Ricursive Intelligence, the Palo Alto chip-design startup founded by Anna Goldie and Azalia Mirhoseini, released its first public model on November 4: Ripple-65B Preview, a 65-billion-parameter dense language model it says it pretrained from scratch. Ricursive claims it scores 80.52 on TerminalBench 2.1, which it calls a new state of the art for U.S. open-weight models, ahead of mixture-of-experts (MoE) models with up to 15 times as many parameters.
Why it matters: the open-weight race has been dominated by ever-larger MoE models that need a server to run. Ricursive says its 65B model runs locally on a desktop-class machine such as NVIDIA's DGX Spark, and it says the whole project cost $32 million. If the benchmark holds up on real workloads, it is evidence that a focused lab can compete on agentic coding without a 500B+ parameter budget. That "if" is doing a lot of work, because every figure below comes from Ricursive.
The claimed scores
Ricursive's Ripple page lists these TerminalBench 2.1 results. It doesn't publish a full benchmark table, and it doesn't say who ran the competitor numbers.
| Model | Parameters (per Ricursive) | TerminalBench 2.1 |
|---|---|---|
| Ricursive Ripple-65B Preview | 65B dense | 80.52 |
| Reflection Beam | 501B | 80.1 |
| Meta Muse Spark Open | 816B | 76.4 |
| Thinking Machines Inkling | 975B | 63.8 |
| NVIDIA Nemotron 3 Ultra | 550B | 56.4 |
Note the margin over Reflection Beam is under half a point. That is a tie in practical terms, not a rout. The bigger story is the gap in size and claimed cost.
What's inside
According to Ricursive's page:
- Architecture: dense Transformer, 80 layers, d_model 5,120, mixing sliding-window and global attention layers, with gated attention and no positional encoding on the global layers.
- Context: up to 1 million tokens.
- Training data: 12 trillion tokens. StartupHub.ai reports it was built on the Allen Institute for AI's Olmo 3 scaling ladder with a code-heavy data mix, and Ricursive's page credits AI2's Olmo team.
- Reinforcement learning: 1.4 million RL rollouts, which Ricursive says is about 70 times fewer than Reflection Beam used.
- Efficiency: about 40% model FLOPs utilization in both pre- and post-training, using simple sharding, no custom kernels and no quantization.
- Cost: $32 million end to end, which Ricursive says covers R&D, data work, evals, synthetic data and the final runs. It estimates that is roughly a tenth of what the previous open-weight leader cost, but that comparison is its own estimate.
The model has a day job. Ricursive says an internal variant powers its AI chip-design platform, and Ripple is a general-purpose version of it. RuntimeWire notes the founders created AlphaChip and that the company reportedly raised $300 million at a $4 billion valuation, citing Data Centre Dynamics.
How to try it
Ricursive says the weights are on Hugging Face as ricursiveAI/Ripple-65B-Preview, with hosted access through Together AI, GMI Cloud and OpenRouter (via GMI Cloud), and Fireworks to follow. It says the model runs efficiently on local machines such as NVIDIA's DGX Spark, but its page gives no memory figures or official quantized builds.
We could not confirm the licence: neither Ricursive's page nor the coverage we read states it, and we could not load the model card. Check the Hugging Face card before using it commercially.
What's still missing
- Independent results. RuntimeWire and StartupHub.ai both point out the scores are company-reported. Nobody outside Ricursive has published a TerminalBench run yet.
- Wider benchmarks. TerminalBench measures terminal-driven agent tasks. One leading score doesn't tell you how Ripple handles your repo, your languages or long multi-file edits.
- "Preview" status. Ricursive labels it a preview, and its page doesn't discuss limitations.
Who should care
Teams that want a self-hostable coding agent model, especially those who can't send code to a hosted API, should put Ripple on their eval list now. Run it on your own tasks against what you use today, read the licence on the model card first, and treat the leaderboard claim as a hypothesis until independent numbers land.
