Ripple-65B open model tops Terminal-Bench 2.1 at 80.5%, trained from scratch on $32M compute
annadgoldie · x · 2026-11-05
Ricursive Intelligence released Ripple-65B Preview, a dense open-weight LLM pretrained from scratch that it claims sets a new SOTA among US open-weight models on agentic coding tasks.
Key facts:
- 80.5% on Terminal-Bench 2.1, narrowly beating the much larger Beam-501B (80.1%) — though Beam is a 501B-parameter MoE activating only 23B, while Ripple uses all 65B densely; the accompanying chart also shows leads on SWE-bench Pro v2.0 Hard (79.1%), DeepSWE 1.1, and HMMT
- Trained on 12 trillion tokens with a $32M compute budget
- Supports up to 1M tokens of context, released under Apache 2.0
- Reportedly runs locally on a DGX Spark
All benchmarks are vendor-reported and await third-party verification.
More from coding & agent
- Weaviate Podcast: Forgetting Is the Key to Scaling Agent Discovery, With Berkeley's Shangyin Tan — CShorten30 · 2026-11-05
- GPT-6 edits Qwen3.5 weights directly: 0.016% changes, 6x math gains, no training — lateinteraction · 2026-11-05
- Dots Team Recruits to Build Out Persistent Agents — sherwinwu · 2026-11-05
- Managing agent workstreams as infrastructure-as-code in production — JoshPurtell · 2026-11-05
- 'You own every line you submit': the coding policy that survives AI — esrtweet · 2026-11-05
- Notion AI agents gain Google Drive access: search files, edit Drive, trigger workflows — technextpreneur · 2026-11-05