Open-Source Models Close the Gap on Speed and Agentic Tasks With Ornith-1.5 and a 1,000 tok/s Diffusion LLM
Two releases — an MIT-licensed 122B model claiming SWE-bench of 0.845 and a diffusion LLM pushing past 1,000 tokens per second — show the open ecosystem attacking the two dimensions enterprises actually care about: agentic reliability and latency.
The open-source frontier advanced on two fronts this week, and both target weaknesses that have historically kept open models out of production. The first is Ornith-1.5-122B, announced by @ornith_ with the tagline 'No Tradeoff Between Strong and Fast.' The release claims a SWE-bench score of 0.845 — a number that, if it holds up under independent grading, puts it in serious contention among open-weight coding models — alongside strong agentic benchmarks, all under a permissive MIT license.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.