October 27 came and the weights didn't. Reuters had pegged the date; Mistral's own October 6 announcement promised "Weights drop end of this month"; the Hugging Face card for Large 4 now shows a new countdown reading October 31 at 15:00Z. Forkast's first fetch at 14:46Z read "Upcoming release" with 969 people on the wait list. A second check at 23:43Z put the wait list at 973, up from 813 on October 14, and the LICENSE path still returned HTTP 404. The raw file returned 401. "The timer said October 27. The fetch says 404."

What's late is a 1.05 trillion-parameter mixture-of-experts, 52 billion active per token (49 billion excluding embedding and output layers, per Mistral's upcoming-release listing), with a 1.6-billion-parameter vision encoder bolted on. VentureBeat framed the preceding three-week preview as reinforcement-learning time to tune the final checkpoint, which means the weights that eventually land may not be the weights the preview advertised. That's a meaningful footnote for anyone benchmarking against the Le Chonk preview that scored 59.9% on AutomationBench at $0.68/M.

Then there's the hardware floor. AI Cybersecurity Review's breakdown puts raw weight storage at 2.10 TB in BF16/FP16, 1.05 TB at FP8/INT8, and roughly 525 GB for an idealized 4-bit checkpoint before KV cache, activations, routing tensors, or the vision encoder. An eight-GPU H200 node offers 1,128 GB of HBM, leaving 78 GB of nominal headroom at FP8 before caches. An eight-B200 node offers 1,440 GB. Eight large GPUs is the floor, not the ceiling.

And the license is the quieter problem. Mistral's existing Modified MIT terms gate commercial use at a $20M revenue line. Until the LICENSE file resolves to something other than 404, no small business can legally plan a deployment, because the terms governing it don't yet exist as a readable document. Compare Reflection AI's Beam 1.0, which shipped 501B open weights under Apache 2.0 with Baseten already listing it. One is deployable today. The other is a countdown and a 404.

For a founder-led business weighing open-weight AI as the affordable path to customer acquisition, the stack is: wait for the weights, wait for the license, source eight H200s or B200s, and absorb the operational overhead before the model does any sales or marketing work. LemonLime sits on the other side of that choice. It studies the company, industry, and competitors and delivers finished sales and marketing work by 9 AM, with no model selection, no HBM arithmetic, and no license file to refresh. Get started.

Open weights are a release artifact. Readable license terms are the actual product.

Sources

https://forkast.news/the-reported-date-arrived-and-the-weights-did-not/
https://forkast.news/the-fetch-is-the-product-96-hours-left-on-mistrals-clock/
https://venturebeat.com/technology/mistral-debuts-large-4-le-chonk-a-1-trillion-parameter-text-output-model-with-high-benchmarks-planned-for-open-weights-release
https://aicybr.com/blog/mistral-large-4-moe-api-pricing-gpu-memory-weights-2026
https://betanews.com/article/mistral-large-4-preview-open-weights/