Mistral's new reasoning models add vision and run locally on Macs

Running a capable reasoning model with vision on your local laptop used to sound like wishful thinking, unless you were willing to settle for a toy-sized LLM or rent an expensive cloud rig.

Mistral just changed the math on that. The French AI lab rolled out Magistral Small 1.2 and Magistral Medium 1.2, framing them on X as mere "minor" updates to its reasoning line.

Don't let the modest label fool you. Both models now feature a dedicated vision encoder, giving them the ability to inspect and analyze user-submitted images alongside their multi-step reasoning capabilities.

A 24B multimodal thinker in your backpack

The real showstopper for developers is Magistral Small 1.2, which packs 24 billion parameters.

As Hugging Face ML growth lead Ahsen Khaliq noted on X, a quantized version of Small 1.2 fits neatly inside a single Nvidia RTX 4090 GPU or an Apple MacBook with 32GB of RAM. That means you can run a private, multimodal reasoning engine completely offline.

Khaliq was impressed enough to update his vibe coding web tool, AnyCoder, making Magistral Medium 1.2 the default model driving the app.

If you want to kick the tires, both models are already live on Mistral's Le Chat web interface, with weights available on Hugging Face. Developers can also tap the Mistral API using the identifiers magistral-small-2509 and magistral-medium-2509.

Pricing sits at $0.50 per million input tokens and $1.50 per million output tokens for Small, while Medium costs $2 input and $5 output. That makes Small substantially cheaper than mid-tier cloud options like Anthropic Sonnet, though pure input-heavy workloads still skew cheaper on DeepSeek.

Why it matters

We think Mistral pulled off a genuinely clever move here by packaging step-by-step reasoning and image understanding into a footprint that developers can actually run on local iron.

For builders and enterprise teams sitting on proprietary visual data, like blueprints, medical scans, or internal dashboards, this is worth testing immediately. You get complex chain-of-thought analysis without pinging an external API.

Across the industry, the race has largely focused on massive, cloud-bound reasoning giants like OpenAI's o-series or DeepSeek R1. Mistral, however, is doubling down on accessibility. By offering Small under an Apache 2.0 license while keeping Medium proprietary for enterprise sales, it keeps open-source loyalists happy without abandoning a commercial monetization path.

The trade-off, of course, is quantization loss. Shrinking a 24B multimodal model down to a 32GB laptop will inevitably introduce some edge-case hallucinations or precision drops in math-heavy tasks.

Even so, having a locally runnable multimodal reasoner sets a new practical benchmark for what edge AI can realistically handle.