Ricursive Intelligence Releases Ripple-65B Preview, Claims Efficient Frontier Open-Weight Model

AI chip design startup Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense language model it describes as a general-purpose variant of the proprietary model powering its AI chip design platform. The company claims the model sets a new U.S. open-weight state-of-the-art on agentic coding while costing roughly one-tenth of comparable efforts.

What was released

Ricursive Intelligence (ricursive.com) published Ripple-65B Preview on November 4, 2026, under an Apache 2.0 license. The model is a dense Transformer—meaning all 65 billion parameters activate on every token—pretrained from scratch on 12 trillion tokens of predominantly English and code data, with a 1-million-token context window. Four checkpoints are available: a BF16 instruction-tuned model, an experimental NVFP4 quantized version targeting local devices like NVIDIA DGX Spark, a BF16 base model, and a DFlash2 drafter model [1][2][4]. Weights are hosted on Hugging Face; inference is offered via GMI Cloud and OpenRouter, with Together AI listing the model as coming soon and Ricursive stating Fireworks follows [1][2][4].

The company frames the release as a "preview" with light post-training focused on agentic coding, noting that further training rounds are planned [1][3].

Benchmark claims and independent context

Ricursive Intelligence self-reports a score of 80.52 on Terminal-Bench 2.1, an agentic coding benchmark, saying this surpasses larger U.S. open-weight models including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4) [1]. The company also reports 79.1 on SWE-bench Pro v2.0 Hard [4] and 76.5 on SWE-bench Verified [2][4].

Benchmark aggregator BenchLM lists the Terminal-Bench 2.1 and SWE-bench Verified scores as "provider exact" and "display only," meaning they are self-reported and not independently verified. The SWE-bench Pro v2.0 Hard score of 79.1 appears only in the company's own materials and is not tracked by BenchLM. BenchLM's best verified score on Terminal-Bench 2.1 is 92.8% (SWE-2), leaving a 12.3-point gap. Ripple-65B Preview trails on several other benchmarks: SciCode (39.4% vs. 60.1% for Sakana Fugu), DeepSWE (44.3% vs. 79.0% for Muse Wonder Full), and τ³-bench (31.3% vs. 96.0% for Mercury 2.5) [2]. Only 17 of 817 tracked benchmark slots have displayable evidence, and the model is excluded from public rankings pending more non-generated benchmark coverage [2].

Architecture and training efficiency claims

Ricursive Intelligence chose a dense architecture over the mixture-of-experts (MoE) designs common at this scale, arguing it reduces infrastructure complexity. The model uses a "thin-and-deep" layout: 80 layers with a narrow d_model of 5,120, overprovisioned attention heads, a high sliding-window-to-global layer ratio with small windows, gated attention, global NoPE, value residuals, and FFN activation clipping [1]. The company reports maintaining 40% model FLOPs utilization (MFU) throughout pretraining and post-training with simple sharding, no custom kernels, and no quantization [1].

Pretraining followed a laddering approach starting from the Olmo 3 recipe. Ricursive says compounded changes yielded up to 16x compute efficiency on code at 4B scale, with the slope suggesting further gains at scale [1]. Post-training used no purchased data; the company built its own synthetic pipelines and a mixed reinforcement learning recipe with truncated importance sampling, score centering, and a linear length penalty. It reports reaching 80.5 on Terminal-Bench 2.1 with 1.4 million RL rollouts—roughly 70x fewer than Reflection Beam's reported 100 million [1][4].

Cost claim

Ricursive Intelligence states the full R&D cost was $32 million, covering pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, post-training recipe development, and the final hero runs [1]. The company estimates this is one-tenth the cost of the prior open-weight SOTA on agentic coding, citing reports that Reflection committed $150 million per month for compute at SpaceX Colossus [1]. The $32 million figure is the company's own estimate "to the best of our knowledge" and has not been independently audited [1].

Real-world deployment friction

A hands-on test by MindStudio found that running Ripple-65B Preview locally required significant engineering effort. A quantized version consumed approximately 75 GB of VRAM on a 96 GB VM with an NVIDIA RTX Pro 6000 Blackwell GPU. The provided inference code had to be rewritten for the test hardware, and a custom serving layer with an agent framework (Hermes) was needed before the model could be evaluated. Total setup took several hours [3].

Once running, the model successfully identified a subtle, planted bug in a full-stack retirement calculator application (a decimal formatting error treating a 4% withdrawal rate as "4" instead of "0.04") and produced a correct fix without guidance. It also handled a complex multi-table SQL/Databricks query task involving nested arrays, window functions, and join logic correctly [3].

API access via OpenRouter currently offers a 16,384-token context window at $0.20/$0.60 per million input/output tokens, well below the documented 1 million-token maximum [2]. Self-hosting at BF16 precision requires roughly 80 GB VRAM (one H100 80 GB) [2].

What remains unknown

• No independent replication of the Terminal-Bench 2.1 or other benchmark scores exists in the available sources.
• It is unclear whether the $32M figure includes opportunity costs or failed runs not counted in the final hero tally.
• The model's proprietary chip-design variant has not been publicly evaluated, to our knowledge.
• Longer-term stability, safety, and generalization beyond coding tasks are unproven; the company acknowledges light post-training and gaps in math, safety, and automation benchmarks [1][3].
• Availability on Together AI and Fireworks is listed as forthcoming [2][4].

Ricursive Intelligence says it is hiring as it pushes toward "full-stack AI from model to silicon" [1]. For now, Ripple-65B Preview demonstrates that a dense, 65B model can achieve competitive agentic coding scores with a reported low training cost, but its broader capabilities and the veracity of its headline numbers await independent verification.