Updated 21 November 2026: the current generations covered here are Mistral Small 4 (March 2026, Apache 2.0, 119B MoE with 6B active), Qwen3.5 (February 2026, 397B-A17B, open weights) and Qwen3.6 (April 2026). The model rows below reflect these; anything not re-verified is marked as such. Mistral and Qwen are the two strongest "second-tier" open-weights model families in 2026 — below Llama on raw ecosystem reach but each leading on a specific dimension that matters in production. Mistral AI (France) leads on function-calling reliability, European multilingual, and Apache-2.0-licensed releases. Qwen (Alibaba) leads on code-specific work via Qwen Coder and on Asian multilingual including Chinese, Japanese, and Korean. Together they cover use cases where Llama and Gemma are the wrong shape.
Mistral AI is the French AI lab founded in 2023 by ex-DeepMind and ex-Meta researchers. The company shipped Mistral 7B in late 2023, Mixtral 8x7B (the first popular open-weights mixture-of-experts model) in early 2024, Mistral Large in 2024, and continues to ship at frequent cadence. The pattern: most Mistral models are released under Apache 2.0 — genuinely open-source, commercially unrestricted — while some flagship variants ship under a commercial-research licence. The mix gives you Apache-licensed open weights for production and the flagship-tier for paid hosted access.
Qwen is Alibaba's open-weights family. Qwen3.5 (February 2026) is a natively multimodal 397B-A17B mixture-of-experts model with open weights. Qwen3.6 (April 2026) added open 35B-A3B releases, while the top Alibaba tiers (Qwen3.6-Plus, Qwen3.7-Max) are API-only. The earlier Qwen 2.5 and Qwen 3 generations, which span edge sizes up to 72B and 235B, are still available. Qwen 2.5 Coder (the last Coder generation checked in this review; newer Coder releases not verified) is the specialised code variant and is widely considered the strongest open-weights coding model in its size class — meaningfully better than Llama or Gemma at code generation, refactoring, and software-engineering tasks. The general Qwen models also lead on Asian-language work given Alibaba's training-data emphasis.
Both families are widely hosted — Ollama, Together AI, Fireworks, Hugging Face, vLLM-via-self-host, and (for Qwen) the official Alibaba Cloud Model Studio. Both run anywhere Llama runs, with the same OpenAI-compatible API surface.
This page treats Mistral and Qwen as a complementary pair because that's how they're typically chosen in production: you don't pick "Mistral or Qwen"; you pick "Mistral for X, Qwen for Y". They cover different shapes of work that Llama and Gemma don't fit cleanly. Combining them keeps the comparison logic together and reflects the way teams actually evaluate the open-weights field. If either becomes load-bearing enough to warrant a deeper dedicated leaf in future, the split is straightforward.
Named by Mistral as the recommended successor to Mistral Medium 3.5 on its model page. Specifications, licence and pricing not verified in this review; check docs.mistral.ai before relying on it.
Released 16 March 2026 (mistral-small-2603), Apache 2.0. 119B mixture-of-experts with 6B active parameters and a 256k context. Consolidates the Magistral reasoning and Devstral coding lines into one model. API price $0.15 / $0.60 per million tokens. The open-weights pick to test first when function-calling matters.
Open weights under a modified MIT licence, released 28 April 2026 with a 256k context. Mistral's model page lists it as deprecated from 31 October 2026 and recommends Large 4 for new integrations. Do not select for new work.
Released 16 February 2026 as an open-weights, natively multimodal mixture-of-experts model: 397B total, 17B active per token, aimed at reasoning, coding and agent work. Strong multilingual including Chinese / Japanese / Korean. Benchmark figures are vendor-published and not independently reproduced here.
Open-weights mixture-of-experts (35B total, 3B active), released April 2026 on Hugging Face and ModelScope and positioned as an agentic coder. Sources differ on the exact April date. Runs on Ollama, vLLM and Together for broad hosting support.
Last Coder generation checked in this review. Earlier vendor benchmarks put it ahead of Llama and Gemma on code at equal size; not re-verified. Apache 2.0. Confirm whether a newer Coder release exists before building on it.
Proprietary and API-only, with no open weights. Qwen3.7-Plus is the companion model. Qwen3.6-Plus and Qwen3.6-Max-Preview are also closed. Not an option for self-hosting.
Earlier revisions of this leaf quoted benchmark leads for Mistral Large 2 and Qwen 2.5 Coder 32B. Those described superseded generations and have been removed until re-verified against the current models. The figures that remain for Qwen3.5 and Mistral Small 4 are vendor-published, so treat them as claims to check rather than settled rankings.
The honest framing: both families are second-tier ecosystems vs Llama. You pick them when their specific strengths matter more than ecosystem reach. Choose Llama by default; choose Mistral or Qwen when their structural advantage matches your use case.
| Use case | Best pick | Why |
|---|---|---|
| General-purpose open-weights workhorse | Llama 3.x / 4 | Largest community ecosystem; broadest hosting |
| Frontier-lab safety tuning + multimodal | Gemma 3 | Google DeepMind training pipeline |
| Function-calling-heavy production agents | Mistral Small 4 | Current Apache 2.0 Mistral release; test against your own tool schema |
| European multilingual (FR / DE / IT / ES) | Mistral Small 4 / Large 4 | Current Mistral line; not re-benchmarked on European languages here |
| Apache 2.0 strict licence requirement | Mistral Small 4 | Genuinely open-source vs Llama Community Licence |
| Code-generating agents on open-weights | Qwen 2.5 Coder (confirm newer Coder release) | Last Coder line checked here; older vendor benchmarks only |
| Asian-language work (CN / JA / KO) | Qwen3.5 | Alibaba training-data emphasis |
| Mixture-of-experts inference economics | Mistral Small 4 | 119B MoE with 6B active parameters per token |
| Frontier reasoning quality | Claude / GPT / Gemini 4 Argon | Closed frontier still leads on hardest reasoning |
For SA studios building production agents on open-weights with function-calling load-bearing — tool dispatch, structured outputs, multi-step workflows — Mistral Small 4 is the current candidate to test against Llama or Gemma at the tool-use surface. Hosted via Together / Fireworks; self-host via vLLM. The Apache-licensed Mistral releases (Small 4) avoid the Llama 700M MAU edge cases entirely and the "Built with Llama" attribution requirement.
For SA studios building internal coding agents (PR review, refactor assistants, codebase Q&A) on open-weights, Qwen 2.5 Coder is the last Coder line checked here, so confirm the current one before committing. Runs on Ollama for dev work; production via Together AI or self-hosted vLLM. The structural cost advantage over Claude Code or GPT-driven coding agents matters at high-volume internal usage — especially when many developers run many agent calls per day.
Mistral and Qwen variants are less consistently available in africa-south1 (Vertex) or af-south-1 (Bedrock) than Llama. Most Mistral variants are reachable via Together AI / Fireworks / Mistral's own la Plateforme; most Qwen variants via Together AI / Fireworks / Alibaba Cloud directly. For strict SA-residency requirements, check current cloud provider model availability before committing — the situation evolves quickly. For development, Ollama on a maxed MacBook handles all three families locally regardless of hosting question.
ollama run mistral or ollama run qwen2.5-coder — one-command access to the canonical variants.langchain-mistralai and langchain-community.