Open Weights, Late October: Mistral's Le Chonk and the Race to Download
The open-weight wall kept cracking this month. Mistral's Large 4 — a trillion-total-parameter multimodal model with 49 billion active and a million-token context — has been live in API preview since October 6, and the downloadable weights are scheduled to drop around October 27 after red-teaming. At launch pricing of $0.68 in / $2.09 out per million tokens, it undercuts most closed rivals on cost.
Reflection AI's Beam, a 501-billion-parameter sparse mixture-of-experts model with a million-token context aimed at coding and reasoning, is expected to open its weights later this month under an Apache 2.0 license. Meta's Muse Wonder, released October 12, sits at $4 in / $20 out and is reportedly near the frontier, though its data defaults — it trains on user submissions by default — deserve a careful read before anyone builds on it.
The pattern is consistent: the closed labs keep the absolute frontier, but the 'good enough and free to self-host' layer is filling in fast. For teams that can't send data to someone else's API, that layer is now where real production work lives, and this month made the menu a lot bigger.
The catch hasn't changed: open weights mean you own the ops — GPUs, serving, guardrails, and the blame when the model says something stupid. Self-host the one workload where data leaving the building is a non-starter, and rent the rest. The hybrid is cheaper either way, and October made it practical.
I'm not tearing down my API subscriptions. I'm moving the sensitive workload to hardware I control, and this month's releases finally made that a real option instead of a research curiosity.