Baseten

the cheapest way to get Opus 5-level agentic performance right now is an open model. GLM-5.3 is 743B params with 40B active. that's about a quarter the size of Kimi K3, and it matches or beats it on intelligence. the benchmarks are clear. the harder questions are whether it holds up on your tasks and what a task actually costs you at production volume. that's what we're covering wednesday 9/2 at 11am PT. members of our technical staff on serving GLM-5.3, benchmarking it on your own evals, and working out per-task cost for your workloads. register here: https://lnkd.in/ee6tWqNt

"If you're not taking care of yourself, you're not writing good code." We were honored to host Bryan Johnson and Sarah Guo at our office for a live discussion with some of our customers and friends. Thank you to everyone who joined us. 💚

As my internship at Baseten wraps up I wanted to show what I've been working on! Over the past year, agents have shown to be remarkably capable of generating high-performance kernels from scratch. But generating an isolated kernel is very different from improving and shipping an optimization for production within a serving engine. We built a framework to bridge that gap, offering a glimpse into the future of kernel engineering. https://lnkd.in/gR89EfWK

Post-train GLM-5.3 and GLM-5.3-Flash on Baseten Loops. Inference + training support on day 0. https://lnkd.in/d2yZpK6y

GLM-5.3 is live on Baseten Model APIs, day 0. - The smartest open-weight model at 743B params - 1M token context - US only - ZDR Stay tuned for more performance optimizations from our engineers.

We're proud to be the fastest inference provider on Artificial Analysis, OpenRouter, and Hugging Face for GLM-5.3-Flash, at 122+ TPS. All served from the US only, starting on day 0, with ZDR by default. Stay tuned for updates as our engineers continue to optimize GLM-5.3-Flash for even higher throughput and lower latency.

Mika Okamoto and I are releasing PACT (Pressure-Applied Compliance Testing): a benchmark for whether enterprise AI assistants keep following workplace rules when breaking them is convenient. We tested 23 models (18 open-weight, 4 closed) on 3,364 multi-turn trials across regulated domains like HIPAA, hiring law, and GDPR. Each trial gives the assistant a standing rule, makes the violating option the easy one, and adds ordinary corporate pressure: a deadline, a manager's verbal OK, "my colleague did it and nothing happened." PACT spun out of our #AIES2026 paper on why AI agents break rules; this time we turned the question into a full benchmark. The two highest-scoring models were open-weight. Kimi K2.7 (0.944) and Qwen3.6 27B (0.943) finished statistically tied, ahead of every closed model we tested. Bigger isn't better here; a 27B you can run yourself co-leads the board, and open models are competitive on all six dimensions (baseline, pressure resistance, transparency, and others) we measure. Not one model aced it. One sentence of pressure raised violation rates 65%, and the best model still missed 1 decision in 18. If you're deploying in a regulated workflow, evaluate on your own workload, and don't assume the closed-source frontier is the frontier for rule-following. Leaderboard, interactive trials, dataset, paper, code: https://lnkd.in/gPmB73Ns Baseten funded the inference for all ~232,000 trials. An inference company paying to measure how models behave, not just how fast they serve, is why I love working here.

GLM-5.3-Flash is live, day 0, on Baseten Model APIs. - Useful for general intelligence + agentic coding - More intelligent and 90% cheaper than GLM-5.2 - Supports vision natively - 1M token context