DiffusionGemma 26B A4B
google/diffusiongemma-26b-a4bHugging FaceChatVisiondiffusionmoereasoningtoolsA 26B mixture-of-experts diffusion language model from Google. It denoises 256-token blocks in parallel instead of writing one token at a time, so a stream arrives in blocks and a reply under one block arrives in one chunk. Sampling controls such as temperature and seed do not apply to its sampler, and structured output is not available.
Providers
Live endpoints, cheapest first. First token is the median wait for a streamed reply; throughput covers the last day and uptime the last 24 hours.
- ImpCCHealthynvfp469% offCA
- Input
- $0.031
- Output
- $0.124
- Context
- 262K66K out
- First token
- 438mslast 30 min
- Throughput
- 156 tok/s
- Uptime 24h
- 99.8%
| Provider | Context | Input | Output | First token | Throughput | Uptime 24h |
|---|---|---|---|---|---|---|
| ImpCCHealthynvfp469% offCanada | 262K66K max out | $0.031 | $0.124 | 438mslast 30 min | 156 tok/s | 99.8% |
Capabilities and data
What each endpoint declares about tools, reasoning and prompt data, beside the marketplace’s own checks. The key below says what each badge means.
- ImpCCnvfp4Tools VerifiedReasoning Effort levelsData Zero retention
- Verified
- a current check called an offered tool with valid arguments
- Effort levels
- effort levels are declared, and none is failing checks
- Zero retention
- prompts are neither stored nor trained on; the claim passed marketplace review
Pricing
USD per million tokens.
| Provider | Input | Output | Cache read | Cache hit | Effective input |
|---|---|---|---|---|---|
| ImpCCnvfp469% off | $0.031 | $0.124 | $0.00775 | 90.2% | $0.010032 |
Price history
Input price · USD per million tokens
Output price · USD per million tokens
Performance
Timing percentiles cover the labelled window: the last 30 minutes when it has enough samples, otherwise the last day. TTFT is the seconds to the first token, measured only on streamed requests; response is the seconds from request to last byte, measured on every request, streamed or not. Throughput is output tokens per second over the last day, counted only over streamed generations long enough to have a rate — a buffered answer has no generation phase, and a two-token one times the scheduler rather than the model. Tool calls is the share of the last day of tool requests this endpoint answered with a tool call, published once the day holds twenty of them; under that the cell is blank rather than a rate nothing measured.
| Provider | TTFT p50 | p75 | p90 | p99 | Response p50 | p75 | p90 | p99 | Throughput p50 | p75 | p90 | p99 | Tool calls 1d | Runaways 1d | Empty 1d | Capacity |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ImpCCnvfp4 | 438mslast 30 min | 941ms | 1.14s | 1.14s | 956mslast 30 min | 1.37s | 5.04s | 5.24s | 156 tok/s | 257 tok/s | 298 tok/s | 364 tok/s | — | 0.0% | 0.8% | 2/2 free |
Time to first token by hour · last 7 days
Hourly median per provider. The line spans hours with no streamed traffic; where samples are sparse, a dot marks each hour actually timed.
Throughput by hour · last 7 days
Hourly median output tokens per second per provider, over generations long enough to have a rate. The line spans hours with no answer long enough to measure; where samples are sparse, a dot marks each hour actually measured.
Benchmarks
Scores from the benchmark board. Each figure comes from the endpoint with the most evidence for that suite, and a provisional one names how much of its suite it covers.
Uptime
Share of successful requests and probes over the last day and week. Reliability is the share of the requests a provider accepted that it answered; turned away is the share it refused with its own 429 because it was full, which uptime counts as failures.
- ImpCCnvfp41d 99.8%7d 99.7%reliability 1d 99.8%7d 99.7%turned away 1d 0.0%7d 0.0%Healthy
Availability by hour · last 7 days
Share of successful requests and probes per hour and provider. The line spans hours with nothing to measure; where samples are sparse, a dot marks each hour actually measured.
Activity
Prompt tokens and completion tokens per UTC day over the last 30 days. Every billed request counts, the platform's own paid probe traffic included.
curl https://api-staging.heimdallm.com/api/v1/chat/completions \ -H "Authorization: Bearer $HEIMDALL_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "google/diffusiongemma-26b-a4b", "messages": [{ "role": "user", "content": "Hello" }] }'Set HEIMDALL_API_KEY to a key from your account. See the quickstart for routing suffixes and provider preferences.
About
A 26B mixture-of-experts diffusion language model from Google. It denoises 256-token blocks in parallel instead of writing one token at a time, so a stream arrives in blocks and a reply under one block arrives in one chunk. Sampling controls such as temperature and seed do not apply to its sampler, and structured output is not available.
- Author
- Family
- google/diffusiongemma-26b-a4b
- Request parameters
- max_completion_tokens (1–65536), max_tokens (1–65536), reasoning_effort (none, medium), stop (text or list), tool_choice (any value), tools (list)
- Licence
- apache-2.0
- Size
- 14.4B parameters, U8 weights
- Tokenizer
- —
- Available until
- —
From the model card
What the repository states about these weights, read from nvidia/diffusiongemma-26B-A4B-it-NVFP4 on Nov 10, 2026.
- Languages
- —
- Papers
- —
- Datasets
- —
- Revision
- ec4ff3dJul 4, 2026
- Downloads
- 27.2K
- Likes
- 147