Mistral vs Qwen: two open model families, compared

Disclosure first: I build Outlier, a Mac app that runs models on-device, and six of its seven tiers are Qwen weights. The seventh is Gemma. I'm not a neutral referee. What I offer instead is the view of someone who shipped one of these families commercially, read the licenses because he had to, and measured what each size costs in RAM.

What each family ships

Mistral AI is French. Its Mistral 3 release in December 2025 put out Mistral Large 3 — 675B total, 41B active, sparse MoE — plus Ministral 3 at 3B, 8B and 14B, all Apache 2.0, in base, instruct and reasoning variants. Mistral Small 4 followed: also Apache 2.0, 119B total with 6.5B active across 128 experts, 256k context. The API catalogue splits in two, and neither list below is exhaustive — read the pricing page for the current split. Open weights include Medium 3.5, Small 4, Large 3, Ministral 3 at all three sizes, Devstral 2 and the Voxtral speech models. Premier, meaning closed, includes OCR 4, Codestral, Codestral Embed, Magistral Medium and Magistral Small.

Alibaba's Qwen team states it flatly in its own repo — all their open-weight models are Apache 2.0. The current run: Qwen3.5 at 397B-A17B in February, then 122B-A10B, 35B-A3B and 27B, then 9B, 4B, 2B and 0.8B in March; Qwen3.6 at 35B-A3B and 27B in April. Qwen3.6-27B is dense, with a native 262,144-token context extending toward a million with RoPE scaling.

The asymmetry that matters isn't the model count. It's the ladder. Qwen gives you one prompt format from 0.8B to a 397B MoE. Mistral gives you a very good small model, a 14B, then a large jump.

Licensing, the part that decides products

Apache 2.0 is a standard OSI license: use, modify, quantize, fine-tune, redistribute, sell. No approval step, no per-seat fee, no acceptable-use appendix bolted on. Both Mistral's open models and every Qwen open-weight release sit there — checked on the model cards, not the marketing pages.

Compare that to the vendor-specific licenses used elsewhere in open weights — Google's Gemma Terms of Use, Meta's Llama Community License. Those aren't OSI licenses. They're vendor documents with their own use policies you have to read, comply with, and pass through to your users. I know because I do it: Outlier's Quick tier is a Gemma model, so it ships under the Gemma Terms of Use while the other six tiers ship Apache 2.0. That's why Apache 2.0 keeps winning inside product teams.

What Apache 2.0 does not give you: training data or a recipe. Neither vendor publishes the corpus. "Open weights" means redistribution rights over a binary, not reproducibility.

Mistral and Qwen, size for size

What runs on a Mac, at what RAM

This is where the families diverge most. A 4-bit MLX build of Ministral 3 8B is 5.6 GB, leaving real headroom for KV cache on a 16 GB MacBook Air — and it includes a vision encoder, so it's multimodal at that size.

Above that, Mistral gets awkward on a laptop. "Mistral Small 4" is a name, not a size: 119B total won't fit 16 GB at any usable quantization, even with 6.5B active per token. On a 64 GB Mac Studio that sparse shape is genuinely flattering — activation cost stays small while quality tracks something much larger.

Qwen fills the middle. Qwen3.6-27B in 4-bit MLX is 16.1 GB on disk; in Outlier it's the Core tier, 15.13 GB with a 24 GB floor, 20.7 tok/s on an M1 Ultra. Nano is Qwen3.5-4B, 2.37 GB, 6 GB floor, 71.7 tok/s on the Ultra and about 32 on an M4 Air. Fuller mapping: how much RAM you need for local AI.

Where Mistral genuinely wins

I ship Qwen. Mistral is still the better choice in places I care about.

• The 8B slot. Apache 2.0, multimodal, 256k context, 5.6 GB at 4-bit. For a 16 GB machine that needs vision, that's a hard package to argue with.
• Sparse efficiency at the top. 119B total with 6.5B active is the architecture that suits Apple's unified memory on a high-RAM Mac.
• Jurisdiction and a readable paper trail. Mistral is EU-based, and the scope of its training exclusion is written down rather than implied. The privacy policy says Mistral does not use your input and output to train its models "when you use Le Chat Enterprise or the paid version of our APIs" — and that is the whole exclusion. Pro and Team conversations aren't in it; they're covered by an account-level opt-out you have to find and switch on. Narrower than the marketing reads, but legible.
• Legible pricing. Flat published per-million rates, a 50% batch discount, a 90% cached-input discount. Model Studio uses volume-tiered pricing on several models and varies by region, which makes an honest quote harder.

Where Qwen genuinely wins

• Apple Silicon as a first-class target. Qwen's README documents mlx-lm and mlx-vlm at release; Mistral's cards lead with vLLM and datacenter GPUs. For a Mac app that's the difference between integrating a release and porting one.
• One ladder, one prompt format. 0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B — same family, same chat template, same tokenizer. You can move a product from a MacBook Air to a Mac Studio to a datacenter without rewriting the prompt layer, and the tier you outgrow always has a next rung. Mistral's general-purpose open line goes 3B, 8B, 14B, then straight to 119B.
• The 24–48 GB middle. That gap is the practical one. On a 24 GB Mac, Qwen3.6-27B fits in 4-bit at 16.1 GB and 35B-A3B at about 19 GB. Mistral's general-purpose open catalogue has nothing between the 14B and an 119B sparse model that wants a 64 GB machine. Its other open releases are specialists — Devstral 2 for code, Voxtral for speech — not general chat models.
• Context on a dense local model. 262k native on a 27B you can run at home changes whole-repo work.

Two things I expected to put on that list didn't survive checking, so they aren't on it. Ollama carries Ministral 3 as an official library entry — 3B, 8B and 14B, 13 tags, 1.3M pulls, vision and tool use — so distribution is not a Qwen-only advantage; the only hole is Mistral Small 4. And Mistral's Vibe CLI is Apache 2.0, documents an offline local-models path, and lists a Mistral account as optional, so "Qwen's tooling is the local-friendly one" is not true as a general claim. I'd rather run a list where every item checks out than pad the column.

Which family to build on

Pick Mistral if you're targeting 16 GB machines and need vision in a small model, if EU jurisdiction is a procurement requirement, or if you want flat published API pricing without a regional asterisk.

Pick Qwen if you're shipping on Apple Silicon, if you need one family to cover a MacBook Air and a Mac Studio without changing prompt formats twice, or if long context on a dense local model decides your product.

Assume the frontier stays closed either way. github.com/QwenLM/Qwen3.7 returns a 404 and the README lists open weights only through Qwen3.6; Mistral holds OCR 4, Codestral and the Magistral models in its closed Premier tier. Both move fast enough — Qwen shipped 3.5, 3.6 and a closed 3.7 in six months — that bundling weights means a permanent re-quantize treadmill. That's my week.

One last note. Alibaba's "never use your data for model training" commitment lives in the Model Studio docs, and that scope is the paid API. I could not verify the consumer Qwen chat app's data handling — those pages render client-side and returned no policy text. Don't accept a claim about Qwen Chat without a primary read. If you'd rather skip the question, that's the case for running weights yourself.