Ripple 65B: A Chip Design Lab Just Beat Every US Open Model

Video thumbnail: Ripple 65B: A Chip Design Lab Just Beat Every US Open Model
Nov 5, 202620m 2s video lengthSam Witteveen

The Signal

Recursive — the chip-design lab behind Google's AlphaChip technology — has released Ripple65B-Preview, a 65B dense model. While it is a significant, Apache 2.0-licensed open-weight contribution with base weights and a transparent training recipe, its headline performance claim of 80.52 on Terminal Bench 2.1 is scoped strictly to US open-weight models, not global benchmarks.

The Case

Release Architecture and Economics

  • The model is a 65B dense architecture with 80 layers, chosen specifically for deployment simplicity over the more complex Mixture-of-Expert systems common in other labs.2:28
  • Recursive reports the total R&D cost at $32 million, a figure they claim is roughly one-tenth the cost of previous open-weight leaders, though this comparison relies on reported compute deals.0:49
  • Training utilized a 'laddering' approach inspired by AI2, where smaller variants were used to test architectural changes before scaling, alongside 1.4 million synthetic RL rollouts rather than purchased post-training data.9:17

Performance and Positioning

  • The 80.52 benchmark score makes it a strong contender among US open-weight models, yet it trails roughly 10 Chinese open models—such as DeepSeek Flash 4.1—that report higher scores on the same benchmark.4:52
  • Performance in local demonstrations shows strong tool-use and agentic coding capabilities after resolving initial SGLang serving issues, though visual design and SVG output remain less polished than competitor models.12:20
  • The company positioned this as a general-purpose release for their own platform needs, keeping the more specialized, chip-making version proprietary.1:46

The 1 Minute Signal Take

Ripple65B is a high-trust release that prioritizes transparency and dense-model efficiency over chasing global state-of-the-art titles. It is a capable tool for local agentic workflows, provided you evaluate it against dense benchmarks and recognize the limitations of its design-task training.

Pro Analysis

Why It Matters

This release signals a maturation of the open-weight ecosystem where transparency is becoming a competitive differentiator. By publishing a full training recipe alongside the weights, Ricursive is forcing a shift in expectations for what "open" means—moving away from black-box binaries and toward actionable technical knowledge.

Strategic Implications

For enterprises and independent developers, Ripple65B offers a high-performance alternative to proprietary models, especially for coding and tool-use tasks. The move away from MoE suggests that for specific high-density applications, the industry may return to simpler architectures that are easier to optimize for dedicated hardware.

Evidence & Hype Audit

This content is based on a verifiable model release and self-reported metrics. The speaker is careful to distinguish between "US open-weight" leaders and overall SOTA, adding a necessary layer of skepticism regarding benchmark-maxxing. The transparency of the $32M R&D figure is refreshing, though it remains a self-reported datum without external audit.

Counterarguments

Critics might argue that a 65B dense model is inefficient for inference compared to MoE models that achieve similar performance with lower active parameter counts. Additionally, the reliance on self-reported benchmarks means we must wait for independent evaluation (e.g., from platforms like VALS) before confirming if the performance lead holds under standardized conditions.

Who Should Care

  • Infrastructure Engineers: Benefit from the simpler dense architecture which avoids the complexity of expert-weight management.
  • AI Researchers: Benefit from the detailed training recipe and laddering methodology.
  • Agentic Workflow Developers: Benefit from the strong tool-use and coding performance demonstrated locally.

What To Do Next

  • Benchmark the NVFP4 version on your specific local hardware to assess true latency.
  • Integrate the model into existing agentic pipelines using DeepSeek Harness.
  • Review the training recipe artifacts to compare your own scaling ladder against their success metrics.
  • Evaluate the model’s SVG generation capabilities against your existing specialized models if visual output is a core requirement.
  • Monitor independent leaderboard updates for non-vendor-provided performance scores.
Time saved:16m 57s

Share this

Tags

Written by: 1 Minute Signal Editorial Team