A general-purpose model from a chip design company
Ricursive Intelligence, a company building AI for chip design, released Ripple-65B Preview on Hugging Face under the Apache 2.0 license on November 4. According to its launch blog, this is a general-purpose variant of the proprietary model that powers its chip design platform: a 65-billion-parameter dense Transformer, not a Mixture-of-Experts (MoE) model, pretrained from scratch on 12T tokens of text-only data that is predominantly English and code, with support for up to 1M tokens of context. The release includes a BF16 instruct model, an NVFP4 quantized version, and a Base model that is pretrain-only, without mid-training or synthetic QA data.
The model card says training started from Ai2's Olmo 3 baseline before switching to a Gemma 3N-like architecture with a "thin-and-deep" shape: a hidden size of 5120 across 80 layers, plus changes such as gated attention, global NoPE and value residuals. The case for dense over MoE was lower infrastructure complexity: the team grew from one to five people during development and kept about 40% MFU through pretraining and post-training with simple sharding, no custom kernels and no quantization.
How big the lead is, and where it trails
Ricursive self-reports 80.52 on Terminal-Bench 2.1, slightly above Reflection Beam 501B's 80.1 and ahead of Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8) and NVIDIA Nemotron 3 Ultra 550B (56.4). The model card is also candid about how hard these numbers are to reproduce: when everything was re-run on the official Terminus 2 harness, Beam dropped to 70.8 and Muse Spark Open scored 78.7.
In the same table, Ripple does not lead everywhere: SWE-bench Verified 76.5 (Beam 81.6), DeepSWE 1.1 44.3 (Muse Spark Open 60.0), SciCode 39.4 (Muse Spark Open 58.5) and MCP Atlas 71.67 (Muse Spark Open 89.2). Two more points matter. The comparison includes only US open-weight models. And Ripple activates all 65B parameters on every token, more than the active parameters of any MoE model in the table (23B to 55B), so "small beats big" holds only when counting total parameters. Ricursive also notes that the Preview has gone through relatively light post-training focused on agentic coding.
What the $32 million covers
The blog puts Ripple-65B Preview's total R&D cost at $32 million, covering scaling-ladder experiments, data mixing, evals, debugging, synthetic data, post-training recipe development and the final training runs. By the company's own estimate, that is about one-tenth the cost of the previous open-weight leader in agentic coding, and its reinforcement learning used only 1.4 million rollouts, roughly one-seventieth of Beam's. These figures are Ricursive's own estimates.
For deployment, 65B fits on local hardware such as the DGX Spark, and Ricursive provides an NVFP4 checkpoint and a Colab notebook; in a data center it reaches about 250 tokens per second on two B200s at batch size 8 before quantization. For teams that want to run coding agents locally or in private environments under a permissive license, it is a new option worth testing on their own tasks.
via: Ricursive: Introducing Ripple-65B Preview, Hugging Face model card: Ripple-65B-Preview