Ricursive's 65B Dense Ripple Preview Tops U.S. Open Weights on Terminal-Bench 2.1

Ricursive Intelligence has published Ripple-65B Preview, a general-purpose open-weight cousin of the proprietary model that drives its AI chip-design platform, and claims the top U.S. open-weight score on Terminal-Bench 2.1 among models it compared. The company announcement and Hugging Face card, both dated for the November 4, 2026 release window, describe a 65B dense Transformer pretrained from scratch on 12 trillion mostly English and code tokens, with support for up to 1 million tokens of context. Weights ship under Apache 2.0. The Preview package includes an instruction-tuned BF16 checkpoint (this card), an experimental NVFP4 build aimed at NVIDIA DGX Spark-class local boxes, and a BF16 base model without midtraining or synthetic QA.

The headline agentic number is 80.52 (rounded to 80.5 in tables) on Terminal-Bench 2.1. Ricursive’s comparison table places that above Reflection Beam at 80.1 (501B MoE), Meta Muse Spark Open at 76.4 (816B), Thinking Machines Inkling at 63.8 (975B), and NVIDIA Nemotron 3 Ultra at 56.4 (550B)—all self-reported or filled by Ricursive’s in-house harness where a vendor card was silent. On the official Terminus 2 harness at 128K with summarization, Ricursive says most rivals dropped while Ripple held 80.5 and Muse Spark Open rose to 78.7. Other instruct scores on the card include SWE-bench Verified 76.5, SWE-bench Pro v2.0 Hard 79.1 (Terminus 2), AIME 2025 98.3, HMMT Feb 2026 90.5, and GPQA Diamond 85.2. Weak spots the lab itself flags include SciCode 39.4, DeepSWE 1.1 44.3, and MCP Atlas 71.67.

Dense on purpose, thin and deep on paper

Ricursive chose a dense stack over Mixture-of-Experts to cut pretraining and serving complexity. The company says the hero runs held about 40% MFU with simple sharding, no custom kernels, and no quantization during training. The aspect ratio is thin and deep: d_model 5120, 80 layers. Relative to an Olmo 3 Llama-like starting point, the recipe moved toward a Gemma 3N-like layout for stability, over-provisioned attention heads for reasoning, raised the sliding-window-to-global ratio while shrinking window size, and added gated attention, global NoPE, value residuals, and FFN activation clipping. Ladder experiments from Olmo 3’s baseline through StepFun scaling, pre+post norm, gated attention, rejected 2-stream mHC at 4B, and a final code-heavy corpus are credited with up to a 16× code compute-efficiency gain at 4B scale on dolma_stack cross-entropy, widening further on the hero run in Ricursive’s telling.

Post-training followed midtraining, long context (including 2M sequence stages), SFT, long-context SFT, and mixed RL. The lab says it bought no post-training data and built its own synthetic pipelines. The mixed RL recipe uses truncated importance sampling, score centering, and a linear length penalty. Ripple reached the Terminal-Bench 2.1 mark with about 1.4 million rollouts—Ricursive puts that at roughly 70× fewer than Beam’s reported figure of more than 100 million. Full R&D cost from scratch, including ladders, data, evals, synthetic work, and hero runs, is stated as $32 million, framed as about one-tenth the prior open-weight agentic coding leader’s estimated spend.

Where it runs and what is still Preview

On datacenter GPUs, Ricursive reports about 250 tokens/s on 2× B200 at batch 8 before quantization. The NVFP4 checkpoint quantizes MLP layers and keeps attention in BF16 for local devices. Hosted routes listed at launch include Together AI, GMI Cloud, and OpenRouter, with Fireworks said to follow the same day. The card is explicit that Preview post-training is relatively light and focused on agentic coding, that harness choice moves scores, and that several competitor cells are Ricursive’s own OpenCode or Terminus 2 fills.

Ripple is positioned as the public wedge of a full-stack lab that co-optimizes models with the silicon it designs. For open-weight builders, the concrete offer is a commercial Apache dense 65B with a published Terminal-Bench lead among the U.S. open set Ricursive tabulated—and a clear list of science and MCP benches where larger MoE rivals still win on the same card.