Qwen3.8-27B with EXL3 on MiMo RL environments: BF16 = 63.2% 2 bpw = ~55.5% (v1.4.2) 1.8 bpw = 2% (v1.4.2) 1.8 bpw = 13.2% (v1.6.4) Roughly...

K3.1 may be as big of a release as K3 I really enjoyed using K3 before Opus 5.5 was released https://twitter.com/kimi_moonshot/status/21134...

I keep wondering why Qwen4, Kimi K3.1, GLM 5.5, and other models are delayed while DeepSeek is shipping Is it just distillation? or are the...

If you could run DeepSeek v4.1 Flash 1017 on a single RTX 3090 + 64GB RAM, would you? This should be possible by ternarizing the model, but...

I haven't properly understood why But something is wrong with most providers serving Kimi K3. I feel quality degradation when using it. Th...

Normal Computing is part of "The World’s Most Groundbreaking AI Startups" Come join us if you are excited about AI x Hardware! https://x.co...

This is a huge unlock for RL Traditionally, it’s a pain to get the right level of reserved compute of the same type in one cluster I hope ...

it's called user experience and moonshot ai perfectly said it in their kimi k3 release. it's the feel of interacting with the model, e.g. h...

I tried DeepSeek's harness desktop app this week. The model side is promising, but the app itself still feels a step behind Cursor's Agents ...

thousands of environments are cooked... this is why open-source should be the default. when you open something up, it lets the world audit ...

Microsoft is launching Surface Spark with Nvidia Honestly an odd form factor and it's $6k for 128GB

Red flags are going off all over the place recently. The speed-maxing obsession has to come to an end. Quality is what defines a model. Th...

Opus 5.5 is the first time I feel comfortable turning down the reasoning knob from Max to High. The acceleration from less reasoning while b...

Inference is more like 80% total if you count inference during post-training https://twitter.com/socialcapital/status/2107239666781208693

you can build something like Bonsai 2 for ~$43k. you too could be like @PrismML, you just need one on-demand B300 node for ~4 weeks. i wrot...

Kimi K2.5-2.7 can be trained with RL on just 16 B300 GPUs (yes, just 2 nodes!) with full context Most post-training frameworks out there wi...

would love to see a Qwen4 27B with 100B+ Engram that offloads to RAM + NVMe runs on small VRAM but has the world knowledge of a big MoE lo...

everyone's complaining karpathy went to anthropic, meanwhile opus 5.5 is one of a kind tips and tricks posts are nice but i want claude to ...

claudeish is absolutely diabolical. happy to see it gone and outperform fable 5.1 at the same time. the question is: will it feel like fabl...

NEVER thought i would hear this from jeff dean he thinks we can compress chip design from 2 years to 3 months with RL + new EDA tooling es...