gpt-oss, Gemma 4 and Mistral Small 4 are open-weight language models from OpenAI, Google and Mistral AI that all use the Apache 2.0 licence, yet they differ widely in size, hardware needs, reasoning controls, modalities and hosted pricing. gpt-oss (gpt-oss-120b and gpt-oss-20b) was released on 5 August 2025, Gemma 4 on 31 March 2026 and Mistral Small 4 on 16 March 2026. This page puts the three families side by side with identical fields: parameters, context window, output limits, knowledge cutoff, licence terms, hardware statements, reasoning and tool controls, hosted availability and price, data use, safety statements and provider-reported benchmarks. It is written for developers and architects choosing a model to self-host or call by API, and for buyers and governance teams who need checked facts. Providers are listed alphabetically (Google, Mistral AI, OpenAI); the order is not a ranking, and the page gives no verdict.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Where a provider publishes nothing on a point, the page says "Not stated" rather than estimating.
Key takeaways
- All three use Apache 2.0. Google's Gemma 4 model card, Mistral AI's Small 4 model card and OpenAI's gpt-oss model pages each state the Apache 2.0 licence; OpenAI's launch post adds usage guidance, and Mistral's Hugging Face card says the model must not be used to infringe third-party rights.
- Sizes span 2.3B effective to 119B total parameters. Gemma 4 ranges from E2B to a 31B dense model; Mistral Small 4 is 119B total with 6.5B active; gpt-oss-120b is 117B total with 5.1B active and gpt-oss-20b is 21B with 3.6B active.
- Context windows differ. Gemma 4 lists 128K (E2B, E4B) or 256K (12B Unified, 26B A4B, 31B); Mistral Small 4 lists 256k; OpenAI lists 131,072 tokens for gpt-oss-120b.
- Hosted prices take three forms. Google lists Gemma 4 on the Gemini API free tier only; Mistral AI lists Small 4 at $0.15 and $0.60 per million tokens; OpenAI's gpt-oss-120b page lists no price and zero rate limits on its API tiers.
- Each family controls reasoning differently. Gemma 4 uses a think token in the system prompt, Mistral Small 4 a reasoning_effort of none or high, and gpt-oss low, medium or high effort set in the system message.
What are Gemma 4, Mistral Small 4 and gpt-oss?
- Gemma 4 (Google): a family of open models in five sizes, E2B, E4B, 12B Unified, 26B A4B (Mixture of Experts) and 31B (dense), with text and image input on every size and audio on the three smallest. Two sizes, gemma-4-31b-it and gemma-4-26b-a4b-it, are also callable on the Gemini API. See the Gemma 4 profile and the Google Gemma Models hub.
- Mistral Small 4 (Mistral AI): a single 119B Mixture-of-Experts model, version 26.03 with API ID mistral-small-2603, that combines chat, reasoning and image input in one model. See the Mistral Small 4 profile and the Mistral AI Models hub.
- gpt-oss (OpenAI): two text-only reasoning models, gpt-oss-120b and gpt-oss-20b, using Mixture of Experts and natively quantised in MXFP4. See the gpt-oss-120b and gpt-oss-20b profile and the OpenAI Models hub.
All three families use or include sparse expert designs, which is why total and active parameter counts are both quoted. See Mixture of Experts Explained. For the largest open model from each provider, which this page does not repeat, see Open-Weight Frontier Models Compared.
The three families at a glance
As of 2026-11-15. Sources: Google Gemma 4 model card and releases page; Mistral Small 4 model card, launch post and Hugging Face card; OpenAI gpt-oss-120b model page, launch post and Hugging Face card.
| Field | Gemma 4 (Google) | Mistral Small 4 (Mistral AI) | gpt-oss (OpenAI) |
|---|---|---|---|
| Released | 2026-03-31 (E2B, E4B, 31B, 26B A4B); 2026-06-03 (12B Unified) | 2026-03-16 | 2025-08-05 |
| Sizes | E2B, E4B, 12B Unified, 26B A4B, 31B | One size, 119B | gpt-oss-120b, gpt-oss-20b |
| API IDs | gemma-4-31b-it, gemma-4-26b-a4b-it (Gemini API) | mistral-small-2603 | gpt-oss-120b, gpt-oss-20b |
| Context window | 128K (E2B, E4B); 256K (12B Unified, 26B A4B, 31B) | 256k | 131,072 tokens (120b model page); 128k (launch post, both) |
| Max output | Not stated by Google on the pages checked (2026-11-15) | Not stated by Mistral AI on the pages checked (2026-11-15) | 131,072 tokens (120b model page) |
| Knowledge cutoff | January 2025 (pre-training data) | Not stated by Mistral AI on the pages checked (2026-11-15) | 1 June 2024 (120b model page) |
| Input modalities | Text and image on all sizes; audio on E2B, E4B, 12B Unified | Text, image | Text |
| Output modalities | Text | Text | Text |
| Licence | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Languages | Over 140 in pre-training; 35+ out of the box | 24 tagged on the Hugging Face card | Not stated by OpenAI on the pages checked (2026-11-15) |
When was each model released?
| Date | Provider | Event | Source |
|---|---|---|---|
| 2025-08-05 | OpenAI | gpt-oss-120b and gpt-oss-20b released with a Red Teaming Challenge carrying a $500,000 prize fund | OpenAI launch post |
| 2026-03-16 | Mistral AI | Mistral Small 4 (mistral-small-2603, v26.03) released, status GA | Mistral model card; launch post |
| 2026-03-31 | Gemma 4 released in E2B, E4B, 31B and 26B A4B sizes | Gemma releases page | |
| 2026-04-16 | Multi-token prediction (MTP) models released for E2B, E4B, 31B and 26B A4B | Gemma releases page | |
| 2026-06-03 | Gemma 4 12B Unified released | Gemma releases page |
gpt-oss is the oldest of the three by about seven months, which shows in its knowledge cutoff. Neither OpenAI nor Mistral AI lists a second version of these models on the pages checked.
How large is each model and what hardware does it need?
| Model | Total / active parameters | Architecture as stated | Hardware stated by the provider |
|---|---|---|---|
| Gemma 4 E2B / E4B | 2.3B / 4.5B effective (5.1B / 8B with embeddings) | Dense; 35 and 42 layers | Not stated by Google on the pages checked (2026-11-15) |
| Gemma 4 12B Unified | 11.95B | Dense; 48 layers | Not stated by Google on the pages checked (2026-11-15) |
| Gemma 4 26B A4B | 25.2B total, 3.8B active | Mixture of Experts; 30 layers | Not stated by Google on the pages checked (2026-11-15) |
| Gemma 4 31B | 30.7B | Dense; 60 layers | Not stated by Google on the pages checked (2026-11-15) |
| Mistral Small 4 | 119B total, 6.5B active (model card); 6B active, 8B with embedding and output layers (launch post) | Mixture of Experts, 128 experts, 4 active | Minimum 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200; recommended 4x HGX H100, 4x HGX H200 or 2x DGX B200 |
| gpt-oss-120b | 117B total, 5.1B active | Mixture of Experts, 36 layers, 128 experts, 4 active; MXFP4 | A single 80 GB GPU, such as NVIDIA H100 or AMD MI300X |
| gpt-oss-20b | 21B total, 3.6B active | Mixture of Experts, 24 layers, 32 experts, 4 active; MXFP4 | About 16 GB of memory |
Weight formats also differ: Mistral AI publishes BF16 and FP8 (F8_E4M3) tensors and links a separate NVFP4 checkpoint; OpenAI ships MXFP4-quantised Mixture-of-Experts weights; Google publishes MTP drafter models for speculative decoding alongside the main Gemma 4 checkpoints. Hardware statements describe the provider's own reference set-up; quantised community builds change the numbers.
What does each licence allow?
| Provider | Licence | Additional terms on the pages checked | Where the weights are |
|---|---|---|---|
| Apache 2.0 | None beyond the licence on the model card | Open weights linked from Google's Gemma documentation | |
| Mistral AI | Apache 2.0 | The Hugging Face card states the model must not be used in ways that infringe third-party rights | mistralai/Mistral-Small-4-119B-2603 on Hugging Face |
| OpenAI | Apache 2.0 | The launch post asks developers not to show chain of thought directly to users and states the models do not replace a medical professional | openai/gpt-oss-120b and openai/gpt-oss-20b on Hugging Face |
A shared licence makes legal review simpler than it is for publisher-specific licences, but the provider guidance above still belongs in internal usage rules. See Open-Weight vs Closed Models Explained and the open-weight model licence tracker.
How much does each model cost?
As of 2026-11-15. Weights are free to download for all three. The table covers each provider's own hosted API.
| Item | Gemma 4 (Gemini API) | Mistral Small 4 (Mistral API) | gpt-oss (OpenAI API) |
|---|---|---|---|
| Input per 1M tokens | Free of charge (free tier); paid tier listed as not available | $0.15 | Not stated by OpenAI on the pages checked (2026-11-15) |
| Output per 1M tokens | Free of charge (free tier); paid tier listed as not available | $0.60 | Not stated by OpenAI on the pages checked (2026-11-15) |
| Context caching | Free of charge, including storage (free tier) | Not stated by Mistral AI on the pages checked (2026-11-15) | Not stated by OpenAI on the pages checked (2026-11-15) |
| API access note | API key from Google AI Studio | Mistral API and AI Studio | Model page lists 0 RPM and 0 TPM for Build, Launch and Grow tiers; Free tier not supported |
For gpt-oss, the practical hosted options are the platforms OpenAI's launch post names, including Azure, AWS, Hugging Face, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare and OpenRouter, each with its own prices. Third-party hosts also serve Gemma 4 and Mistral Small 4 at their own rates. Worked example: 100,000 input and 10,000 output tokens on Mistral Small 4 cost $0.015 plus $0.006, about $0.02, at list price. Reasoning tokens count as output on every model that produces them. See the LLM Pricing Tracker and Self-Hosted LLMs vs Hosted LLM APIs.
What is the context window of each model?
| Limit | Gemma 4 | Mistral Small 4 | gpt-oss |
|---|---|---|---|
| Context window | 128K (E2B, E4B); 256K (12B Unified, 26B A4B, 31B) | 256k (vLLM example uses max-model-len 262144) | 131,072 tokens on the 120b model page; 128k for both in the launch post |
| Max output | Not stated by Google on the pages checked (2026-11-15) | Not stated by Mistral AI on the pages checked (2026-11-15) | 131,072 tokens on the 120b model page |
| Long-context evaluation published | MRCR v2 8-needle 128K: 66.4% (31B), 44.1% (26B A4B) | AA LCR: 0.72 with about 1.6K characters of output | Not stated by OpenAI on the pages checked (2026-11-15) |
Self-hosted deployments can set a lower limit than the model supports to save memory. See Context Windows Explained.
Reasoning, tools and modalities
| Feature | Gemma 4 | Mistral Small 4 | gpt-oss |
|---|---|---|---|
| Reasoning control | Thinking on when the system prompt starts with the <|think|> token; on the Gemini API, thinking level high or minimal | reasoning_effort none or high | Effort low, medium or high, set in the system message |
| Function calling | Native support for structured tool use | Function Calling listed | Few-shot function calling listed |
| Structured output | Function declarations on the Gemini API | Structured Outputs listed | Structured Outputs listed |
| Built-in tools named | Google Search grounding on the Gemini API | Built-In Tools, Agents and Conversations, Document QnA | Web browsing and Python code execution |
| Prompt format note | Recommended sampling: temperature 1.0, top_p 0.95, top_k 64 | Temperature 0.7 with high effort; 0.0 to 0.7 with none; strong system prompt adherence | Must use the harmony response format |
| Image input | All sizes | Yes | Text only |
| Audio input | E2B, E4B, 12B Unified | Not stated by Mistral AI on the pages checked (2026-11-15) | Text only |
The prompt format matters when self-hosting: OpenAI's Hugging Face card says gpt-oss will not work correctly without harmony formatting, which the Transformers chat template applies automatically. For reasoning controls across providers, see Reasoning Models Explained and Reasoning LLMs Compared Across Providers.
Where can you run each model?
| Channel | Gemma 4 | Mistral Small 4 | gpt-oss |
|---|---|---|---|
| Provider API | Gemini API (31B and 26B A4B) | Mistral API and AI Studio | Model page lists API endpoints with zero rate limits on paid tiers |
| Weights download | Open weights linked from Google's Gemma documentation | Hugging Face | Hugging Face |
| Inference engines named by the provider | Not stated on the model card checked (2026-11-15) | vLLM (recommended for production), llama.cpp, LM Studio, SGLang, Transformers | Transformers, vLLM, SGLang, PyTorch and Triton reference code, Ollama, LM Studio, Docker Model Runner, llama.cpp builds |
| Partner platforms named by the provider | Not stated on the pages checked (2026-11-15) | NVIDIA build.nvidia.com, NVIDIA NIM, NVIDIA NeMo | Azure, AWS, Hugging Face, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, OpenRouter |
| Fine-tuning statement | Gemini API tuning listed as not available on the paid tier | Axolotl named on the Hugging Face card | gpt-oss-120b fine-tunable on a single H100 node |
Google's ai.google.dev Gemma pages now carry a notice that the documentation is deprecated and point to gemma.ai, so check there for current details. Self-hosting moves patching, network exposure and access control to your team: see the Self-Hosted LLM Security Guide and the vLLM, Triton and NVIDIA NIM Inference Server Security Guide.
Are my prompts used for training?
- Self-hosted (all three): prompts stay on infrastructure you control; no provider terms apply to the data, only the licence.
- Google (Gemini API): the pricing page marks Gemma 4 free tier content as used to improve Google's products; the paid tier, where content is not used, lists every Gemma 4 line as not available.
- Mistral AI: its help centre says that in Mistral Studio free mode Mistral may use input and output for training, with opt-out at any time, and that pay-as-you-go customers can also opt out at any time.
- OpenAI: with zero rate limits on OpenAI's own API tiers, data terms are those of the third-party host chosen.
Read the host's data processing terms before sending personal or confidential data. See the AI Training Opt Out Settings Guide.
Benchmarks
As of 2026-11-15. Provider-reported figures only, from each provider's own model card or launch page (Gemma 4 model card; Mistral Small 4 Hugging Face card, March 2026; OpenAI launch post, August 2025). The three providers publish different suites, so only one evaluation appears on more than one page.
| Evaluation | Gemma 4 | Mistral Small 4 | gpt-oss |
|---|---|---|---|
| GPQA Diamond | 84.3% (31B); 82.3% (26B A4B) | 71.2 | Not stated as a figure by OpenAI on the pages checked (2026-11-15) |
| How the provider frames results | Nine evaluations with figures for 31B and 26B A4B, including MMLU Pro 85.2% and LiveCodeBench v6 80.0% (31B) | Compared with gpt-oss-120b on output length; 40% lower completion time and 3x requests per second against Mistral Small 3 | Compared with OpenAI o3-mini and o4-mini; gpt-oss-120b described as near parity with o4-mini on core reasoning benchmarks |
Settings, harnesses and reasoning levels differ between providers, so these figures are not a like-for-like ranking. Independent leaderboard results are not included because none were checked against the leaderboards' own pages for this comparison. Test on your own prompts at the effort level you plan to use. See How to Read LLM Benchmarks.
Model cards and safety statements
| Provider | Documents | Safety statements on the pages checked |
|---|---|---|
| Gemma 4 model card; Gemma releases page; 2026 technical report cited | Major content safety improvements over Gemma 3 and 3n, low unjustified refusals, and minimal policy violations when tested without safety filters | |
| Mistral AI | Mistral Small 4 model card; Hugging Face card; launch post | No dedicated safety section on the pages checked (2026-11-15); the Hugging Face card states the model must not be used to infringe third-party rights |
| OpenAI | Model pages; launch post; Hugging Face card | Harmful CBRN data filtered in pre-training; deliberative alignment and instruction hierarchy in post-training; adversarially fine-tuned versions did not reach high capability under OpenAI's Preparedness Framework |
Once weights are downloaded, provider-side safeguards no longer apply, and fine-tuning can change model behaviour. Plan your own guardrails and logging. See the Open-Weight Model Security Guide and the Hugging Face Security Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Image input needed | Gemma 4 (all sizes) and Mistral Small 4; gpt-oss is text only |
| Audio input needed | Gemma 4 E2B, E4B and 12B Unified |
| Single 80 GB GPU stated | gpt-oss-120b |
| Edge or 16 GB memory stated | gpt-oss-20b; Gemma 4 offers E2B and E4B sizes, with no memory figure on the pages checked |
| Context above 131K tokens | Gemma 4 12B Unified, 26B A4B and 31B (256K); Mistral Small 4 (256k) |
| Published per-token price on the provider's own API | Mistral Small 4 ($0.15 / $0.60); Gemma 4 free tier only |
| More recent knowledge cutoff stated | Gemma 4 (January 2025) against gpt-oss (1 June 2024); Mistral AI states none |
| Reasoning can be switched off | Gemma 4 (omit the think token) and Mistral Small 4 (none); gpt-oss offers low effort |
For the full open-weight field, see Open-Weight Frontier Models Compared; for closed frontier models, Frontier AI Models Compared.
Practitioner checklist
- Record the exact checkpoint (for example gemma-4-31b-it, mistral-small-2603 or gpt-oss-120b) and its source repository in the AI inventory.
- Verify checkpoint files come from the provider's official Hugging Face organisation before deployment.
- Apply each provider's prompt requirements: harmony format for gpt-oss, the think token for Gemma 4, reasoning_effort for Mistral Small 4.
- Size hardware from the provider's stated set-up, then load-test with your real context lengths.
- Write the provider guidance (OpenAI's chain-of-thought note, Mistral's third-party rights note) into internal usage rules.
- Check the data terms of any hosted endpoint, including free-tier training use, before sending sensitive data.
- Put authentication, rate limits and logging in front of self-hosted inference servers, and store any hosted API keys in a secrets manager.
- Run your own evaluation at the effort level you plan to use instead of relying on provider benchmarks.
Standards and references
- Google: Gemma 4 model card (accessed 2026-11-15)
- Google: Gemma releases (accessed 2026-11-15)
- Google: Gemma on the Gemini API (accessed 2026-11-15)
- Google: Gemini API pricing (accessed 2026-11-15)
- Mistral AI: Mistral Small 4 model card (accessed 2026-11-15)
- Mistral AI: Mistral Small 4 launch post (accessed 2026-11-15)
- Mistral AI: Mistral-Small-4-119B-2603 on Hugging Face (accessed 2026-11-15)
- Mistral AI Help Center: Do you use my user data to train your models? (accessed 2026-11-15)
- OpenAI: gpt-oss-120b model page (accessed 2026-11-15)
- OpenAI: Introducing gpt-oss (accessed 2026-11-15)
- OpenAI: gpt-oss-120b on Hugging Face (accessed 2026-11-15)
Related NHI Mgmt Group resources: Gemma 4, Mistral Small 4, gpt-oss-120b and gpt-oss-20b, Open-Weight Frontier Models Compared, Open-Weight Model Security Guide.