Today we release DeepSeek-TNG R1T2 Chimera. This new Chimera is a Tri-Mind Assembly-of-Experts model with three parents, namely R1-0528, R1 and V3-0324. R1T2 operates at a sweet spot in intelligence vs. output token length. It appears to be... * about 20% faster than R1, and more than twice as fast as R1-0528 * significantly more intelligent than R1 in benchmarks such as GPQA Diamond and AIME-24/25, albeit not quite on R1-0528 level * much more intelligent than our first R1T Chimera, and also think-token consistent, which is a major improvement We perceive it as generally well-behaved and a nice persona to talk to. The weights are on @huggingface under the MIT licence. We are looking forward to your experiments and feedback! Thanks to @deepseek_ai for giving their models to the world, to @chutes_ai and @openrouter for hosting R1T, to @WolframRvnwlf for benchmarking it, to @xlr8harder for beta-testing the new Chimera, and to @natolambert for constructive discussions at @aiDotEngineer.

We are thrilled to announce that @natolambert is joining Arcee as a Research Advisor. Nathan’s work and thought leadership have been instrumental to the open model ecosystem, and his guidance comes at a critical time as open builders face growing pressure. This is a major addition for Arcee and the American OS movement. Nathan brings the conviction, taste, and technical depth this moment calls for.

Heard in the market: "Our Claude costs are greater than zero."

Xeophon on the art of benchmarking

Our robots are in the metal hospital right now.

GPU acquisition prioritization, your opinion:

New walking policy for our walker.

0xsero.github.io/blackwell-g… <—— I address all this (lots of slop but the data comes from months of digging)

Experiments using @Zai_org 's GLM-5.1 on Hopper and Blackwell, using @opencode with 50k input tokens and 500 output tokens on average. The RTX 6000 Pro x 8 node seems not to scale well with concurrency, while the B200 x 8 is just very strong. Are these measurement errors or reality? With luck and new kernels, how much could the RTX be improved?

Registration for #MuniHac 2026 is open! munihac.de/2026.html#registr… #MuniHac will take place October 9–11, 2026 at the @tngtech office in Munich. Keynote by @TacticalGrace confirmed, more to come. Want to contribute? Check out our CfC munihac.de/2026.html#cfc! See you in Munich!

Here a very dry observation by GPT-5.4 from right now about using rather expensive inference hardware: "If the software is bad, even an NVL72 can be operated astonishingly inefficiently." "Wenn die Software schlecht ist, kann man eine NVL72 auch erstaunlich ineffizient betreiben." 8)

BAR - branch adapt route. Congrats and thank you, AI2, for this interesting approach. We've performed similar experiments, namely transplanting experts from a new LLM version (DS-V3-0324) to an earlier base mode (DS-V3). Domain-specific performance, such as on subsets of SimpleQA, can be improved selectively. Implanting more and more experts, e.g. up to 1% of total experts, steadily enhances benchmark performance. The selection method which experts to transplant is crucial.

Last year, we introduced FlexOlmo, a novel way to train parts of a model independently then combine them later. BAR builds on that idea for a harder problem: how to keep improving a model without having to retrain each time. 🧵

Q: Why is @opencode nowhere to be seen on @OpenRouter? Is it just not being shown or truly not present on the plattform? Turns out: it seems to be delisted on OpenRouter on request. Looking at github stars does tell a different story: - 144k @opencode - 98k @NousResearch Hermes (congrats!) - 23k @roocode

Happy Easter, @natolambert @latkins @dmsobol @0xSero @teortaxesTex @willccbb @xeophon @kalomaze @xlr8harder @jon_durbin @airesearch12 @danielhanchen @polynoamial @reach_vb @_xjdr !

Introducing GLM-5-Turbo: A high-speed variant of GLM-5, excellent in agent-driven environments such as OpenClaw. Coding Plan Max: z.ai/subscribe OpenRouter: openrouter.ai/z-ai/glm-5-tur… API: docs.z.ai/guides/llm/glm-5-t…

Preliminary tests of Weight Offloading V2 of @vllm_project v0.17.0 with @Zai_org's GLM4.7-FP8 on RTX Pro: Median TTFT: without offloading 16.8s, with offloading 32.3s, x 2 Median inter-token latency: without offloading 27ms, with offloading 805ms, x 30 (very slow!) 50,000 input, 500 output tokens It required a vLLM pull request (37178) to fix weight-prefetch. Alternative measurements, e.g. on B200, corrections and/or feedback mucho appreciado.

Munich today in the sunshine: Taking the @UnitreeRobotics dog for a walk.

Greetings @_xjdr : We did some preliminary tests with your Noumena nmoe trainer - thanks for all the work & code! On our 8xB200 systems, we were not able to get significantly different results than from regular Megatron. Is that plausible or wrong? Any ideas how to tweak it?

Great feature! And badly needed. Because: @opencode is not in the list yet? How is that possible... Would be great if you made apps a first-class citizen on @openrouter. Currently, apps are listed per model only, and the ranking page details the top 20: openrouter.ai/rankings#apps Would love to be able to search for apps on openrouter