ICLR 2027 submission: we benchmarked all frontier models and our model performs... Anthropic & OpenAI today: no you don't. ICLR deadline is 9/25, so i guess no excuse not including opus5.5 and gpt6-luna & sol in the paper.

wow, Opus 5.5 looks very promising! On my CRPG benchmark, Opus 5.5 takes only 40 steps to get out of the spawn house and onto the world map, much faster than Fable5.1. Let's see if it can speedrun past gpt6-astra. 4-hour gameplay livestream starting now:

Model-specific inference engines are basically what everyone autoresearches and hill-climbs the hell out of for TPS after each Qwen release. In the early days I worked with MLX too, but these days I just trust oMLX to do the job.

yeah this actually looks really bad for zhipu/glm given the recent situation, so far no response from their official account neither Zibo @ZixuanLi_ or @louszbd respond, probably their first crisis PR.

thoughts on Typesafe/Jev: I'm surprised that a general-purpose classifier (or discriminative model) can be just as interesting to the public as a general-purpose generative model. I've worked on search and discriminative models for years, so maybe I should have called this upfront, tbf I never thought about making them general-purpose beyond search. Typesafe also positions Jev really well as "System 1", a complement to the generative models we already have. Humble, unassuming, and sidestepping the frontier-lab warzone. Jev's success could be disruptive to today's agentic systems in many ways, such as tool-calling and routing patterns. Over the last three years the consensus was to use a small generative model for routing, tool calling, MCP etc. Time to pause and rethink such architecture, because we'll probably see tool calling and routing move back to discriminative models. Whoever ships the next Jev-like open-weight base model will likely win the community.

Super excited about this new model! We don't ship a lot of generative models, so every time we release one I'm hyped. On M3 Ultra I managed to hit 600 tps aggregated at decoding with swift-metal, available in omni-macos:

not sure it’s even remotely related but back in 2024 I have this weird post about chain-of-thought classification using embedding model to achieve some kind of test time intelligence

people are sleeping on this release. I finally got a chance to test it today: qwen3.8-flash-next-q4 now runs at 50-70 tps on long-horizon tasks on M3 ultra, with no regression in quality or thinking. It used to be 20-30 tps. Very impressive work!

Every time I see a small LM or encoder-only initiatives, my first guess is that they must from Europe. honestly that's fine, it's way more credible than some eu AI companies claiming frontier this and frontier that on LLMs to get eu/gov funding while they're really just post-training Chinese open weights. The problem of SLM startups is scaling and fundraising: the SLM story is hard to sell in the EU because the VCs there lack the knowledge (or still dreaming about AI sovereignty), and also hard to sell in the US because VCs want to throw money at categories with huge upside and get that compounding effects. imo, the upside of SLMs is upperbounded to, say, dev tools? meaning the best SLM can probably be a pretty smart dev tools, and when compounding with some network effect it can be cool, but it's still nowhere near the LLM story and its upside.