ModelsNov 4 · ricursive.com

Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026, and it is the company's first model. <cite index="33-2">On Terminal-Bench 2.1, a test of how AI agents handle real Bash environments, Ripple scores 80.5, narrowly ahead of Reflection AI's 501-billion-parameter Beam at 80.1.</cite> <cite index="33-4">Ripple is dense, so all 65 billion parameters work on every request, while Beam activates 23 billion of its 501 billion parameters per task.</cite> The company also lists Meta's Muse Spark Open, Thinking Machines' Inkling and NVIDIA's Nemotron 3 Ultra as models it beats on the same test.

<cite index="8-4">Ripple-65B Preview was trained on 12 trillion tokens, supports up to 1M context length</cite>. <cite index="6-7,6-8">Ricursive puts full R&D cost at $32 million, covering pretraining ladder development, data mixing, evals, synthetic data generation, posttraining and the final run, which it estimates is one-tenth the prior open-weight state of the art on agentic coding.</cite> <cite index="29-10,29-11">It reached 80.5 with 1.4 million RL rollouts, which Ricursive says is roughly 70 times fewer than the over 100 million reported for Beam.</cite> <cite index="14-3">The starting point was the Olmo 3 recipe from the Allen Institute for AI, after which Ricursive moved to a Gemma 3N-like architecture to keep large-scale training stable.</cite>

<cite index="14-4">The package includes an instruction-tuned BF16 model, an NVFP4 build for NVIDIA GPUs, and a base checkpoint.</cite> <cite index="17-8">It is available on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks slated to go live later.</cite> Ricursive says the model fits on a MacBook or a DGX Spark. <cite index="11-4">Ripple-65B Preview is a general-purpose variant of the proprietary model behind Ricursive's chip-design platform.</cite> <cite index="33-5">The benchmark numbers are self-reported and have not been independently verified.</cite>

Why it matters

For teams building on open weights, the release offers a 65-billion-parameter base under a permissive license, with a claim that it runs on a single desktop-class machine. If the local performance holds in practice, a serving budget that once required a large cluster could shrink for coding agents. The reported cost also matters: <cite index="6-7">the $32 million full R&D figure</cite> would suggest a small team can reach frontier-adjacent coding results without a hyperscale training run, though the figure is the company's own accounting.

Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Why it matters

Analysis

The release arrives as U.S. labs try to close the gap with Chinese open-weight models on agentic coding in particular. <cite index="33-6">The claim that a smaller, denser model matches or beats systems with up to fifteen times its parameter count places Ripple in the middle of that race.</cite> <cite index="22-4">VentureBeat's coverage frames the lead as a claim on two benchmarks, Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard, the latter introduced in September.</cite> The same reporting notes that <cite index="22-5">its own results show it trailing other open models on additional benchmarks.</cite> Ricursive's table shows weaker results on some tests, including <cite index="24-8">IFBench and GPQA Diamond</cite>, and <cite index="29-7">SciCode, at 39.4, is the lowest of the five models in its own table.</cite>

Several caveats narrow the lead. <cite index="31-8">Several rivals' scores in the table are from Ricursive's in-house runs, marked in italics on the card, not from those labs.</cite> <cite index="22-8">The card says its tables combine published results with internal evaluations and acknowledges difficulty reproducing agent benchmarks.</cite> Results also depend on the harness. Under the Terminus 2 setup, <cite index="24-5">Muse Spark Open rises to 78.7 and Beam drops to 70.8, while Ripple stays at 80.5.</cite> The hosted context ceiling is also lower than advertised. One explainer notes that <cite index="21-3">the advertised context is up to 1 million tokens, but the limit currently listed by OpenRouter is 16,384 tokens.</cite> Commenters have also argued that the U.S. framing is narrow, since the comparison set leaves out strong non-U.S. models.

Second-order effects could matter beyond this model. If a dense 65B model holds up on independent tests, the argument for mixture-of-experts at this size weakens, and small labs gain a reason to invest in data and training recipes rather than parameter count. Ricursive credits its own posttraining stack and synthetic data pipelines for the efficient RL gains, which points to data work as a moat. <cite index="22-6">Its coding benchmarks, however, do not establish the effectiveness of the separate chip-design platform.</cite>

What to watch: independent reruns of Terminal-Bench 2.1 under other harnesses, since <cite index="30-10,30-11">independent tests will come when someone reruns Terminal-Bench with their own harness and their own machine</cite>. Watch whether Fireworks goes live, whether the preview becomes a full release with more post-training, since <cite index="25-4">post-training is still light in this preview release</cite>, and whether a version of the chip-design model is ever published. Teams that want to test the claim should check whether the ladder and RL recipe reproduce the terminal results on their own workloads.

Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Why it matters

Discussion · 0

Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding

Why it matters

Discussion · 0

Key points

Analysis