Ricursive releases Ripple-65B Preview, an open-weight dense model aimed at agentic coding

Ricursive Intelligence, a Palo Alto startup best known for applying AI to chip design, released Ripple-65B Preview on November 4. The model has 65 billion parameters, uses a dense architecture rather than a mixture of experts, and is published with open weights under the Apache 2.0 licence on Hugging Face. RuntimeWire reports that Ricursive presents it as a general-purpose version of the proprietary model behind its chip-design work.

What the model card says

According to the model card, Ripple was pretrained from scratch on 12 trillion tokens of mostly English text and code, using 8,000-token sequences, and supports a context window of up to one million tokens. It has 80 layers and draws on a design similar to Gemma 3n, with several additions such as gated attention and value residuals. The repository holds the instruction-tuned model in BF16; a base model and an NVFP4 quantised version sit in separate repositories. Ricursive notes that post-training was relatively light and concentrated on agentic coding.

On its own evaluations, the company lists 80.52 on Terminal-Bench 2.1, a benchmark of command-line tasks for agents, along with 76.5 on SWE-bench Verified and 61.42 on SWE-bench Pro. It also reports high scores on general tests such as MMLU and GPQA Diamond. RuntimeWire says Ricursive puts the training cost at about $32 million. The card claims large gains in compute efficiency compared with the Olmo 3 model family, and says the model can run on hardware such as Nvidia's DGX Spark.

Why caution is warranted

All of these numbers come from Ricursive. The model card itself acknowledges that agentic benchmark scores are hard to replicate, and RuntimeWire stresses that the Terminal-Bench result needs independent confirmation before developers rely on it. The "Preview" label also signals that this is not a final release. Because the weights are public, outside researchers can now test the claims directly.

Who is behind it

Ricursive was co-founded by Anna Goldie and Azalia Mirhoseini, whom RuntimeWire identifies as the creators of AlphaChip, an AI method for chip design. The outlet describes Ricursive as a $4 billion startup, citing a report of a $300 million raise. In a post on X quoted by RuntimeWire, the company claims a new state of the art for US open-weight models. Whether a chip-design company can sustain a line of general-purpose models, and how Ripple compares in everyday use with larger open models, is not yet clear.