This week's AI race took an unusual turn: instead of bigger closed models locked behind APIs, the action was in models you can download and run yourself. Europe's Mistral shipped its biggest model ever — a trillion-parameter beast nicknamed "Le Chonk." A Brooklyn startup called Reflection answered with Beam, a model that does more with far less compute. Anthropic slashed its small model's price by 90%. And China's DeepSeek, the team everyone's chasing, is reportedly raising a ~$12 billion war chest. Here's what changed, in plain language.
Unveiled on October 6, Mistral Large 4 is the French lab's biggest release ever: ~1 trillion total parameters with about 49 billion active (a mixture-of-experts design, so only a slice fires per token — keeping serving costs sane), multimodal, and built for reasoning. Training took roughly 3,800 Grace Blackwell GPUs — about 52 full NVL72 racks — in Mistral's own European datacenters, across a corpus spanning 160+ languages, using supervised pretraining plus reinforcement learning.
The positioning is blunt: Mistral calls it "state of the art on critical workloads" — cyber defence, manufacturing, finance — and the strongest open-weights model from the US or Europe yet. CEO Arthur Mensch says it's above the Chinese models on certain aspects, including cyber. Independent testing by Artificial Analysis is cooler: on raw intelligence, Le Chonk sits between DeepSeek V4.1 Flash and OpenAI's entry-level GPT-6 Luna. But among non-Chinese downloadable models, it beats Thinking Machines Lab's Inkling — the most capable US open model — by a significant margin.
The model is in preview now (available via Mistral's API), with weights expected on Hugging Face later this month. And it's only the start: Le Chonk is the first of a series funded by Mistral's €3 billion Series D last month, and the company says continued reinforcement learning should make it meaningfully stronger — a 4.1 may not be far off.
Sources: The Register, CNBC
Brooklyn-based Reflection AI — an Nvidia-backed startup founded by former Google DeepMind researchers — announced its first open model this week: Beam, a sparse MoE with 501 billion total parameters and just 23 billion active, built for coding, reasoning, and agentic workloads.
The headline isn't the size; it's the training discipline. Reflection pretrained on 23.8 trillion tokens, then ran what it calls one of the largest reinforcement-learning runs of any open lab: 100 million+ rollouts on 10,500 Nvidia GB300 GPUs over four weeks, graded across 1.3 billion sandboxes. The scores: 80.9% on SWE-Bench Verified and 77.2% on SWE Bench Pro v2-Hard. Reflection says Beam matches GLM-5.2-class reasoning while using 3–4× less inference compute — "more intelligence per token," in their phrase. The weights, technical report, and model card land later this month under Apache 2.0.
Source: Reflection AI
Anthropic launched Claude Haiku 5.5 on October 7 with a headline-grabbing cut: for prompts up to 100,000 tokens, pricing falls from $1/$5 per million input/output tokens to $0.10/$0.50 — a 90% reduction. Above 100K tokens the cut is 50% ($0.50/$2.50), and Anthropic says the model costs roughly 75% less to run on average across workloads (about 90% of the old Haiku's requests fell in the cheaper band). Context doubles to 1 million tokens, output to 128K, and — a first for the Haiku line — a five-level adjustable effort setting trades thinking depth against cost.
The interesting part is that this isn't a bargain-bin cut. On Anthropic's own benchmarks, Haiku 5.5 leaps from 0.0% to 39.2% on Terminal-Bench 4.0 and from 15.7% to 72.4% on OSWorld 2.1 (computer use). The small model can now run an agent loop without being the weak link.
Sources: TechGig, DEV Community
Meanwhile, the team everyone is chasing is raising money at a pace that dwarfs the West. Bloomberg and Reuters reported on October 6 that China's DeepSeek is close to securing at least 80 billion yuan (~$12 billion) in fresh funding, possibly approaching 100 billion yuan — blown past its original 50-billion-yuan target. Tencent and battery giant CATL are among the largest backers. The round is expected to close in October, and DeepSeek is preparing for a possible domestic IPO in early 2027 (CITIC Securities is reportedly advising).
This would be its second raise of the year, after a ~$7.4 billion maiden external round in June. DeepSeek has also partnered with Huawei on programming tools optimised for Huawei's Ascend AI chips — a push to cut reliance on Nvidia's ecosystem — and released DeepSeek-V4.1-Flash last month.
Source: Tech Startups (Bloomberg/Reuters reporting)
Open weights have become the counter-model to rented intelligence: download it, run it yourself, keep your data in-house. Two serious Western drops in a single week, a $12 billion raise into the Chinese champion, and a 90% price cut on the budget model all point the same way — the race is no longer just about the smartest model. It's about the best model you can actually own.