FULL STORY

Ripple-65B: From Open Release to Local Agentic Coding

Ricursive Intelligence open-sourced Ripple-65B Preview, claiming US open-source coding SOTA; it later landed on Ollama for local agentic coding.

2026-11-05 ~ 2026-11-17 · 3 episodes · 47 posts

Episode 1 · Chip-design firm Ricursive open-sources Ripple-65B, claiming US open-weight coding SOTA (2026-11-05, 42 posts)

Ricursive Intelligence released its first large model, Ripple-65B Preview: a 65B (64.7B-parameter) dense model fully pretrained from scratch on 12 trillion tokens, supporting up to 1M context, open-sourced under Apache-2.0 (safetensors with customcode). Weights went live on HuggingFace and briefly topped the trending list. The company claims a new US open-weight SOTA on agentic coding, with TerminalBench 2.1 score 80.5, significantly beating far larger models including Muse Spark Open-816B, Beam-501B, and a 975B-scale giant MoE, at a total compute cost of about $32 million—claimed to be one-tenth of the prior open-source agentic coding SOTA. Notably, Ricursive is a chip-design company rather than an LLM startup; this is its first LLM, reportedly built by a 5-person team.

Confirmed

  • Ripple-65B Preview is a 65B (64.7B) dense model, 12T tokens pretrained from scratch, 1M max context, Apache-2.0 open weights live on HuggingFace (verified by @coder543) and briefly trending (@ricursiveAI)
  • Claimed TerminalBench 2.1 score of 80.5 (80.52), a US open-weight agentic coding SOTA over 816B/501B/975B models; @Sam Witteveen also relayed the SOTA claim in his video overview
  • Claimed total cost of $32M, about one-tenth of the prior open SOTA; investor Guru Chahal (Lightspeed), relayed by @Azaliamirh, claims a 16x compute-efficiency training recipe
  • Per @Azaliamirh: trained on NVIDIA B300 GPUs, post-training uses open NVIDIA Nemotron tech, runs locally on DGX Spark; reposted by NVIDIA
  • Architecture is 'thin and deep' per @Sam Witteveen, who noted Ricursive is a chip-design company and this is its first LLM
  • Community evaluation (@returnity) suggests it competes with US frontier open-weight models of any size; Reddit user @DeltaSqueezer highlighted it as a rare new coding model under 100B
  • Founded by ex-Google TPU team members Anna Goldie and Azalia Mirhoosinghalamsi et al.; model available on OpenRouter, Together AI, Fireworks, and GMI Cloud (@teortaxesTex, @annadgoldie), with GMI Cloud as Day 0 launch partner (@DZhang50)

Unconfirmed

  • @teortaxesTex questions the comparison methodology behind the claimed SOTA and cost advantage; details await independent verification
  • The 16x compute-efficiency claim comes only from investor/promotional channels without independent validation

Why it matters

  • If the cost and performance claims hold, a far smaller budget could reach the open-source agentic coding frontier, reshaping the price-performance landscape of open coding models
  • A chip-design AI company shipping its first LLM with a reported 5-person team, backed by academics like Anna Goldie, Lightspeed, and NVIDIA reposts, makes the entry itself noteworthy

22 more related posts →

Episode 2 · Ripple 65-B Lands on Ollama for Local Agentic Coding (2026-11-10, 2 posts)

Ollama has made Ricursive AI's Ripple 65-B Preview available for local use with zero data retention. The 65B dense model, trained on 12T tokens under Apache 2.0, targets agentic coding with up to one million tokens of context.

Episode 3 · Ricursive Releases Open-Weight Ripple-65B, Claims Agentic Coding SOTA (2026-11-17, 3 posts)

Ricursive Intelligence, founded by Stanford professor Azalia Mirhoseini, released its first open-weight model Ripple-65B Preview, scoring 80.52 on TerminalBench 2.1 to claim SOTA among US open-weight models on agentic coding.