How DeepSeek-R1 thinks: Activation patterns flowing through its Mixture-of-Experts layers while processing the token stream "What happened last year at Berlin Wall?". The Matrix...

DeepSeek uploaded a new model on huggingface: DeepSeek-Prover-V2 It seems, the architecture is identical to V3 and R1 models, because: model_config.py shows no difference, also the safetensor index files are the same. One minor diff is a new experimental feature in

this is kind of absurd, and totally flew under my radar you can selectively blacklist ~0.07% of DeepSeek r1's experts which are most associated with refusing this makes the model... better, according to Mt-Bench, most prominently for coding (tho fyi, mt-Bench is a bit dated)

8x256 GB!

Having a blast in beautiful Estes Park, CO. It was a real pleasure sharing my perspective on how LLMs are transforming software engineering, and why running your own GPU stack can be a game-changer. Thanks to everyone who came with smart questions and big ideas! #LambdaConf2025

x.com/i/article/1928…

More evidence for the effectiveness of the Chimera construction method: Taking DeepSeek's R1-0528 release, we started benchmarking new Chimera variants on AIME-24 and SimpleQA. R1-0528 significantly improves AIME performance from 79.8 to 91.4 while doubling the amount of output

Assembly of Experts: Our linear-time 671B Chimera LLM construction paper should soon appear on arxiv.org. We are at the AI Engineer fair in SFO until tomorrow, so for a chat -> DM ;-)

We post our new paper "Assembly of Experts: Linear-time construction of the Chimera LLM variants with emergent and adaptable behaviors" on Hugging Face, while waiting for arXiv.org. We explain how we constructed the 671B R1T Chimera child model from the great DeepSeek V3-0324

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

R1T2 Chimera is available on Chutes Thanks to Jon Durbin and team for making it available within 24 hours!

DeepSeek-TNG-R1T2-Chimera is currently the #1 trending model on OpenRouter The platform which Andrej Karpathy called the "transfer switch of AI". Our Assembly-of-Experts method pushes the Pareto frontier between model intelligence and inference cost. Thanks again to the Open

Fun screenshot. Let's hand over the #1 trending position to #Grok4.

Elon Musk R1T2 Chimera #3 trending model today Right behind Grōk and Kimi K2

FAZ article: "Der Boxkampf der KI-Tiger" with assessments from TNG A recent article in the Frankfurter Allgemeine covers the World Artificial Intelligence Conference in Shanghai, providing insights into the new and significant Chinese open source #AI models such as Qwen3, K2, and DeepSeek. TNG

News from the Aider discord regarding DeepSeek-TNG R1T2 Chimera's performance in the Aider Polyglot benchmark, courtesy of benchmark wizard neolithic5452 and the magic Unsloth AI quantizations: - 2 bit UD-IQ2_M: 60.0% - 4 bit Q4_K_XL: 62.7% - 8 bit: 64.4% This seems to be the