One (of over a dozen) on Datacurve: deepswe.datacurve.ai/data/v1… The second is from Astra huggingface.co/buckets/evals… The latest ds4 flash does it too

DeepSeek did something scary. Look! Here's the precise traces and reproducability so you can see for yourself. github.com/huggingface/atif-… fast-agent.ai/benchmarks/run… huggingface.co/buckets/evals… (Check out the HF Trace viewer btw 😎).

Short example of discovering MCP Servers and Skills on Hugging Face and Vercel using the AI Catalog. Hosting resources at .well-known makes this easy for you - and your agents.

Bumped in to the GOAT Den Delimarsky with his keystone award. Well deserved!

Thrilled to see these new projects join the AAIF. Great additions.

Go @techgirl1908 at AGNTCon / MCPCon. Hugging Face mentioned!

Hard to explain what's happened with xAI Grok 4.7. 4.6 was competitive price/performance with 5.6 Sol - now it's 60x more expensive for similar performance. This is a subset of terminal-bench 4.0. Couple of notes 1 - The OAuth route was used, and it returns price info of 0.34x list price. I have used API pricing here. 2 - The tb-4 leaderboard Grok Build run doesn't price reasoning tokens; I've recalculated for this comparison.

PyTorch on TPU. Charts on the right for the inference margin speculators.

I ran DeepSeek 4.1 1017 against my terminal-bench 4 subset: 45.2% for ~$25.48. It's wild though: - One run it started a background process to monitor for verifier logs - Probed the Kubernetes infrastructure for task hints in 6 trials. Info and traces here: fast-agent.ai/benchmarks/run… I'd estimate a full tb-4 run would cost $163 off-peak, double for peak.

Shiny huggy day. Have a great time merve!