Chip Design Start-Up Ricursive Releases Ripple-65B, an Open-Weight Coding Model

The weights are on Hugging Face under the Apache 2.0 licence, which allows commercial use. According to the model card, Ripple is a general-purpose version of the proprietary model behind Ricursive's AI chip design platform. The company was co-founded by Anna Goldie and Azalia Mirhoseini, who led the Google research that produced the AlphaChip chip-layout system.

The release matters because of its size. Every model Ricursive compares it against on Terminal-Bench uses a mixture-of-experts design with far more parameters: Reflection's Beam has 501 billion, Meta's Muse Spark Open 816 billion, Nvidia's Nemotron 3 Ultra 550 billion and Thinking Machines' Inkling 975 billion. Ripple is dense, so all 65 billion parameters are used on every token. The company says an experimental compressed NVFP4 version is aimed at desktop hardware such as Nvidia's DGX Spark, while the full BF16 model runs at about 250 tokens a second on two B200 GPUs at a batch size of eight.

On the model card, Ripple scores 80.5 on Terminal-Bench 2.1, against 80.1 for Beam, 76.4 for Muse Spark Open, 63.8 for Inkling and 56.4 for Nemotron 3 Ultra. It also leads the table on SWE-bench Pro v2.0 Hard with 79.1, and scores 98.3 on the AIME 2025 maths test. It is weaker elsewhere. It scores 76.5 on SWE-bench Verified, behind Beam, and 85.2 on the GPQA Diamond science test, where Beam and Muse Spark Open lead. According to MarkTechPost, Ripple's 39.4 on SciCode is the lowest of the five models in Ricursive's table.

All the figures are self-reported, and several come from Ricursive's own runs of rivals' models. The company says agentic benchmark results are hard to reproduce and depend heavily on the test harness, and it calls its comparisons "best-effort". MarkTechPost reported that when Ricursive reran Terminal-Bench 2.1 with the official Terminus 2 harness, most rivals scored lower, Beam falling to 70.8, while Ripple stayed ahead.

Ricursive trained the model from scratch on 12 trillion tokens of mostly English text and code. The design is "thin and deep", with 80 layers and a hidden size of 5,120. The company chose a dense transformer over mixture-of-experts to keep its infrastructure simple. It says training ran at about 40 per cent model FLOPs utilisation without custom kernels, and that the model supports a context of up to one million tokens. The model card says the team grew from one person to five during development and claims a 16-fold gain in compute efficiency on code at small scale, rising to 55-fold for the full run.

According to MarkTechPost, Ricursive puts the full research and development cost at $32 million, including experiments, data work, evaluations and final training runs. The company estimates this at one-tenth the cost of the previous open-weight leader in agentic coding. MarkTechPost also reported that the company's reinforcement learning stage reached its Terminal-Bench score with 1.4 million rollouts.

The release includes three checkpoints: the instruction-tuned BF16 model, the NVFP4 version and a BF16 base model, plus a DFlash2 companion. The "Preview" label reflects what the company calls relatively light post-training focused mainly on agentic coding. MarkTechPost said Together AI, GMI Cloud and OpenRouter are listed as launch hosts, with Fireworks to follow, but the Hugging Face page did not yet show any inference provider serving the model, and Together AI listed it as coming soon.