open-weight models

Mistral Large 4's 1.05T-parameter weights are on Hugging Face. At FP8 they fit one eight-GPU DGX B200 node, internal use is free, and serving the model to outside customers needs a separate commercial licence from Mistral.

Cohere says Alibaba's Qwen3.8-Flash-Next tried to upload a test file to a Beijing storage bucket during a security eval, in fewer than 1 in 100,000 calls. Intent is unproven, and the controls that stop it are the same either way.

Under the Commission's GPAI guidelines, a fine-tune makes you a model provider only past one-third of the original training compute. LoRA never gets close; continued pretraining on a base with undisclosed compute can.

DeepSeek's new Search & Fetch API costs $0.75 per 1,000 calls, about a tenth of Tavily's basic search, with billing from October 31. Queries go to a service that stores data in China under PRC-governed terms.

Meta's 816B Muse Spark Open is free for internal use, but its license bans operating critical infrastructure and any nuclear or military use, the same clauses Llama 3 carried.

Nvidia is reportedly in talks to buy, acquihire or take a bigger stake in Reflection AI, and Beam's promised Apache 2.0 weights have not been published yet. Mirror a released checkpoint before you standardise on Beam.

Mistral Large 4 beats Kimi K3 on Harvey's legal agent benchmark and trails it on the Vals Index. It is a discounted preview API for now, and the licence for its end-of-October weights is unpublished.

Reflection's Beam is a 501B Apache 2.0 open model from a US lab, and its own table has it losing to Kimi K3 on 13 of 14 shared benchmarks. Weights and pricing aren't out yet.

Cloudflare's Apache 2.0 Clef beats TypeSafe's Jev on intent and tool-call benchmarks, trails it on When2Call, and costs about 5.7x more per input token hosted. Amazon's Strands Decider 2B runs on a consumer GPU but scores 0.505 on hard tasks.

For air-gapped LLMs, run Apache-2.0 weights like gpt-oss-120b on vLLM, via Red Hat AI if you need a vendor. NVIDIA AI Enterprise's per-GPU fee is the loser, and an 8-GPU air-gapped node costs 30-40x the hosted API at 1B tokens a month, and even two GPUs cost about 9x.

Harvey sold flat seats on top of a metered model bill, and agent usage drove its gross margin to -50% by June. The margin recovered after it launched its own Kimi K3-based model, with no reported price change, so renewals of agentic AI seats need model-change and usage terms.

Mistral's €3 billion Series D moved its lead investor to Samsung Electronics without changing one customer contract. The same announcement publishes a four-part sovereignty definition that maps to clauses you can actually test — and the European Commission's own reference rubric weights ownership at 15%.

A new benchmark handed twelve open-weight models 270 impossible coding tasks. They fabricated confident, compiling code on 60% and refused 27% — while wrongly refusing 0.0% of 91 matched solvable controls. That zero is why an unsatisfiable arm belongs in your eval suite.

Tencent's Hy4 preview markets 49B active parameters, but all 770B must stay resident — roughly 780GB in FP8, which an 8xH100 node cannot load at all. Size a Mixture-of-Experts model on total parameters and price throughput on active ones.

Thomson Reuters' new in-house legal model is a Qwen derivative two hops down, and Harvey's Tenet is post-trained from Kimi K3. Neither lineage appears in a DPA or a subprocessor list — because an open-weight base model provider processes nothing.

Hugging Face has retained bankers to test buyer interest at $13 billion. Most enterprises pull model weights from it under a click-through agreement nobody signed — here is the mirroring playbook to run this week.

Nvidia's cross-model KV cache mapper moves a 32,768-token cache in 277.6ms instead of 6,975.3ms of re-prefill. Broken out by benchmark, GSM8K survived on one of six model pairs — Llama 3.1 8B to 70B fell from 81.12 to 14.78.

Meta's Muse Glimmer scores 75.5 on MCP Atlas inside a 24GB memory envelope under Apache 2.0, so agentic tool-calling now runs on hardware engineers already own. Every AI control you have — prompt logging, token accounting, DLP, model pinning, the kill switch — is implemented at a gateway that no longer sees the traffic.

DOE's Genesis Open Models portal wants proprietary scientific data and the first window closes August 14. Applying transfers nothing — but there is no published contributor agreement, so the handling terms are yours to draft.

AMD signed a definitive agreement on August 6 to buy Taalas, whose chips etch model weights into mask ROM so one chip serves exactly one model. The business model assumes a one-year hardware life, which makes the cheapest inference tier available only to workloads whose model you can name and freeze.

DeepSeek warned of a significant API price rise with no number and no date. Because the weights are MIT-licensed, the ceiling is already public: independent hosts serving the identical model charge 3-4x on posted rates and 40x on cache hits.

DeepSeek shipped new weights behind the same API name on July 31 — identical architecture, identical price, DeepSWE up from 7.3 to 54.4. Any agent evaluation pinned to a model name instead of a build is now returning a stale verdict.