So much folklore of injecting text like checking docs or extra think for models to improve performance. Folks, just say you are forcing a certain distribution of behavior, it ain't that deep

Bytedance has allegedly surpassed 1 million GPUs of H100/H800 sku, putting them in the same league as Google (1-1.5M H100 level) and Microsoft (750k - 900k). Combining megascale, Verl, and joint efforts with AReaL, it is some incredibly exciting times ahead

With releases like MiMo it's clear the future for production AI systems will be a strong and small reasoning model paired with a strong retrieval and knowledge recommendation system.

We used RL to train a model for MCP! Connect any MCP client to any MCP server - you can run MCP workflows fully with local models (+ tune it further). It works with Ollama / any MCP client that supports Qwen3 models - download it below 👇1/

The path to ASI already exist in the data we just gotta figure out how to sample ourselves to it

For people that have tried post training the MiMo RL models, have anyone noticed that this model is *incredibly* verbose? Even for simple tasks it will go on for 20k+ tokens before an answer.

Many frameworks have claimed to support RL for LLMs. Only one that would actually work in production and even it lacks support on standardizing things like tool invocation and reward policies. It's like selling you a fully functional car with no gas and the nearest gas station is

Did Anthropic just make pass@n "parallel time-time compute" or is it different

Mythical Waymo pull

Don't use structured output mode for reasoning tasks. We’re open sourcing Osmosis-Structure-0.6B: an extremely small model that can turn any unstructured data into any format (e.g. JSON schema). Use it with any model - download and blog below!

ollama run Osmosis/Osmosis-Structure-0.6B

Am I a boomer in the ai agent world... I don't care about your fancy abstractions show me the jinja template and your tools k? Thx

Nvidia releases Python DSL for CuTe and within weeks there's already a 200+ WeChat group chat about this discussing use cases and findings. Recommending a blog post here: veitner.bearblog.dev/bridging-math-…

It’s easy to fine-tune small models w/ RL to outperform foundation models on vertical tasks. We’re open sourcing Osmosis-Apply-1.7B: a small model that merges code (similar to Cursor’s instant apply) better than foundation models. Links to download and try out the model below!

Yesterday, we had a chance to talk about RL and what we're building at Osmosis (YC W25) at the Vercel HQ! Thanks for the invite Guillermo Rauch!

Anyone coding with OSS AI IDE's should have at least a perplexity MCP and a language server MCP for linting/type checking etc