5-person team trains Ripple-65B from scratch, sets US open-weight SOTA on Terminal-Bench 2.1
annadgoldie · x · 2026-11-05
RicursiveAI released Ripple-65B Preview, a dense 65B LLM pretrained entirely from scratch by a five-person team, claiming a new US open-weight state of the art on Terminal-Bench 2.1 for agentic coding — beating much larger models like Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Key design choices:
- Dense over MoE: for long agent sequences of model and tool calls, they prioritized strong reasoning and low latency, plus a simpler training/serving stack without expert-routing communication overhead.
- No purchased posttraining data: all data came from in-house generation and filtering pipelines, built on open-source training and inference stacks.
More from coding & agent
- Running Qwen 27B Q4 at 0.7 tok/s on a 4GB-VRAM laptop as a main coding model — Felix_455-788 · 2026-11-05
- His agent now auto-labels his Google Calendar in the background — steren · 2026-11-05
- The security startups racing to rein in AI agents — pstAsiatech · 2026-11-05
- OpenAI October dev roundup: Decisions API, 12-hour caching, 8x faster Codex tokens — OpenAIDevs · 2026-11-05
- Datadog previews Canvas, a new way to collaborate with AI agents on observability — farooqib · 2026-11-05
- One hour with Opus 5.5 fully fan-translates a 2004 GBA game, graphics included — ezshine · 2026-11-05