Mistral Small 4: one model, three jobs, and what ‘reasoning effort’ means for serving
“Four models, four deployments, four scaling policies, four monitoring dashboards. Or: one model with a dial.”
TL;DR
Mistral Small 4 (March 16, 2026) folds the Magistral, Pixtral and Devstral lines into the Mistral Small family as one 119B MoE model with 6B active params. The reasoning_effort parameter replaces model routing with a per-request switch. 256K context, Apache 2.0, runs on vLLM. The ML systems story: consolidation from many specialized deployments to one.

What did Mistral consolidate?
Three flagship models, each previously requiring its own deployment:
| Previous model | Capability | Now in Small 4 |
|---|---|---|
| Magistral | Mathematical and logical reasoning | reasoning_effort: high |
| Pixtral | Vision and multimodal understanding | Image input support |
| Devstral | Agentic coding, tool use | Code + function calling |
Each had its own deployment target, its own GPU allocation, its own scaling policy, and its own monitoring. An organization using all three, plus Mistral Small for general chat, ran four separate inference services. Mistral Small 4, which Mistral calls the next major release in the Small family, absorbs all three capabilities.
Mistral Small 4 is one model that handles all four jobs. Inside, a router picks 4 of 128 experts for each token. Mistral does not say that particular experts specialize in reasoning or vision, so do not assume a clean split. What matters for serving is that the model weights, the serving infrastructure and the API endpoint are shared.
What does reasoning_effort change about serving?
The conventional approach to serving models of varying capability is model routing: a classifier examines incoming requests and routes simple ones to a cheap model and complex ones to an expensive model. This works but requires maintaining multiple deployments and building the router itself (which must be fast, accurate, and maintained).
Mistral’s reasoning_effort parameter flips this. Instead of choosing which model handles a request, you choose how much reasoning a single model applies.
reasoning_effort: none: fast responses for extraction, formatting, simple classification. Equivalent to Mistral Small 3.2 behavior. The model generates directly with minimal deliberation.reasoning_effort: high, extended reasoning for math, multi-step logic, and complex analysis, with output as verbose as previous Magistral models. Per-token compute does not change. The model writes a longer reasoning trace before answering, so a request costs more tokens, more GPU time and more latency.
Mistral’s launch post documents two positions, not a gradient. You either want reasoning or you do not. This simplicity is a feature: no tuning a continuous dial, no wondering whether “medium” is right. For capacity planning, treat high-effort requests as long-output requests.
The infrastructure implication: one deployment serves all tiers. Scaling is one autoscaling policy, not three. Monitoring is one dashboard. Model updates are one rollout. The operational surface area shrinks proportionally to the number of consolidated models.
What are the MoE economics?
119B total parameters sounds large. 6B active parameters per token is small (8B counting embedding and output layers, per Mistral), comparable to Llama-3.1-8B in compute cost but with access to 119B parameters of learned knowledge through the expert routing layer.
The MoE trade-off is always the same: memory for quality. You need enough GPU memory to hold the full 119B parameter set (even though only 6B activate per token), because any expert might be needed for the next token. At 16-bit precision the weights alone take about 238 GB (119B × 2 bytes), which already needs three 80 GB GPUs before any KV cache. Mistral’s launch post gives the minimum as 4x HGX H100, 2x HGX H200 or 1x DGX B200. At 4-bit quantization the weights drop to roughly 60 GB, which fits on one 80 GB GPU with room left for cache. For picking the format, see the quantization decision tree.
The compute cost per token, however, scales with the active parameters (6B), not the total. This is why MoE models offer better quality-per-FLOP than dense models: you get the pattern recognition capacity of 119B parameters at the inference cost of 6B.
For the KV cache implications of serving MoE models, see KV cache for MoE. The memory savings from MoE’s sparse compute do not extend to the attention cache, which remains dense.
When does consolidation make sense?
Consolidate when:
- You run 3+ model deployments for different capabilities
- Operational complexity (monitoring, scaling, updates) is a bigger cost than inference
- No single specialized model is dramatically better than Small 4 on your benchmarks
- You want a single API contract for clients, regardless of task type
Keep separate when:
- One task dominates (90%+ of traffic) and a specialized model significantly outperforms
- Latency requirements differ dramatically between tasks (vision at 2s is fine, chat at 200ms is required)
- You need different scaling policies for different workloads (batch vision processing vs real-time chat)
The 256K context window and Apache 2.0 license remove two common blockers. The context handles long-document tasks. The license permits commercial deployment and fine-tuning, with only attribution and notice requirements.
Consolidation is one input to a wider choice between self-hosting and an API. The build vs buy framework covers when running your own model is worth it at all.
Key takeaways
- Three flagships, one successor. Reasoning, vision, coding, all in a single 119B MoE deployment with 6B active params.
- reasoning_effort replaces model routing. One API parameter instead of a routing classifier and multiple deployments.
- Operational surface area shrinks. One pipeline, one monitor, one scaling policy.
- MoE trade-off applies. 238 GB memory for 6B of compute per token. GPU memory is the binding constraint.
- Evaluate before migrating. Consolidated models trade peak specialization for breadth. Benchmark on your specific workloads.
Further reading
- KV cache for MoE: why MoE’s compute savings do not extend to attention cache
- Mistral Small 4 announcement: Mistral’s launch post with architecture, hardware and reasoning_effort details
- Cost management for agents: FinOps for optimizing model serving costs