Ripple-65B Preview is now on Hugging Face. Ricursive Intelligence, a US lab that builds AI for chip design, says the model is a dense Transformer with 65 billion parameters, pretrained on 12 trillion tokens, with support for up to 1 million tokens of context. The company presents it as a general-purpose variant of the proprietary model that powers its own AI chip design platform.
The release is a preview, and every headline number in Ricursive’s announcement is the lab’s own. On its site and the Hugging Face model card, Ricursive reports a TerminalBench 2.1 score of 80.52 and calls that US open-weight state of the art. It also says it developed and trained the model for $32 million in full research and development cost from scratch — its estimate of about one-tenth the cost of the previous open-weight SOTA on agentic coding.
In brief
- Ripple-65B Preview weights are on Hugging Face, with serving listed on Together AI, GMI Cloud and OpenRouter.
- Ricursive reports 80.52 on TerminalBench 2.1 versus other US open-weight models, and a $32M full R&D cost.
- The release is a preview; the benchmarks and cost figures are Ricursive’s, with no third-party audit cited.
What Ripple-65B Preview includes
Architecturally, Ripple-65B Preview is a modernised dense Transformer rather than a mixture-of-experts design. Ricursive lists 80 layers and a d_model of 5,120, a thin-and-deep shape it says was chosen to train and serve efficiently on NVIDIA hardware. The Hugging Face page for this repo hosts the BF16 instruction-tuned model; Ricursive also points to NVFP4 and base checkpoints, and says the preview went through relatively light post-training with a focus on agentic coding.
On availability, Ricursive lists Hugging Face, Together AI, GMI Cloud and OpenRouter, and said Fireworks would go live later the same day. It credits open-source work from the Olmo team at Ai2, Prime Intellect, RadixArk and NVIDIA as foundations it built on.
On TerminalBench 2.1, Ricursive’s comparison table is limited to US open-weight peers. The figures it publishes include Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4. Those are larger models by total parameter count; Ricursive’s point is that its 65B dense model matches or beats them on this agentic coding bench in its own tests. The model card also notes that reproducing modern agentic scores is hard, and that some of its comparison numbers come from its in-house eval flow when official cards omit a score.
Cost claims and what stays unproven
The $32 million figure is unusually specific. Ricursive says it covers full R&D from scratch: pretraining ladder work, data mixing, evals, debugging, synthetic data, post-training recipes and the final training runs — not only the last pretraining bill. The “one-tenth” comparison is the company’s own estimate against what it believes prior open-weight SOTA training cost. TSN cannot check either number from public filings, and readers should treat both as vendor claims.
The same caution applies to the leaderboard story. Self-reported TerminalBench scores have become a common way for new open-weight coding models to announce themselves, as with Moonshot’s Kimi K3.1 Pro, but third-party harnesses and prompt settings often move the numbers. Until an independent board or a shared harness run is published, 80.52 is best read as Ricursive’s result under its chosen setup, not a settled ranking.
Why a chip lab opening weights matters
Ricursive’s framing is that the model and the silicon should improve each other: an internal variant designs chips, and a public preview lets outsiders probe the same stack. That co-design story is distinctive next to pure model labs. A dense 65B envelope that fits local hardware such as NVIDIA DGX Spark, as Ricursive claims, also speaks to teams that want agentic coding without calling a frontier API for every loop.
The open-weight race itself is crowded. Mistral’s Large 4 weights and Meituan’s LongCat-2.5 show how quickly new checkpoints land, often with vendor benches attached. Ripple-65B Preview adds a US dense model that leans hard into agentic coding and cost efficiency, and that tries to make a virtue of a smaller total parameter count rather than chasing trillion-parameter MoE envelopes.
For buyers and researchers the practical questions are narrower than the marketing. Can a shared harness reproduce the TerminalBench number? How much of the agentic strength is locked in the preview’s light post-training versus the base checkpoint? And does co-design with a chip platform produce models that behave differently under long-horizon tool use than models trained only for chat and code completion? Ricursive has not said when a non-preview version will appear, so those answers will have to come from the weights that are public today.
Sources
Related stories
- Thinking Machines Ships Inkling 2 Small Via Tinker
- Moonshot’s Own Tests Put Kimi K3.1 Pro at 65.7% on Terminal-Bench
- Mistral Opens Large 4 Weights: a 1-Trillion-Parameter Model
- Meituan’s LongCat-2.5 AI Model Was Trained Only on Chinese Chips
- Ocient Raises $100m Series C Co-Led by BOKA and OCA
- Baidu Releases ERNIE-5.5-Flash-Base as Open Weights
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
