Ricursive's Ripple-65B Preview: An Apache 2.0 Coding Model That Fits on a DGX Spark, Built for $32 Million

Ricursive Intelligence, which builds AI for chip design, has released a general-purpose version of the model behind its own platform. It isn't Recursive, the company that released the Flint coding model in October. The headline claim is narrow on purpose: best among US open-weight models on agentic coding, at a fraction of the size and cost of the models it beats.

The model

• Size and shape: a 65B dense transformer, not a mixture of experts, with 80 layers and a narrow 5,120-wide model dimension.
• Training: 12 trillion tokens of mostly English text and code at 8K length, then mid-training, long-context training up to 2 million tokens, fine-tuning and reinforcement learning. The card says post-training was relatively light and focused on agentic coding.
• Cost: about $32 million, which Ricursive says covers all research and development from scratch, not just the final run. It says it reached its Terminal-Bench score with 1.4 million RL rollouts, about 70 times fewer than Reflection's Beam.
• Team: the card says the team grew from 1 to 5 people during development.
• Downloads: BF16 and 4-bit NVFP4 instruction-tuned models, plus a BF16 base model. Apache 2.0 licence.
• Where it runs: Ricursive says it fits on an NVIDIA DGX Spark and reaches 250 tokens a second on two B200 GPUs. It's on Together AI, GMI Cloud and OpenRouter, with Fireworks listed as coming the same day.

Where it leads and where it doesn't

The caveats in Ricursive's own card

• The harness matters. Ricursive says Terminal-Bench 2.1 scores were lower for most models when it used the official Terminus 2 harness. Under that harness, Beam drops from 80.1 to 70.8 and Muse Spark Open rises from 76.4 to 78.7.
• Competitor numbers are mixed. Where a lab didn't publish a score, Ricursive ran its own evaluation. Muse Spark Open's SWE-bench Pro v2 Hard score is 52.3 in Ricursive's run, while Cognition's table the same day lists it at 82.4.
• "US open-weight" is the scope. Cognition's 4 November table lists DeepSeek V4.1 Flash 1017, an open-weight model from China, at 91.0 on Terminal-Bench 2.1, above Ripple's 80.5.
• It's a preview. The card describes light post-training, so later versions may change the picture.

What this means if you pick coding tools

• If you want a coding model you can run and fine-tune yourself under a permissive licence, Ripple is one of the few US options that fits on a single desktop-class machine.
• Test it with your own harness before trusting the leaderboard position. Its own card shows scores moving by up to 10 points with the harness.
• For long agentic tasks, Muse Spark Open still leads it on DeepSWE and MCP Atlas, at roughly 12 times the parameter count.