Mistral vs Qwen: the open model families compared

I'm Matt Kerr, and I build Outlier, a Mac app that runs models on-device. Six of its seven tiers are Qwen-family weights and one is Gemma-family, so I'm an interested party here, not a neutral one — read this knowing that. What I can offer instead of neutrality is the view from someone who has shipped these weights commercially and read the licenses closely enough to sign off on redistribution.

Who ships what, and in what sizes

Qwen's open catalogue is the wider of the two: Qwen3.6-27B and Qwen3.6-35B-A3B (April 2026), sitting on top of the Qwen3.5 line — 397B-A17B, 122B-A10B, 35B-A3B, 27B, then 9B, 4B, 2B and 0.8B, all released between February and April 2026. One family, one prompt format, from phone-class up to a 397B mixture-of-experts.

Mistral is shaped differently. Mistral 3 (December 2025) brought Mistral Large 3 — 675B total, 41B active — plus Ministral 3 at 3B, 8B and 14B, each in base, instruct and reasoning variants. Mistral Small 4 followed in March 2026: 119B total, 6.5B active across 128 experts, 256k context. The gap that matters if you're building locally is 14B to 119B. Mistral has nothing open in that band, and it's exactly the range a 24GB Mac can hold.

Licensing: the axis that decides what you can ship

Qwen's repo says it flatly — all their open-weight models are Apache 2.0 — and the Qwen3.6-27B card carries the apache-2.0 tag. Mistral Small 4 and Ministral-3-8B-Instruct-2512 are Apache 2.0 on their cards too. That means bundling in a paid app, quantizing, fine-tuning and redistributing the result: no revenue trigger, no user-count threshold, nobody to ask.

That isn't the norm. Other open-weight families ship a bespoke community license carrying its own acceptable-use policy and conditions that switch on above a certain scale. Gemma weights come under Google's Gemma Terms of Use rather than a standard open-source license, with restrictions that flow down to whoever you redistribute to. I know that one first-hand: Outlier's Quick tier is a Gemma model, and it's the single tier where I track a vendor license separately from the other six.

Two caveats. Open weights aren't open training data — neither vendor publishes a corpus or a recipe, so you're getting redistribution rights over a binary, not reproducibility. And licensing is per-model, not per-vendor: as of mid-2026 some third-party write-ups describe one of Mistral's speech models as non-commercial rather than Apache 2.0. I couldn't confirm that from the card, so I won't state it as fact — but open the card for the exact model you intend to ship, every time.

What actually runs on a Mac, and at what RAM

Qwen treats Apple Silicon as a shipping target in its own documentation: the README names mlx-lm and mlx-vlm alongside GGUF/llama.cpp, Ollama, LM Studio, vLLM and SGLang. Community 4-bit MLX builds land immediately — mlx-community/Qwen3.6-27B-4bit is about 16.1 GB and runs from one mlx_lm.server command. Mistral's cards lead with vLLM and datacenter GPUs; the Ministral 3 8B card frames FP8 around a single H200. MLX conversions come from the community instead, and they do arrive — 4-bit Ministral 3 8B is 5.6 GB — but each release is a wait. Ollama shows the same split: Qwen 3.5 and 3.6 are in the library, while the newest Mistral entries are mistral-medium-3.5 and mistral-large-3.

In practice, on a 16GB Mac that 5.6 GB Ministral 3 8B fits with room for KV cache and is multimodal at 256k context — a genuinely good small-model story. Qwen's 4B and 9B fit the same slot; Outlier ships them at 2.37 GB and 5.04 GB. On 24-32GB, Qwen3.6-27B at 4-bit is the interesting option (our Core tier is that model, 15.13 GB, 24 GB minimum). Mistral Small 4 will not fit 16GB at any useful quantization — 119B total is 119B total, even at 6.5B active. Its sparse shape does flatter a 64GB machine, which is the same bet our Plus tier makes with Qwen3.5-397B-A17B: a 209 GB download that streams experts from SSD at roughly 11 GB peak RSS.

Strengths by task, and the language question

Context favours Qwen at local sizes: Qwen3.6-27B is natively 262,144 tokens, extensible toward about 1,010,000 with RoPE scaling, against 256k for Mistral Small 4 and Ministral 3 8B. On agents, the tooling gap is sharper than any model gap. Both vendors open-source a CLI coding agent under Apache 2.0, but Qwen Code points at local models via Ollama or vLLM and speaks the OpenAI, Anthropic and Gemini protocols, while the mistral-vibe README requires a Mistral API key and documents no local endpoint at all.

Mistral's counterweight is specialist breadth — Voxtral for speech and transcription, a dedicated OCR line, Devstral 2 and Codestral for code on the API. Both do vision: Ministral 3 8B bundles a 0.4B encoder into a ~9B package, and Outlier's Vision tier is Qwen3.6-35B-A3B at 19.0 GB.

On language coverage I'll be straight with you: I found no clean, comparable multilingual benchmark on either vendor's primary pages that I'd be willing to quote, and I won't invent one. Test your own languages on your own prompts. What I can point at is jurisdiction, which is usually the real question underneath — Mistral is a French company operating under EU terms, and Alibaba's Model Studio runs in Singapore, US-Virginia, Beijing, Hong Kong, Tokyo and Frankfurt, with non-interchangeable API keys and region-specific pricing.

Side by side

Where each family genuinely wins

Mistral wins on jurisdiction and on the small multimodal slot. If a customer objects to Chinese-headquartered infrastructure, Mistral is an EU company with EU consumer terms, and that alone closes deals. Ministral 3 8B is a strong 16GB package: Apache 2.0, vision included, 256k context, 5.6 GB at 4-bit. Mistral is also more legible commercially — published per-model list prices (Small 4 at $0.15/$0.60, Large 3 at $0.50/$1.50, Medium 3.5 at $1.50/$7.50 per million tokens), a 50% batch discount, 90% off cached input, and clear consumer tiers at $14.99 and $24.99. Their data policy is quotable in a way Qwen's consumer side isn't: API traffic isn't used for training, and Vibe Pro, Team and Enterprise conversations aren't either by default.

Qwen wins on the size ladder and on Apple Silicon. Eight open sizes in one family means a single integration story spans a MacBook Air and a Mac Studio without changing vendors, prompt formats or eval harnesses. MLX support is documented by the vendor at release rather than reconstructed weeks later, which for a Mac app is the difference between shipping on day one and shipping next month. A 262k native context on a 27B you can hold in unified memory is a real advantage. And Qwen Code being provider-neutral suits anyone who doesn't want a cloud key in the loop.

Both lose the same argument. "Open family" doesn't mean the best model is open — QwenLM/Qwen3.7 is a 404, and Mistral's OCR, Codestral and moderation models are closed. Both also move fast enough to hurt: Qwen shipped 3.5, then 3.6, then a closed 3.7 in roughly six months; Mistral went Large 3 to Small 4 to Medium 3.5 in about four. Bundle weights into a product and you inherit a permanent re-quantize-and-revalidate treadmill. I do, and it's the most under-discussed cost of building on open weights.

Who should pick which

Pick Qwen if you're building on a Mac, you need one family to cover several RAM budgets, you want a 27B-class model with a very long context you can hold in memory, or you want an open agent CLI that can point at a local endpoint.

Pick Mistral if your buyers care about EU jurisdiction, you need a small multimodal model under a clear commercial license, you want published per-token pricing and a stated no-training-on-API-data policy for the cloud half of a hybrid app, or you need speech and OCR models Qwen doesn't release openly.

Use both if you're doing this seriously. They're both Apache 2.0 — nothing stops you running Ministral 3 8B for a fast multimodal path and a Qwen 27B for heavy reasoning inside the same product. That's a licensing luxury you don't get everywhere in this market.

My own choice is on the record: Outlier runs Qwen3.5-4B, Qwen3.5-9B, Gemma-4-26B-a4b-it, Qwen3.6-27B, Qwen3.6-35B-A3B and Qwen3.5-397B-A17B — six Apache 2.0, one under Gemma's terms. That was driven by the size ladder and the MLX documentation more than by any benchmark. On our measurements the local 27B matched Claude Opus on 98.9% of rubric checks across a 54-prompt head-to-head and scored about 45% on a blind slice of SWE-bench Verified. I haven't run that harness against Ministral 3 or Mistral Small 4, so don't read those numbers as a verdict on Mistral. They aren't one.