Mercury 3 is live: the most capable diffusion LLM released to date. It outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running at over 1,000 tokens per second. More than 40x the size of Mercury 1, 3x more intelligent across standard benchmarks, and just as fast, with a 1M-token context window. Available today on the Inception API, @baseten, and @OpenRouter .

Scaling past a trillion parameters requires inference engineers to solve 3 issues: - Width: Routing between hundreds of sparse experts - Depth: Mixing global and local linear attention - Length: Processing 1M-token sequences I asked Ali and Bryce to draw it up. They delivered.

AI-generated fake documents grew 500% last year, and humans can no longer tell the difference That's why we’re launching Extend Verify, the first document fraud detection API built for agents Verify catches the fastest growing types of document fraud, including: - fully AI-generated documents - localized AI edits (down to a single modified character) - patched text, overlays, and in-painting from Photoshop, PDF editors, and coding agents - digital tampering after a document was signed - and many more In our internal evals, Verify detected over 96% of AI-generated and tampered documents. Verify is available today: extend.ai/verify/request-acc…

public benchmarks rarely tell you how a model does in your repo and the full story. so how do you decide which model is optimal for each engineering task? next week, molly and I will walk through the framework we use with development teams to evaluate open models for coding agents, powering the fastest growing ai applications in the world. we'll cover which models to shortlist, how to run a successful pilot without changing your harness, and how to measure cost per successful task on your own merged PRs. we hope you can join us! baseten.co/resources/webinar…

Fast is better than slow. We've got two new PhoneLLM models optimized for conversational voice AI. PhoneLLM Mini (30B) — When you need the lowest possible latency (sub-100ms TTFAT), smallest model size, and lowest cost. Good multi-turn conversation behavior and tool-calling accuracy in a model that can run >100 concurrent agents on a B200 and fits on an RTX 5090. PhoneLLM Max (120B) — A "microthinking" model for complex conversational tasks. Sub-200ms TTFAT. SOTA task completion performance at 1/10th the latency and cost of models with similar capabilities. Both of these models are full-weights fine-tunes of @NVIDIAAI Nemotron models. Mini is based on Nemotron 3 Nano. Max is based on the brand new Nemotron 3.5 Super VL. Both of these models are available as open weights on HuggingFace, and from four inference partners: AWS, Baseten, Cloudflare, and Modal.

harnesses are everything. and they help smooth out the jagged edges of intelligence. here's a primer on them. a few months ago, I wrote a series documenting learnings and useful patterns around building and working with harnesses with my friend @zakkeener42 at @baseten. last week, we compiled them and turned it into a small book. it covers how to optimize existing harnesses, how to choose the best one for you, how they are co-evolving with models. as a bonus, we added some thoughts on what the future might look like. link in description:

Open models can come from anywhere. Securing them shouldn’t depend on where you get them. Today we're launching Project Beacon to make safety a default for open intelligence. We're partnering with @GoodfireAI to bring inference and in-line monitoring together for the first time, so safety signals are caught proactively. Read more in our announcement: baseten.co/blog/safe-open-mo…