Moritz

Extremely positive results for @Zai_org's GLM-5.3 in @sam_paech's EQ-Bench v4 benchmark: It appears to be the best model by a wide margin, not just compared to the previous open-weights champion K3, but also compared to the best commercial models such as Fable. It's personality is the "direct analyst", not Qwen-3.8 the "harmony-seeking advisor", and also not V4F-0731, the "people pleaser". However, with an unheard-of analytical score of 9.1, GLM-5.3 may be almost too machine-like. More testing will show.

Sleeper Agent and How to Tame Them

Prototyping vision for Nemotron 3.5 Lightning Big thanks to @NVIDIAAI for an early-access version of 3.5 Lightning for tinkering. We followed the great example set by @part_harry_ from @baseten in their GLM-5.2 with vision, which demonstrated how a frozen Vision Tower from Kimi K2.6 could be injected into an LLM with little training effort. We adapted this approach to 3.5 Lightning, using one node of 8 x RTX 6000 Pro GPUs for training. We found this quite practical and Nemotron 3.5 Lightning to be very adaptable. This bodes well for all kinds of experiments with the new model. huggingface.co/blog/tngtech/… huggingface.co/blog/tngtech/…

Real-world robotics problems, #1.

We benchmarked the new @deepseek_ai V4 Flash 0731 with our TL;DR-bench via @OpenRouter. The model scores very well for its model size. In our benchmarks, it's significantly stronger than the older DeepSeek V4 Flash version that we have deployed, and comparable to GLM 5.2 with HIGH reasoning settings.

Our OpenArm robot arms in action. With only 100 demonstrations collected via our VR teleoperation setup, we fine-tuned a pi0.5 robot foundation model for this task. The teleoperation and inference pipeline is implemented in @rosorg ROS 2 which makes it easily transferable to other robots in our growing robot family.

New walking policy for our walker.

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

Happy 25th birthday, TNG - coding since January 1st 2001!

Yesterday, the TNG's Chimera family of 671B child models, constructed from @deepseek_ai-parent LLMs (谢谢), processed in excess of 10 billion tokens/day on @openrouter for the first time. It is running on @chutes_ai backends.

Zen Driving Études after a long day of robotics work.

DeepSeek-TNG-R1T2-Chimera is currently the #1 trending text-generation model on @huggingface It got added to @openrouter a couple of hours ago, available for everyone to use. Thanks to @pingToven at @openrouter for adding the model and to @chutes_ai for providing the @nvidia H200s nodes.

Want to try out R1T2 Chimera? Here is the direct link to chutes: chutes.ai/app/chute/4fa0c7f5… Despite heavy attacks against its platform, @chutes_ai stabilised the deployment. Thx @jon_durbin

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.

Self-reflecting its own birthday...

Today we release DeepSeek-R1T-Chimera, an open weights model adding R1 reasoning to @deepseek_ai V3-0324 with a novel construction method. In benchmarks, it appears to be as smart as R1 but much faster, using 40% fewer output tokens. The Chimera is a child LLM, using V3s shared experts augmented with a custom merge of R1s and V3s routed experts. It is not a finetune or distillation, but constructed from neural network parts of both parent MoE models. A bit surprisingly, we did not detect defects of the hybrid child model. Instead, its reasoning and thinking processes appear to be more compact and orderly than the sometimes very long and wandering thoughts of the R1 parent model. Model weights are on @huggingface, just a little late for #ICLR2025. Kudos to @deepseek_ai for V3 and R1!

At TNG, we handle 5,000+ #LLM requests per hour and generate 10+ million tokens every day. Learn how our team optimizes inference serving for low-latency responses in high-traffic environments in the second article of our series on LLM performance: huggingface.co/blog/tngtech/…

Load testing @nvidia RTX 5090 💯% GPU-utilization 87% vRAM-usage Burning 550 watts while running @Google gemma-3-27B at 105 tokens/s total throughput with @ollama Waiting for @vllm_project to ship a @Docker image compatible with the RTX 5090