Ricursive opens Ripple-65B: a dense Apache-2.0 agentic coding model
A 65B dense model claiming U.S. open-weight best on Terminal-Bench (80.5), runnable on one DGX Spark and trained for $32M — a tenth of the prior open SOTA.

Copy markdown
A 65B dense coder you can self-host
Ripple-65B Preview is an Apache-2.0, 65B dense Transformer with up to 1M context. BF16, NVFP4, and base checkpoints are all on Hugging Face now — Ricursive says it fits consumer hardware like an NVIDIA DGX Spark and reaches ~250 tok/s on 2×B200. Hosted routes are already live on Together, Fireworks, OpenRouter, and GMI Cloud.
SOTA on Terminal-Bench, mid-pack on SWE-bench
It claims U.S. open-weight SOTA on Terminal-Bench 2.1 (80.5), nosing past Reflection Beam 501B (80.1) at a fraction of the size, and tops SWE-bench Pro v2.0 Hard at 79.1. But on SWE-bench Verified it scores 76.5 — behind Beam (81.6), Muse Spark Open (78.7), and Inkling (77.6). All numbers are self-reported, and Ricursive admits replication "has been challenging."
Trained for $32M, a tenth of prior open SOTA
Ricursive puts full from-scratch R&D at $32M — about one-tenth the cost of the last open-weight coding SOTA — using ~70× fewer RL rollouts than Reflection Beam. The dense-over-MoE bet is deliberate (d_model 5120, 80 layers, 12T pretrain tokens). The pitch: frontier-class agentic coding you can fine-tune and run without a cluster.
Vercel auto-applies Token Pass credit in AI Gateway
As of today, your Token Pass credit covers included model usage through Vercel's AI Gateway automatically — whether you authenticate with an API key, an OIDC token, or CLI login. One fewer billing surprise if you route models through the Gateway.
Vercel Schedules hits public beta
Vercel Schedules is now in public beta: fire tasks on a cron cadence or at a one-off time, dispatched to a Vercel Function or a Queue topic in any environment. Useful for cron-driven agent runs and digest jobs without standing up your own scheduler.