; `default_effort` and `default_enabled` describe known defaults; `supports_max_tokens: true` declares a token budget. An absent domain is unknown, never a promise of every API effort. Switches and budgets can be present without effort choices. See reasoning controls.
`default_parameters`	object	The sampling the model's author recommends: `temperature`, `top_p`, `top_k`, `repetition_penalty`. Always present, `{}` where nothing is known, and a member nobody stated is absent rather than null.
`family`	object	`{ id, name }` of the model's family. (Heimdallm)
`author`	string	The creator of the weights — never the serving provider. (Heimdallm)
`tags`	string[]	Short provider-declared tags. (Heimdallm)
`endpoint_count`	integer	Live endpoints currently serving this model. (Heimdallm)
`links.details`	string	Path of the endpoints listing for this model.
`supported_voices`	string[] \| null	The voices a speech model loaded, as the union across its live endpoints. `null` says the model synthesizes no speech; `null` on one that does says no live endpoint of it has declared an envelope. What a single endpoint speaks is `speech_envelope` on the endpoints listing.
`supported_dimensions`	integer[] \| null	The truncated vector widths a `dimensions` request on this model may name, read off its cheapest embeddings endpoint. `null` says the model produces no embeddings; `null` on one that does says its cheapest endpoint declares no width. (Heimdallm)
`supported_languages`	string[] \| null	The languages a `language` request on this model may name, read off its cheapest transcription or speech endpoint. `null` says the model transcribes or speaks nothing; `null` on one that does says its cheapest endpoint declares no list. (Heimdallm)

default_parameters is what a request that leaves a member out is served with. The values are the model author's own recommendation, read off the generation_config.json of the repository hugging_face_id names, except where the provider declares the field themselves, which is published exactly as they declared it. A serving endpoint applies a listed default to a member your request omits — a member you send, whatever its value, is always what reaches the engine instead. A member missing from default_parameters is served with whatever the engine runs, since the marketplace has no recommendation to fill it with.

hugging_face is attribution rather than a marketplace figure. Every member inside it is what the weights repository stated about itself at the moment read_at names, so you can tell a repository's own claim from a figure this catalogue computes. The two absences say different things. The member itself is absent on a model naming no repository and on one whose repository has not been read, so its presence is the statement that a reading exists and read_at is when it was taken. Inside it, a fact the repository states no value for is absent rather than null or an empty array, the only meaning absence can carry within a single reading. A missing languages is a card that states none, never a card nobody has read.

Reasoning controls

Read the endpoint metadata when choosing a deployment; the model summary describes only its selected endpoint. supported_efforts is optional and never null. When present, it lists the explicitly declared values in descending order. Its absence does not imply universal support or prove the model cannot reason.

mandatory: false means the declaration includes none or a boolean thinking switch. Otherwise mandatory is omitted: an absent disable declaration does not prove thinking is mandatory. default_effort appears only for a valid declared default. default_enabled uses a declared boolean default, or whether the effort default differs from none when there is no independent thinking switch. A switch without a default does not inherit the effort default. Missing or conflicting defaults stay absent.

A thinking switch or token budget can exist without an effort dial. supports_max_tokens: true requires a declared integer or range budget. These fields describe declarations; they do not extend an adapter's request handling. The generic OpenAI-compatible adapter forwards effort, but not the nested enabled, budget or exclude fields. See chat forwarding.

Conditional prices

The native media scalar keys (image_output, image_output_token, image_megapixel, and video_second) appear only for one unambiguous, unconditional output tariff. A conditional or windowed tariff is omitted from those scalar keys. The image endpoints listing preserves the structured tariff lines and their conditions; the general projection below cannot express every media condition. An omitted scalar price is unknown, not free.

A pricing key is what a request carrying no stated condition pays. An endpoint that charges a different rate under a condition publishes that rate in overrides. Each entry carries a price under the same key name the object beside it uses. All conditions on an entry must match; among matching entries, the last price for each key wins. Time-window rates follow default-rate overrides, so an active window takes precedence even when the provider lists its default rate last. A prompt-length override within a window carries both conditions:

"overrides": [
  // A long prompt, charged at a higher prompt rate above the threshold.
  { "min_prompt_tokens": 200000, "prompt": "0.0000002" },
  // A daytime rate on weekdays, as an HHMM UTC window that may wrap midnight.
  { "utc_start": 800, "utc_end": 1630, "utc_days": ["monday", "tuesday"], "completion": "0.0000008" }
]

min_prompt_tokens applies where the prompt is strictly longer than the figure. Only a lower bound on prompt length and a UTC clock window are published: an endpoint may declare a condition on another quantity, and a price shown without the condition it depends on would read as one every request pays, so such a price is left out of overrides entirely rather than published bare. Cache-write prices for unasked writes or lifetimes other than the default five minutes and the separately named hour are also omitted from this projection because its keys cannot identify those qualifiers.

The prompt length is read twice, and a request is charged the tier it meets both times. The credit hold reserves against the prompt estimate, measured over everything the endpoint will tokenize with tools included, so a request over the threshold is held at the tier's rate rather than at the rate beside it. The settlement reads the prompt count the endpoint reported with its usage, which is the count the charge is taken over, so what is billed is the tier the measured prompt met rather than the one the estimate predicted. A reply that reported no prompt count is charged the entry's own price and its prompt is recorded unmeasured: a threshold nothing measured is unknown, not met. Prompt length is the one condition evaluated on the token rails; the image and video rails resolve their rendering settings when a render is quoted, before it is drawn.

One example per family of filters — every predicate inside one call combines with AND:

# Text search
GET /models?q=deepseek

# Capability, price and privacy together — an endpoint must satisfy all three
GET /models?supported_parameters=tools&max_price=1&zdr=true

# Modality and context floor
GET /models?input_modalities=text,image&context=32768

# Provider, author and datacenter region
GET /models?providers=impcc&model_authors=deepseek&region=us

# Age and measured reliability
GET /models?min_age_days=30&min_tool_success_rate=90

# Ordering and pagination
GET /models?sort=price-low-to-high&limit=20&offset=20

# A filter this catalogue does not compute — accepted, answers empty, and says so
GET /models?category=roleplay
# { "data": [], "unsupported_filters": ["category"] }

GET /models/count

Returns { "data": { "count": 2 } }, where count is the length of the complete, unfiltered GET /models list across all modalities. Pagination does not change this total; a catalogue without served models returns zero.

GET /models/user

GET https://api-staging.heimdallm.com/api/v1/models/user

Requires an API key. The complete unfiltered GET /models list, narrowed by the presented key's own account preferences. Today that is exactly one stored preference: the workspace's default zdr routing preference. When it is true, every model with no live endpoint the ZDR fold keeps is omitted. A key whose workspace has not set this preference, or has set it false, sees the same list GET /models serves. Returns 401 without a key.

{ "data": [/* ModelSummary, filtered */] }

GET /models/:author/:slug

GET https://api-staging.heimdallm.com/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated

Returns { "data": model }, where model is exactly the same object as the matching entry in GET /models, including its pricing, supported parameters and family. The model must be active and have a live endpoint belonging to an approved provider. An unknown model or one without a served endpoint returns 404 with the standard error envelope. The singular /model/:author/:slug spelling serves the same object and applies the same visibility rules.

SDK retrieval

GET https://api-staging.heimdallm.com/api/v1/models/deepseek%2Fdeepseek-v4-flash-0731-abliterated

The OpenAI Node and Python SDKs interpolate a model id into one path segment rather than two: models.retrieve("deepseek/deepseek-v4-flash-0731-abliterated") requests /models/deepseek%2Fdeepseek-v4-flash-0731-abliterated, one percent-encoded segment. This path resolves through the same catalogue projection and the same visibility rules as GET /models/:author/:slug — same 404 for an unknown or unserved id — but answers the bare object those SDKs read directly off the result, not a data envelope:

{
  "id": "deepseek/deepseek-v4-flash-0731-abliterated",
  "canonical_slug": "deepseek/deepseek-v4-flash-0731-abliterated",
  "object": "model",
  "created": 1785611388,
  "owned_by": "deepseek",
  "hugging_face_id": "deepseek-ai/DeepSeek-V4-Flash-0731",
  "name": "DeepSeek: DeepSeek V4 Flash 0731 (abliterated)",
  "description": "A mixture-of-experts model tuned for speed.",
  "context_length": 131072,
  "architecture": {
    "modality": "text->text",
    "input_modalities": ["text"],
    "output_modalities": ["text"],
    "tokenizer": null,
    "instruct_type": null
  },
  "hugging_face": {
    "id": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "read_at": 1789030800,
    "availability": "public",
    "languages": ["en", "zh"],
    "papers": ["2606.19348"],
    "datasets": ["deepseek-ai/DeepSeek-ProverBench"],
    "weights_revision": "0e9e39f249a16976918f6564b8830bc894c89659",
    "weights_updated_at": 1785697788
  },
  "pricing": { "prompt": "0.0000001", "completion": "0.0000004", "discount": 0 },
  "top_provider": { "context_length": 131072, "max_completion_tokens": 32768, "is_moderated": false },
  "per_request_limits": null,
  "supported_parameters": ["temperature", "top_p", "tools"],
  "default_parameters": { "temperature": 1, "top_p": 1 },
  "family": { "id": "deepseek/deepseek-v4-flash-0731", "name": "DeepSeek: DeepSeek V4 Flash 0731" },
  "author": "deepseek",
  "tags": ["abliterated", "moe"],
  "endpoint_count": 1,
  "links": { "details": "/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated/endpoints" }
}

object is always the literal "model" and owned_by equals author; every other field is the same as the matching entry in GET /models. GET /models/:author/:slug keeps answering { "data": model } unchanged — this path exists beside it for a client that cannot spell two path segments.

GET /models/:author/:slug/endpoints

GET https://api-staging.heimdallm.com/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated/endpoints

One model with every live endpoint serving it, ordered by provider slug. family is the { id, name } pair the GET /models summary carries. benchmarks_note is the grading disclosure every score in an endpoint's benchmarks is read under. 404 when the id names no active model.

{
  "data": {
    "id": "deepseek/deepseek-v4-flash-0731-abliterated",
    "name": "DeepSeek: DeepSeek V4 Flash 0731 (abliterated)",
    "created": 1786677579,
    "description": "...",
    "architecture": {
      "modality": "text->text",
      "input_modalities": ["text"],
      "output_modalities": ["text"],
      "tokenizer": null,
      "instruct_type": null
    },
    "family": {
      "id": "deepseek/deepseek-v4-flash-0731",
      "name": "DeepSeek: DeepSeek V4 Flash 0731"
    },
    "endpoints": [
      {
        "name": "ImpCC | deepseek/deepseek-v4-flash-0731-abliterated",
        "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
        "model_name": "DeepSeek: DeepSeek V4 Flash 0731 (abliterated)",
        "provider_name": "ImpCC",
        "tag": "impcc/fp4",
        "provider_slug": "impcc",
        "provider_id": "0b6f3c1e-8d2a-4f7b-9c1d-5e4a3b2c1d0e",
        "context_length": 131072,
        "max_completion_tokens": 32768,
        "max_prompt_tokens": null,
        "quantization": "fp4",
        "supported_parameters": ["max_tokens", "seed", "stop", "temperature", "tools", "reasoning_effort"],
        "refuted_reasoning_efforts": [],
        "pricing": {
          "prompt": "0.0000001",
          "completion": "0.0000004",
          "input_cache_write": "0.000000125",
          "input_cache_write_1h": "0.0000002",
          "discount": 0.1
        },
        "status": 0,
        "uptime_last_1d": 99.8,
        "uptime_last_7d": 99.5,
        "ttft_last_30m": { "p50": 0.42, "p75": 0.55, "p90": 0.81, "p99": 1.9 },
        "ttft_last_1d": { "p50": 0.45, "p75": 0.58, "p90": 0.86, "p99": 2.1 },
        "response_last_30m": { "p50": 2.61, "p75": 3.4, "p90": 4.9, "p99": 8.1 },
        "response_last_1d": { "p50": 2.74, "p75": 3.6, "p90": 5.2, "p99": 8.6 },
        "throughput_last_1d": { "p50": 85.1, "p75": 92.4, "p90": 101.7, "p99": 118.2 },
        "throughput_last_1d_samples": 214,
        "render_timing": null,
        "render_timing_samples": 0,
        "render_timing_sizes": 0,
        "supports_implicit_caching": false,
        "health": "healthy",
        "compliance": { "zdr": true, "zdr_status": "approved", "hipaa": null },
        "data_policy": { "training": false, "retention_days": 0 },
        "datacenters": [{ "country_code": "CA" }],
        "tool_call_success_1d": 98.2,
        "tool_call_success_1d_samples": 340,
        "runaway_rate_1d": 1.4,
        "runaway_rate_1d_samples": 620,
        "benchmarks": [{ "dataset": "gsm8k", "score": 0.71, "sample_count": 400 }],
        "capacity": { "current": 1, "total": 4, "updated_at": "2026-08-16T12:00:15Z" }
      }
    ],
    "benchmarks_note": "Scores are sampled and mechanically graded..."
  }
}

Per endpoint:

Field	Type	Notes
`name`	string	`Provider name \| model id`.
`model_id`, `model_name`	string	The model this endpoint serves.
`provider_name`	string	Display name; also what the inference response's `provider` field carries.
`provider_slug`	string	The identifier used in provider preferences. (Heimdallm)
`provider_id`	string	The provider's primary key. It holds when the provider's `slug` is changed, so a client that stores endpoints keyed by provider follows a renamed provider rather than seeing a new one. (Heimdallm)
`tag`	string	`slug` or `slug/quantization` when a quantization is declared.
`context_length`	integer	This endpoint's context window.
`max_completion_tokens`	integer \| null	Output cap, when declared.
`max_prompt_tokens`	integer \| null	Prompt cap declared on the endpoint's `text` input, when stated in tokens; `null` when not declared.
`quantization`	string \| null	`int2` … `fp32`, or `null` when not declared.
`supported_parameters`	string[]	Request parameters this endpoint declares support for.
`input_modalities`	string[]	This endpoint's inputs, with current semantic qualification observations applied.
`declared_input_modalities`	string[]	Inputs the provider's document lists, before qualification.
`capability_qualification`	object[]	Current semantic observations per capability, each carrying its latest status and an unexpired decisive verdict. An entry is absent where a capability is untested, where its evidence does not apply, and where a check refuted something the endpoint's own document never declared: a refutation of an undeclared capability withholds nothing and is published as nothing. See reasoning routing for how a refuted reasoning level reads through it.
`reference_inputs`	object	The image, video and speech reference inputs this endpoint accepts, each with its formats, roles and item limit.
`render_parameters`	object	Full descriptors for every declared image/video render parameter this endpoint states a domain for, keyed by modality then name.
`image_envelope`	object \| null	The image sizes and formats this endpoint's `image_generation` route accepts; matched against `size` on a request. `null` on an endpoint that generates no images.
`video_envelope`	object \| null	This endpoint's declared video render limits — `frame`, `max_seconds`, `frame_step`, `audio` among them. `null` on an endpoint that generates no video.
`mesh_envelope`	object \| null	This endpoint's declared face range — `max_faces`, `min_faces` and `texture`. A floor as well as a ceiling, because a decimator handed a target below the surface it can close lands above the target in silence. `null` on an endpoint that generates no mesh.
`speech_envelope`	object \| null	What this endpoint speaks: its voices, codecs and speed range, and the only voices a speech request routed to it may name. `null` on an endpoint that synthesizes no speech.
`reasoning`	object	This endpoint's optional declared reasoning controls, with the same fields and absence semantics as the model summary.
`video_generation`	object	This endpoint's optional declared video controls, with the same fields and absence semantics as the model summary.
`refuted_reasoning_efforts`	string[]	Currently declared levels of `reasoning_effort` whose latest per-level check refuted the value; empty when none is. The declaration in `supported_parameters` stands — a refuted level filters the value, not the name — so a request naming a listed level routes past this endpoint while the failed observation matches its serving revision and probe version and is less than 24 hours old. See reasoning routing. (Heimdallm)
`pricing`	object	This endpoint's own prices as a request is charged them, the same keys as on `GET /models`, plus `discount`: the fraction taken off its list to reach them, so a list price is `price / (1 - discount)`. Always present, `0` where nothing is taken off and negative where the endpoint charges above its list, so no reader has to tell a declared zero from a figure that failed to arrive. The endpoints array is ordered by this charged price, cheapest first.
`status`	integer	`0` when `health` is `healthy`, `-1` otherwise.
`uptime_last_5m`, `uptime_last_30m`	number \| null	Percent of counted attempts that succeeded in each window; `null` below 10 success/error attempts. Other outcomes do not count.
`uptime_last_5m_samples`, `uptime_last_30m_samples`, `uptime_last_1d_samples`	integer \| null	Counted attempts in the named window. `0` is a scanned empty window; `null` means it has not been computed.
`latency_last_30m`	object \| null	Alias of `ttft_last_30m`: p50/p75/p90/p99 time to first token in seconds, not whole-response latency.
`latency_last_30m_samples`	integer \| null	Streams carrying a first-token observation in the last 30 minutes; the figure is `null` below 5.
`throughput_last_30m`	object \| null	Recent completion tokens per second under the daily observation rules and a separate five-stream gate. Percentiles are performance floors: p90 is met or exceeded by 90% of qualifying streams; p99 is lower, not higher. The daily field retains ascending statistical percentiles.
`throughput_last_30m_samples`	integer \| null	Qualifying streams in the last 30 minutes, including counts below the gate.
`perf_last_30m_by_workload`	null	Workload-specific performance is not measured.
`supports_tool_choice`	boolean	`tools` is declared and neither a standing capability refutation nor measured tool fidelity below the routing floor withholds it. An inconclusive check does not clear a standing refutation.
`supports_voice_cloning`	boolean	Whether the endpoint declares `input_references` on its `speech` output modality, which is a reference clip accepted for a voice. `false` on every endpoint whose document declares no reference clip.
`uptime_last_1d` / `_7d`	number \| null	Success rate (percent) over the window; `null` below 10 attempts in the window.
`ttft_last_30m` / `_1d`	object \| null	Time to first token in seconds: `p50`/`p75`/`p90`/`p99` over the window; `null` below 5 samples. Only a streamed request has a first token, so an endpoint that never streams — an embeddings one — reports `null`.
`response_last_30m` / `_1d`	object \| null	Whole request in seconds, start to last byte, same percentiles and gate. Every completed request contributes, streamed or not; this is the only timing a buffered endpoint can offer.
`throughput_last_1d`	object \| null	Completion tokens per second during generation over the last day, same percentiles and gate. A generation contributes only if it streamed — a buffered response has no generation phase to measure — and carried at least 16 completion tokens, since on a shorter answer that phase is one gap between tokens, which times the scheduler and not the model. The 30-minute window uses the same observation rules and its own gate; sparse recent traffic leaves that window null.
`throughput_last_1d_samples`	integer \| null	Streams behind `throughput_last_1d`, published whether or not the figure is. `0` means the day held no stream long enough to time; a number below 5 beside a `null` figure means the endpoint is still earning the figure; `null` means nothing has been measured for this endpoint at all. (Heimdallm)
`render_timing`	object \| null	What a render costs an endpoint in seconds: `fixed_seconds` before it draws anything, plus `seconds_per_megapixel` for the size asked for, so the wait for a size can be worked out before asking. A percentile could not answer that — it reads as whichever size the endpoint served most, and that is the health check's cheapest one. `null` for everything that renders no images, and until the endpoint has served at least 2 distinct sizes and 5 renders across them. (Heimdallm)
`render_timing_samples` / `_sizes`	integer \| null	Renders behind `render_timing`, and the distinct geometries among them. Two counts because the fit has two gates: `render_timing: null` beside 40 samples and 1 size says the line is waiting on a second geometry rather than on more traffic. (Heimdallm)
`supports_implicit_caching`	boolean	`true` when the endpoint declares a cache-read price.
`health`	string	Judged on the endpoint's newest 20 counted attempts, not on a clock window, so it recovers as fast as new evidence arrives: `healthy` (≥ 90% of them succeeded), `degraded` (≥ 60%), `down` (below 60%), `unknown` (fewer than 5). A run of consecutive probe failures short-circuits it downwards before the sample turns over — two in a row is `degraded`, four is `down` — so an outage shows within minutes. One isolated failure moves nothing: at this scale it is noise, and a request that fails is retried once before it is counted at all. (Heimdallm)
`compliance`	object \| null	`hipaa` is the provider's declaration, `null` when not declared — and so is the whole object when nothing is declared. `zdr` is the marketplace's verdict on the provider's zero-data-retention declaration, not the declaration itself: `true` once the claim passed review, `null` while it is under review (or nothing is declared), `false` when review denied it or the provider declared false. Wherever a claim stands, `zdr_status` names the verdict: `pending`, `approved` or `denied`. See Data & privacy. (Heimdallm)
`data_policy`	object \| null	Declared `training` (prompts may train models) and `retention_days`. (Heimdallm)
`datacenters`	object[] \| null	Where inference runs: `[{ "country_code": "CA" }]`. (Heimdallm)
`tool_call_success_1d`	number \| null	Percent of the last day's counted tool requests that produced a tool call; `null` below 20 of them, meaning too few to measure rather than none answered. Requests the caller abandoned, requests whose stream broke, and requests whose outcome was never recorded are not counted either way. Measured over a whole day because the figure swings hour to hour, and an endpoint measuring under 85% stops receiving tool traffic until it measures above again. (Heimdallm)
`tool_call_success_1d_samples`	integer \| null	Demanded tool asks behind `tool_call_success_1d`: asks made under a required or named tool choice, which are the only ones the endpoint owns the outcome of. `0` means no such ask reached it in the day. (Heimdallm)
`runaway_rate_1d`	number \| null	Percent of the last day's long replies that ended still repeating their own tail. The one percentage here that reads worse the higher it is, uptime and `tool_call_success_1d` both reading better the higher they are. The denominator is the replies whose text was read — no embedding, transcription or render is in it — carrying at least 256 completion tokens, which is several times the roughly 64 tokens of repeated tail the reading needs before it says anything — below that, whether a reply looped is a fact about how long the answer happened to be rather than about the endpoint; `null` below 50 of them, meaning too few to measure rather than none looping. A reply that stopped of its own accord and one that ended by calling a tool are outside the numerator, and so is one the engine's own repetition detector ended, which stays in the denominator. The reading is taken as the reply passes, over a bounded window of its tail, and no reply text is retained. (Heimdallm)
`runaway_rate_1d_samples`	integer \| null	Long replies behind `runaway_rate_1d`, published whether or not the figure is: replies whose text was read, carrying at least 256 completion tokens. `0` means the day held no reply long enough to judge; a number below 50 beside a `null` figure means the endpoint is still earning the figure; `null` means nothing has been measured for this endpoint at all. (Heimdallm)
`benchmarks`	object[]	This endpoint's published scores: `{ dataset, score, sample_count }`. Earned per endpoint, not per model — the same weights at a different quantization score differently. Also available alone at `GET /models/:author/:slug/benchmarks`. (Heimdallm)
`capacity`	object \| null	The provider's own live report: `total` concurrent request slots across its replicas, `current` in use, `updated_at` when it was heard. `null` when it never reported or the last report is over a minute old. (Heimdallm)

Capability evidence applies to its tested serving configuration and probe version for at most 24 hours. An inconclusive check cannot extend an earlier failure's expiry. HTTP refusals and aborted responses remain diagnostic observations: they do not prove that the endpoint lacks the feature. A request can be rejected because of marketplace translation or an unsuitable test.

Chat image-reading checks apply only to chat endpoints. They do not assess image references for video or image generation, embeddings or reranking. An absent qualification result means unverified, not unsupported; listed image inputs remain available unless an applicable check refutes them.

status, health and the uptime figures are what routing orders endpoints by, and what the status page renders — there is no separate status API. Reported capacity orders them too: endpoints with free slots are tried before endpoints that report nothing, and saturated endpoints last.

GET /endpoints/zdr

GET https://api-staging.heimdallm.com/api/v1/endpoints/zdr

Every live endpoint the ZDR fold keeps, across the whole catalogue, in the same object GET /models/:author/:slug/endpoints publishes per endpoint. An endpoint appears here only for a ZDR claim whose provider passed marketplace review — a claim on a provider still under review or one denied review is not here, exactly as provider.zdr and GET /models?zdr=true read it. No key required.

{ "data": [/* one endpoint object per entry, as above */] }

GET /providers

GET https://api-staging.heimdallm.com/api/v1/providers

Approved providers, ordered by name.

{
  "data": [
    {
      "id": "0b6f3c1e-8d2a-4f7b-9c1d-5e4a3b2c1d0e",
      "name": "ImpCC",
      "slug": "impcc",
      "privacy_policy_url": "https://impcc.example/privacy",
      "terms_of_service_url": null,
      "status_page_url": null,
      "headquarters": "CA",
      "datacenters": ["CA"],
      "website": "https://impcc.example",
      "model_count": 3
    }
  ]
}

Field	Type	Notes
`id`	string	The provider's primary key. It holds when `slug` is changed.
`name`, `slug`	string	Display name and the identifier routing preferences use.
`privacy_policy_url`, `terms_of_service_url`, `status_page_url`, `website`	string \| null	As declared by the provider.
`headquarters`	string \| null	Two-letter country code.
`datacenters`	string[] \| null	Country codes across the provider's live endpoints.
`model_count`	integer	Distinct models with at least one live endpoint at this provider.

GET /families

GET https://api-staging.heimdallm.com/api/v1/families

Model families with at least one live endpoint. (Heimdallm extension — OpenAI-style clients can ignore it.)

{
  "data": [
    {
      "id": "deepseek/deepseek-v4-flash-0731",
      "name": "DeepSeek: DeepSeek V4 Flash 0731",
      "author": "deepseek",
      "hugging_face_id": "deepseek-ai/DeepSeek-V4-Flash-0731",
      "models": ["deepseek/deepseek-v4-flash-0731-abliterated"],
      "endpoint_count": 1,
      "providers": ["impcc"]
    }
  ]
}

Field	Type	Notes
`id`, `name`	string	The family — the base model the variants group under.
`author`	string	Creator of the base weights.
`hugging_face_id`	string \| null	The base model's weights repository, when published.
`models`	string[]	Served model ids in the family, sorted. Pass one as `model`.
`endpoint_count`	integer	Live endpoints across the family.
`providers`	string[]	Slugs of the providers serving any model in the family.

GET /families/:author/:slug

GET https://api-staging.heimdallm.com/api/v1/families/deepseek/deepseek-v4-flash-0731

One family with the full summary of every served model in it, the shape GET /models returns per model, and every live endpoint across those models. Each endpoint row is exactly the row the endpoints listing publishes for that endpoint, and it names its own model_id. Rows are ordered by model_id, then provider_slug, then endpoint id. There is no family-level figure: uptime, timings, prices and scores stay per endpoint. benchmarks_note is the grading disclosure every score in a row's benchmarks is read under. A family none of whose models is served answers with empty models and endpoints arrays rather than a 404, so a family link stays valid between listings.

{
  "family": {
    "id": "deepseek/deepseek-v4-flash-0731",
    "name": "DeepSeek: DeepSeek V4 Flash 0731",
    "author": "deepseek",
    "hugging_face_id": "deepseek-ai/DeepSeek-V4-Flash-0731"
  },
  "models": [],
  "endpoints": [],
  "benchmarks_note": "Scores are sampled and mechanically graded..."
}

GET /models/:author/:slug/benchmarks

GET https://api-staging.heimdallm.com/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated/benchmarks

The published benchmark scores of one model's endpoints, with the grading disclosure. Scores attach to endpoints, not models — the same weights at a different quantization, or served with different settings, answer the same exam differently — and grading is mechanical, never by another model. The response is cached server-side for about a minute.

{
  "data": [
    {
      "dataset": "gsm8k",
      "score": 0.71,
      "sample_count": 400,
      "grading": "graded mechanically from sampled generations"
    }
  ],
  "note": "Scores are sampled and mechanically graded..."
}

Field	Type	Notes
`dataset`	string	The exam the score belongs to.
`score`	number	Correct fraction, 0 to 1.
`sample_count`	integer	How many items earned it.
`grading`	string	How the score was produced — the disclosure line, human-readable.

(The per-endpoint benchmarks array on the endpoints listing carries the same scores without repeating the note.)

GET /models/:author/:slug/series

GET https://api-staging.heimdallm.com/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated/series?days=7

Per-endpoint hourly measurement series of one model over the last 7 days — the data behind the model page's performance and price-history charts. days is 1 to 7, default 7. Hours are UTC and sparse: an absent hour recorded nothing, so a quiet endpoint is honest about it. Each row names its model_id, the one row shape the family series shares.

{
  "data": {
    "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
    "window_days": 7,
    "endpoints": [
      {
        "endpoint_id": "9e8d7c6b-…",
        "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
        "provider_slug": "impcc",
        "provider_id": "0b6f3c1e-8d2a-4f7b-9c1d-5e4a3b2c1d0e",
        "provider_name": "ImpCC",
        "quantization": "fp4",
        "hours": [
          {
            "hour": "2026-08-16T12:00:00.000Z",
            "ok": 412,
            "error": 3,
            "ttft_p50": 0.41,
            "ttft_p90": 0.79,
            "throughput_median": 84.2
          }
        ],
        "prices": [
          {
            "captured_at": "2026-08-10T00:00:00.000Z",
            "prompt": "0.0000001",
            "completion": "0.0000004",
            "cached_prompt": "0.000000037",
            "request": null,
            "image": null,
            "audio": null,
            "character": null,
            "image_output": null,
            "image_output_token": null,
            "video_second": null
          }
        ]
      }
    ]
  }
}

Per endpoint: hours (ascending) with request outcomes (ok, error) and, where samples earned them, time-to-first-token percentiles in seconds and median completion tokens per second — both null in an hour that did not stream enough to measure. prices (ascending) records changes to the supported scalar prices a request is charged, as exact decimal strings, one member per unit the change log keeps; a document that changes nothing but its discount moves the series. Native media rates use image_output, image_output_token, or video_second; image retains the megapixel rate. Conditional, windowed, or ambiguous native tariffs have no scalar history value. A price reading null on a point was unknown at that moment, not zero; older points are not converted between units.

GET /models/:author/:slug/activity/series

GET https://api-staging.heimdallm.com/api/v1/models/deepseek/deepseek-v4-flash-0731-abliterated/activity/series?days=30

One model's billed usage per UTC day: requests, prompt and completion token sums, image requests with the megapixels they drew, and audio seconds. days is 1 to 30, default 30; only days with traffic appear. Every billed request counts, the platform's own paid probe traffic included, like any other.

{
  "data": {
    "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
    "window_days": 30,
    "daily": [
      {
        "date": "2026-08-16",
        "requests": 5120,
        "prompt_tokens": 161052110,
        "completion_tokens": 40881205,
        "images": 0,
        "megapixels": 0,
        "audio_seconds": 0
      }
    ]
  }
}

GET /families/:author/:slug/series

GET https://api-staging.heimdallm.com/api/v1/families/deepseek/deepseek-v4-flash-0731/series?days=7

Per-endpoint hourly measurement series across every served model in a family over the last 7 days, the data behind the family page's performance and price-history charts. Each row is the row the model series publishes for that endpoint and names its model_id. Rows are ordered by provider_slug then endpoint id, and no bucket is merged across endpoints or models. days is 1 to 7, default 7. Hours are UTC and sparse.

{
  "data": {
    "family_id": "deepseek/deepseek-v4-flash-0731",
    "window_days": 7,
    "endpoints": [
      {
        "endpoint_id": "9e8d7c6b-…",
        "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
        "provider_slug": "impcc",
        "provider_id": "0b6f3c1e-8d2a-4f7b-9c1d-5e4a3b2c1d0e",
        "provider_name": "ImpCC",
        "quantization": "fp4",
        "hours": [
          {
            "hour": "2026-08-16T12:00:00.000Z",
            "ok": 412,
            "error": 3,
            "ttft_p50": 0.41,
            "ttft_p90": 0.79,
            "throughput_median": 84.2
          }
        ],
        "prices": [
          {
            "captured_at": "2026-08-10T00:00:00.000Z",
            "prompt": "0.0000001",
            "completion": "0.0000004",
            "cached_prompt": "0.000000037",
            "request": null,
            "image": null,
            "audio": null,
            "character": null,
            "image_output": null,
            "image_output_token": null,
            "video_second": null
          }
        ]
      }
    ]
  }
}

GET /families/:author/:slug/activity/series

GET https://api-staging.heimdallm.com/api/v1/families/deepseek/deepseek-v4-flash-0731/activity/series?days=30

A family's billed usage per UTC day, the model route's counts summed over every active model in the family, a model with no live endpoint included. days is 1 to 30, default 30; only days with traffic appear. Every billed request counts, the platform's own paid probe traffic included.

{
  "data": {
    "family_id": "deepseek/deepseek-v4-flash-0731",
    "window_days": 30,
    "daily": [
      {
        "date": "2026-08-16",
        "requests": 5120,
        "prompt_tokens": 161052110,
        "completion_tokens": 40881205,
        "images": 0,
        "megapixels": 0,
        "audio_seconds": 0
      }
    ]
  }
}

GET /providers/:slug/series

GET https://api-staging.heimdallm.com/api/v1/providers/impcc/series

Hourly request outcomes (ok, error) across one provider's live endpoints over the last 7 days — the data behind the provider page's availability chart. days is 1 to 7, default 7; hours are UTC and sparse.

GET /status

GET https://api-staging.heimdallm.com/api/v1/status

The live status board — what the status page renders: one row per live endpoint with its health tier, uptime over 1 and 7 days, median response time (last 30 minutes, falling back to the last day), and the time and classification of its last error. Cached server-side for about half a minute. There is no separate status API beyond this one.

{
  "data": [
    {
      "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
      "endpoint_id": "9e8d7c6b-…",
      "health": "healthy",
      "uptime_1d": 99.8,
      "uptime_7d": 99.9,
      "response_p50": 2.6,
      "response_1d_p50": 2.7,
      "last_error": null,
      "last_success_at": "2026-08-17T12:00:03Z"
    }
  ]
}

last_error, when present, is { "at", "status_code", "error_type" } — status_code and error_type are null when the failure did not come from an HTTP exchange.

GET /stats/public

GET https://api-staging.heimdallm.com/api/v1/stats/public

Marketplace-wide usage over the last 30 days. Every billed request counts — the platform's own paid probe traffic included — and no monetary figure ever appears here. This is what the activity page renders. The response is cached server-side for about a minute, and — unlike the other endpoints — it is not wrapped in a data envelope. (Heimdallm extension.)

{
  "totals": { "requests": 128441, "tokens": 5013882211, "models_served": 3 },
  "window_days": 30,
  "daily": [
    {
      "date": "2026-08-16",
      "requests": 5120,
      "prompt_tokens": 161052110,
      "completion_tokens": 40881205,
      "requests_by_api_type": { "chat": 4890, "images": 230 },
      "tokens_by_api_type": { "chat": 201933315 }
    }
  ],
  "api_types": [
    { "api_type": "chat", "requests": 118200, "tokens": 5013882211 },
    { "api_type": "images", "requests": 10241, "tokens": null }
  ],
  "models": [
    {
      "model_id": "deepseek/deepseek-v4-flash-0731-abliterated",
      "author": "deepseek",
      "name": "DeepSeek: DeepSeek V4 Flash 0731 (abliterated)",
      "api_type": "chat",
      "requests": 91002,
      "tokens": 4410882140
    }
  ],
  "authors": [{ "author": "deepseek", "requests": 91002, "tokens": 4410882140, "share": 0.88 }]
}

Field	Type	Notes
`totals`	object	`requests`, `tokens` (prompt + completion, every request that reported either) and distinct `models_served` in the window.
`window_days`	integer	Always 30.
`daily`	object[]	Per-UTC-day series, oldest first; only days with traffic appear. `requests_by_api_type` and `tokens_by_api_type` split the day, and omit a type with none.
`api_types`	object[]	One row per API type served in the window, most requests first. `tokens` is `null` for a type none of whose requests reported a token count.
`models`	object[]	Top models by requests, at most 50. `api_type` is the one that served most of the model's requests; `tokens` is `null` where no request reported a count.
`authors`	object[]	Requests and tokens per model author, most requests first; `share` is the fraction of the tokens reported across authors, and `null` for an author with none.

Requests are the one measure every API type has. A token count exists only where an endpoint reported one, so an image, video or 3D generation has tokens: null rather than a zero, and nothing about it is ranked on tokens.

Video soundtrack controls

Video model summaries and endpoint rows may include video_generation.generate_audio. supported_values lists the declared boolean choices; default is present only where a default is known. Missing fields do not establish either choice or a universal default. The endpoint’s video_envelope.audio says whether generated sound is a capability, not whether every returned clip contains sound. See video generation.