Ricursive Ships Ripple-65B, a 65 Billion-Parameter Open-Weight Model That Claims the US Agentic Coding Crown for $32 Million

Ricursive Intelligence, a Palo Alto startup building AI for chip design, has released Ripple-65B Preview, an open-weight language model it says sets a new US open-weight state of the art on agentic coding. The model is on Hugging Face under Apache 2.0, and is also served by Together AI, GMI Cloud and OpenRouter, with Fireworks support promised. Every performance and cost figure below is company self-reported; no independent evaluation was available at release.

The headline claim is a score of 80.52 on Terminal-Bench 2.1, which would put the 65-billion-parameter dense model just ahead of Reflection's Beam (80.1) — a 501-billion-parameter model — and well ahead of Meta's Muse Spark Open 816B (76.4), Thinking Machines' Inkling 975B (63.8) and NVIDIA's Nemotron 3 Ultra 550B (56.4). Ricursive also reports 79.1 on SWE-bench Pro v2.0 Hard against 76.5 for Beam.

Cost is the other half of the pitch. Ricursive says the full R&D bill — from ladder experiments and data mixing to synthetic data generation, evaluations and the final runs — was $32 million, which it estimates at roughly one-tenth of what the previous open-weight agentic-coding leader required. The company says it reached its Terminal-Bench score with 1.4 million reinforcement learning rollouts, about 70 times fewer than Beam, having bought no post-training data at all and built its own synthetic pipelines instead.

Architecturally, Ripple is a modernized dense Transformer rather than a mixture-of-experts model: 80 layers with a narrow width of 5,120, gated attention, global NoPE, value residuals and FFN activation clipping. Ricursive says the design kept model FLOPs utilization at 40% during training with simple sharding, no custom kernels and no quantization. The recipe started from AI2's Olmo 3 model ladder, and the company reports a 16x improvement in compute efficiency for code tasks at the 4B scale of its experiments.

The model card tells a less uniform story. Ripple trails Beam on SWE-bench Verified, scores 44.3 on DeepSWE 1.1 against 60.0 for Muse Spark Open, and posts 39.4 on SciCode — the lowest of the five models in Ricursive's own table. It lags on knowledge benchmarks like GPQA Diamond. Ricursive itself notes that replicating agentic benchmark scores is difficult and calls its comparison a best-effort attempt, and the model card acknowledges discrepancies with its own launch graphic on SWE-bench Pro v2.0 Hard. OpenRouter's current endpoint also caps context at 16,384 tokens, below the advertised 1 million-token capacity.

The team is a notable part of the story: chief executive Anna Goldie and CTO Azalia Mirhoseini previously helped build Google's AlphaChip, the reinforcement-learning system for chip floorplanning, before raising a $300 million Series A in January at a $4 billion valuation. Ripple-65B Preview is a general-purpose variant of the proprietary model that powers Ricursive's AI chip-design platform. Pricing on OpenRouter is $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.05; Ricursive reports about 250 tokens per second on two B200 GPUs, and an NVFP4 build that runs on a single NVIDIA DGX Spark.

For buyers of open models, the release sharpens an argument that has been building all year: benchmark leaders are increasingly not the largest models, and dense architectures remain competitive at a fraction of the parameter count. The "Preview" label and relatively light post-training suggest the model will keep changing, and teams weighing it should treat the Terminal-Bench headline as a claim to reproduce on their own repositories — not a specification.