Mistral's next open-weight MoE hits early access — and the specs I actually care about are missing
Mistral's new open-weight MoE hits July early access, but the specs that matter for local use — params, VRAM, quant support — are still undisclosed.
Mistral is putting a new open-weight Mixture-of-Experts model into early access this month with research, government, and industry partners. That's the headline. What's missing is everything a person who runs models locally needs to know: parameter count, VRAM footprint, and quant support. Until those land, "open-weight" is a promise, not a checkpoint I can pull.
Arthur Mensch called it "very exciting" and confirmed a €4B data-center buildout behind it, with a broader release later this summer, according to the July early-access disclosure. Cool. None of that tells me whether it fits on a 24GB card.
What does "fat but sparse" actually mean for local?
Sparse MoE means total parameters and active parameters are two different numbers, and vendors love to quote the flattering one. A model can be 200B+ total but only fire 20-30B per token. That's great for throughput on a cluster. It's rough for local, because you still need enough memory to hold all the expert weights resident, even if only a fraction activate per forward pass.
So the number I want isn't "active params." It's total weight size on disk, and how it maps to VRAM once quantized. A sparse 8x22B-style layout can eat more memory than a dense 70B while feeling faster. People get burned by this constantly.
What I want to see on day one
Three things, in order:
- **Total vs active param split.** Say both numbers plainly. If the release only quotes active params, assume the total is uncomfortable.
- **Quant support at launch.** GGUF and a working llama.cpp path, plus GPTQ/AWQ for the GPU crowd. MoE quantization is still fiddly — expert routing plus low-bit weights can tank quality in ways a dense model wouldn't. I want to see day-one Q4/Q5 numbers, not "community quants coming soon."
- **A license I can read.** "Open-weight" from Mistral has historically meant a mix of Apache-2.0 and their own research/commercial split. Which bucket this one lands in decides whether it's usable or just viewable.
Will it run local before the hype?
Probably not on consumer hardware at full precision. If the total footprint is in the hundreds of billions, you're looking at multi-GPU or heavy quantization to get it onto anything under a workstation. The honest answer is: local-runnable depends entirely on numbers Mistral hasn't published yet.
The €4B data-center detail tells you where the target audience is — partners with racks, not a 4090 in a closet. That doesn't kill local use, but it sets expectations. Expect the first genuinely local-friendly variant to be a smaller sibling, not this flagship.
Verdict
I'm interested, not sold. Mistral has a real track record of shipping weights that the local community actually uses, so the intent is credible. But an early-access announcement with undisclosed params is a press beat, not a release. I'll judge it the day I can download a GGUF, run it on a single card, and check whether the MoE routing survives 4-bit. Until then it's a placeholder. Ping me when there's a file to pull.