Ricursive AI Releases Ripple-65B Preview: A 65B Dense Open-Weight Model Scoring 80.5 on Terminal-Bench 2.1
Ricursive Intelligence has released Ripple-65B Preview, a 65B dense open-weight model pretrained from scratch for agentic coding. It is a general-purpose variant of the proprietary model that powers Ricursive's AI chip design platform.
The headline claim: a self-reported score of 80.52 on Terminal-Bench 2.1 — the top score among US open-weight models. Notably, every MoE rival it outperforms on that benchmark carries between 501B and 975B total parameters.
TL;DR
- Size: 65B dense (65B active), pretrained on 12T tokens, with up to 1M context.
- Runs on: BF16 on datacenter GPUs (250 tok/s on 2x B200). An experimental NVFP4 build targets NVIDIA DGX Spark.
- Performance: Frontier-level agentic coding for its size; trails larger MoE models on knowledge and science coding.
- Best result: Terminal-Bench 2.1 at 80.5 (self-reported).
- Weakest result: SciCode at 39.4 — the lowest of the five models in Ricursive's own comparison table.
What Is Ripple-65B Preview?
Ripple-65B Preview is a dense, open-weight model built specifically for agentic coding workloads. Unlike mixture-of-experts (MoE) designs that activate only a fraction of their parameters per token, Ripple-65B activates all 65B parameters on every forward pass — a design choice Ricursive frames as a trade-off favoring predictability and deployment simplicity over raw scale.
The model was pretrained from scratch on 12 trillion tokens and supports context windows of up to 1 million tokens, positioning it for long-horizon agentic tasks such as repository-scale code editing, multi-step terminal workflows, and tool-use chains.
A General-Purpose Spin-Off
Ripple-65B Preview is a general-purpose derivative of Ricursive's internal model for AI-driven chip design. That lineage helps explain its emphasis on structured reasoning and tool interaction — capabilities that matter as much in hardware design automation as in software engineering agents.
Deployment and Hardware
Ricursive is shipping the preview in two configurations:
- BF16 on datacenter GPUs: The primary build, running at roughly 250 tokens per second on 2x NVIDIA B200 GPUs.
- Experimental NVFP4 build: A quantized variant targeting NVIDIA DGX Spark, aimed at local and edge-adjacent agentic deployments.
The dual-track release reflects a broader 2026 trend: open-weight labs increasingly ship both a high-throughput datacenter configuration and a compressed build for workstation-class hardware, letting developers prototype locally before scaling to cluster inference.
Benchmark Positioning
Ricursive's self-reported results place Ripple-65B Preview at frontier-level agentic coding performance relative to its parameter class. The comparison is deliberately framed around efficiency: the model's 65B dense footprint is roughly an order of magnitude smaller than the 501B–975B MoE models it beats on Terminal-Bench 2.1.
The picture is more mixed elsewhere. On knowledge-intensive evaluations and science-focused coding tasks, Ripple-65B trails larger MoE competitors — an expected outcome given the parameter gap. Its SciCode score of 39.4 is the lowest among the five models in Ricursive's own table, a limitation the company appears to disclose transparently rather than obscure.
Reading the Numbers Carefully
All figures cited are self-reported. As with any vendor benchmark release, independent reproduction — particularly on Terminal-Bench 2.1 under standardized harness conditions — will be the real test of the 80.5 claim. The gap between self-reported and third-party-verified agentic coding scores has narrowed in 2026, but it has not closed.
Why It Matters
The release adds a credible US-origin entrant to the open-weight agentic coding segment, which through 2026 has been dominated by large MoE architectures from both US and Chinese labs. A 65B dense model that runs at production speed on two B200s — and can be quantized for a DGX Spark — lowers the infrastructure bar for teams building terminal-native coding agents.
Whether the Terminal-Bench 2.1 result holds up under independent evaluation will determine how quickly Ripple-65B Preview moves from preview to production adoption.
via MarkTechPost