
Mistral Small 4 vs Mistral Large 3: Specs, Price and Which to Deploy
By Nathan Cole
MyClaw Editorial
MyClaw
Run Best-in-Class AI Agents Now
Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.
AI Takeaway
- Is Mistral Small 4 better than Mistral Large 3?For most new projects, yes: the newer Small 4 adds switchable reasoning, costs less than a third as much per input token and edges Large 3 on the index run by Artificial Analysis, while Large 3 keeps its appeal for fast, non-reasoning multilingual chat.
- How much do the two cost on Mistral's API?Mistral lists Small 4 at $0.15 per 1M input and $0.60 per 1M output tokens, against $0.50 and $1.50 for Large 3, and its pricing page offers 50% off batch jobs and up to 90% off cached input.
- Can Mistral Large 3 reason step by step?Not in a dedicated mode: Large 3 is an instruct model, and Mistral's own model card warns that dedicated reasoning models can beat it on strict reasoning tasks, whereas Small 4 turns reasoning on per request.
- Which is easier to self-host?Small 4, which activates about 6.5B of its 119B parameters and, per Mistral, runs on as little as a single DGX B200, while Large 3's 675B weights need a full eight-GPU node.
- Should I wait for Mistral Large 4 instead?Test Large 4 if you need a flagship, since Mistral opened a public preview on October 6, 2026, but for cheap, quick agent work Small 4 is the safer production choice today.
Two Generations, Two Jobs
Mistral Large 3 arrived on December 2, 2025 as the headline of the Mistral 3 launch. Mistral called it the company's first mixture-of-experts model since the Mixtral series, trained from scratch on 3,000 NVIDIA H200 GPUs. At launch it placed second among open-weight non-reasoning models on the LMArena leaderboard, a ranking Mistral itself reported. Mistral promised a reasoning version as "coming soon", but none appears in its current model list.
Mistral Small 4 followed on March 16, 2026. Mistral describes it as the first model to fold three of its specialist lines into one: Magistral for reasoning, Pixtral for images and Devstral for agentic coding. Instead of choosing among a quick chat model, a thinking model and a vision model, you call one endpoint and set the reasoning level.
Both ship under Apache 2.0, so weights are free to download, modify and run commercially. The same small-versus-flagship question comes up in other families too; our GLM 5.3 Flash and GLM 5.3 comparison works through a similar split.
Mistral also put Large 4 into public preview on October 6, 2026, with a 1M-token window and 52B active parameters out of 1.05T. It is still a preview, so this guide sticks to the two generally available models.

Spec Sheet: Small 4 and Large 3 Side by Side
Every figure below comes from Mistral's documentation and model cards, checked October 14, 2026.
| Item | Mistral Small 4 | Mistral Large 3 |
|---|---|---|
| API ID | mistral-small-2603 | mistral-large-2512 |
| General availability | March 16, 2026 | December 2, 2025 |
| License | Apache 2.0 | Apache 2.0 |
| Architecture | Mixture of experts, 128 experts with 4 active per token | Granular mixture of experts plus a 2.5B vision encoder |
| Parameters | 119B total, about 6.5B active | 675B total, 41B active |
| Context window | 256K tokens | 256K tokens |
| Inputs | Text and images | Text and images |
| Reasoning | reasoning_effort set to none or high | No thinking mode |
| API extras | Function calling, structured outputs, agents, batching, predicted outputs | Function calling, structured outputs, agents, batching |
A small inconsistency: Mistral's launch post gives Small 4's active count as 6B (8B with embedding and output layers), while the docs and model card say 6.5B. Either way, it uses a fraction of the compute Large 3 needs per token.
Small 4's reasoning switch is the practical headline. With reasoning_effort="none" it answers in the style of the older Small 3.2; with "high" it works step by step, with output lengths Mistral likens to its Magistral models. For a broader view of how models rank for agents, see our OpenClaw model picks.
![]()
Benchmarks: Thin Head-to-Head Data
Mistral has not published a table pitting Small 4 against Large 3, so any comparison relies on independent testing.
Independent lab Artificial Analysis currently shows:
| Measure (Artificial Analysis, Oct 14, 2026) | Small 4 (reasoning) | Large 3 |
|---|---|---|
| Intelligence Index | 11 | 9 |
| Output speed | ~165 tokens/sec | ~76 tokens/sec |
| Output tokens used on the index | 54M | 9.8M |
Small 4 scores a little higher and generates more than twice as fast, but it writes about five times as many tokens on the test suite because it reasons. Comparison sites quote other index values for these models, apparently taken from earlier index versions, so stick to one source and one date when you compare.
Mistral's own claims for Small 4, labelled as vendor-reported: 71.2 on GPQA Diamond with reasoning, and 0.72 on the AA-LCR long-context test while writing about 1.6K characters, where Mistral says Qwen models needed 3.5 to 4 times more output for similar scores. Mistral also says Small 4 beats GPT-OSS 120B on LiveCodeBench while producing 20% less output. Large 3's model card presents its scores only as charts against similar-sized open models. Our look at DeepSeek's Flash and Pro tiers shows a similar pattern of an efficient model closing on its bigger sibling.
Pricing on Mistral's API
Mistral's docs quoted these per-million-token rates on October 14, 2026:
| Rate (Mistral) | Small 4 | Large 3 |
|---|---|---|
| Input | $0.15 | $0.50 |
| Output | $0.60 | $1.50 |
| Batch jobs | 50% off | 50% off |
| Cached input | Up to 90% off | Up to 90% off |
Our own back-of-envelope math: 1,000 jobs that each read 20K tokens and write 2K cost about $4.20 on Small 4 and $13.00 on Large 3. The catch is reasoning. With reasoning_effort="high", Small 4 writes far longer answers, so its output bill grows; Artificial Analysis's token counts above suggest the gap narrows a lot on hard tasks. Keep reasoning off for routine calls and turn it on only where accuracy pays for itself. Our guide to trimming agent token spend covers more ways to keep bills down.
Self-Hosting and Deployment
Serving the open weights is where the two differ most.
Small 4. Mistral lists a minimum of four NVIDIA HGX H100s, two HGX H200s or one DGX B200, and recommends four HGX H100s, four HGX H200s or two DGX B200s for best performance. It runs on vLLM, SGLang, llama.cpp and Transformers, and an NVFP4 checkpoint is available.
Large 3. The FP8 build fits one 8×H200 node, and an NVFP4 checkpoint brings it to a single 8×A100 or 8×H100 node through vLLM. Mistral's model card still says Transformers support has not been added, and lists complex deployment as a known limitation.
Large 3's card also gives tuning advice: keep temperature below 0.1 for production, limit the tool list to what the task needs and keep images close to square. At launch, Mistral listed the Mistral 3 family on Amazon Bedrock, Azure Foundry, IBM watsonx and several inference hosts; for Small 4, Mistral's launch post names its own API, Hugging Face weights and a day-one NVIDIA NIM container.
Which Model Fits Which Workload
| Workload | Better pick |
|---|---|
| High-volume chat and customer support | Small 4 with reasoning off |
| Coding agents, maths and multi-step problems | Small 4 with reasoning on |
| Multilingual drafting without a thinking pass | Large 3 |
| Deterministic extraction at low temperature | Either; Large 3's card is explicit about settings |
| On-premises with limited GPUs | Small 4 |
| Flagship-level quality | Trial Large 4 while it is in preview |
Mistral Small 4
Pros: about a third of Large 3's input price, reasoning switch on one endpoint, roughly twice the output speed, fits on far less hardware.
Cons: verbose when reasoning, which erodes the price edge; active-parameter figure differs between Mistral sources; Mistral's launch post names fewer third-party clouds for it than for Large 3.
Mistral Large 3
Pros: far more total parameters, concise answers on independent tests, Mistral claims best-in-class multilingual conversation outside English and Chinese, wide launch availability across clouds.
Cons: no reasoning mode, lower independent index score, needs a full eight-GPU node, now overshadowed by the Large 4 preview.
Verdict: Small 4 Is the Default, Large 3 Is the Specialist

Small 4 is newer, cheaper, faster and more flexible, and it matches or beats Large 3 on the independent numbers we found. Large 3 still makes sense where you want short, non-reasoning answers in many languages, or where your stack already serves it well. If you need Mistral's strongest model, the Large 4 preview is the one to watch. For model picks in the other agent framework, see our list of top models for Hermes.
Run Your Agent Around the Clock With MyClaw
Whichever model you settle on for your own stack, an agent is only useful if it keeps running. MyClaw keeps your agent online in the cloud, whichever framework you pick, so there is nothing to maintain.
- Register on MyClaw; the framework can be OpenClaw or Hermes Agent, your call.
- Choose a model for it. Connect your chat app, which is where replies will land.
- Give the agent a recurring task and let it run while you are offline.
New to the framework? Read our explainer on OpenClaw.
Skip the Setup, Run Best-in-Class AI Agents Now
Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.