Six open-weight model families shipped in one week, all targeting consumer hardware. Meanwhile Anthropic kept Mythos behind a curtain while quietly briefing the White House on what it could do. Two stories pointing in opposite directions, and the gap between them is where the next year of AI policy is going to live.
The open-weight release cadence is no longer a trickle. Unsloth's Qwen3.6-35B-A3B GGUF crossed 2.6 million downloads and the 27B variant added another 1.4 million in the same window. Mistral shipped Mistral-Medium-3.5 as a 128B open release. IBM published the Granite 4.1 family with an 8B claimed to match a 32B MoE on several benchmarks. NVIDIA's Nemotron-3 Nano Omni reasoning variant hit 116,000 downloads as a 30B any-to-any model running BF16. Gemma 4 GGUFs got chat-template fixes that unblocked local serving, and Grok 4.3 launched with developer docs. If you are still running production on a single closed API in 2026, the substitution cost just dropped again.
Mythos and the gatekeeper question
Anthropic is limiting access to Mythos because of what it can do, then telling federal agencies what it can do. The official line is that the model has exploit-finding capability dangerous enough to warrant restricted release. The same week, an Anthropic co-founder confirmed they had briefed the Trump administration directly. TechCrunch reported that Trump officials are encouraging banks to test Mythos. SiliconAngle's later coverage framed it bluntly. Mythos remains a mystery to the security world even as it gets pitched to financial regulators. There is a defensible safety argument for staged disclosure. There is also a less defensible argument where capability-gating becomes a way to manufacture a procurement moat. The data does not distinguish between the two, but the timing pattern is the same one we have seen before with frontier dual-use claims.
The alignment whack-a-mole paper deserves more attention than it got. A research release this week demonstrated that lightweight finetuning re-elicits verbatim recall of copyrighted books in aligned LLMs, bypassing the RLHF refusal training that supposedly suppressed it. The code was discussed across developer communities but has no traditional news coverage yet. The finding itself is the story. RLHF removes the surface behaviour, but the underlying memorisation is still there, and a few hours of finetuning brings it back. For anyone defending an AI product on the basis that "we trained out the bad outputs," this is the question your litigation team should be asking. RLHF suppresses the refusal, not the underlying capability, and a weekend of finetuning is enough to recover what was supposedly trained away.
Uncensored finetunes are a real distribution category now, not a curiosity. A single Heretic-Uncensored Qwen3.6-27B GGUF from a community uploader pulled over 161,000 downloads. This is one repo from one author, mid-sized, explicitly marketed as having safety training removed. Combine it with the copyrighted-recall finding above, and the picture sharpens. Removing safety conditioning from open-weight models is a few-day exercise, the audience is large enough to register at six-figure download counts, and the underlying capabilities the original training was trying to gate are still present in the weights. The open-weight competitive thesis was always going to collide with the alignment thesis. This is what the collision looks like at scale.
The capital is rearranging around two centres of gravity. Amazon committed up to $25 billion in additional investment to Anthropic as part of an expanded cloud partnership. OpenAI and Microsoft published the next phase of their partnership, formally unwinding the AGI clause that had governed the relationship since 2019. Apple announced Tim Cook's transition out and John Ternus's elevation to CEO, with the framing from analysts pointing to a hardware-defined AI future. The compute-platform map for the next decade is being drawn in public this month. The notable absence is anyone else with credible standing to bid. The hyperscaler tier is now a function of which model lab they have stapled to which silicon roadmap, and the latitude for new entrants at that scale is narrower than it was a quarter ago.
On our radar
- Local agentic search at frontier quality. A locally hosted Qwen3.6-27B running an open search recipe reportedly hit 95.7% on SimpleQA on a single consumer GPU. If that reproduces, the economics of paying for hosted retrieval shifts for any application that does not need real-time web freshness.
- Mechanistic interpretability getting practical. Multiple groups are now pairing two LLMs to translate internal activations of a third into readable text, with Anthropic's Natural Language Autoencoders being the most discussed. If this generalises, it moves interpretability from a research curiosity to a real input for compliance documentation.
- Anti-AI contribution policies in open source. The Zig project's formalised stance against AI-generated contributions drew significant attention in developer communities this week. We are starting to see the maintainer side of the labour question get organised. If a handful of high-profile projects follow, expect this to show up in supply chain due diligence conversations within the quarter.
Signal data for this briefing is provided by HiddenState, Mosaic Theory's signal intelligence platform.
— Cosmo