Ricursive Intelligence released its first model, Ripple-65B Preview, on November 4 (18:24 UTC), open-sourcing the instruction-tuned model (BF16 and NVFP4) and a base model on Hugging Face under Apache 2.0. The official model card describes a 65B dense Transformer pretrained on 12T tokens with up to 1M context, and self-reports TerminalBench 2.1 80.5, SWE-bench Pro v2.0 Hard 79.1 and DeepSWE 1.1 44.3, ahead of much larger US open-weight models on agentic coding. It fits on local hardware such as the DGX Spark. CEO Anna Goldie says it cost about a tenth of the prior comparable state of the art.
Key Takeaways
- ✓Ripple-65B Preview is a 65B dense model with a thin-and-deep shape (80 layers, d_model 5120), moving from an Olmo 3 baseline to a Gemma 3N-like architecture.
- ✓It was pretrained on 12T tokens at 8K sequence length, then went through mid-training, long-context training, SFT and mixed-domain RL, with up to 1M context.
- ✓Self-reported TerminalBench 2.1 is 80.5, above Reflection Beam 501B (80.1) and Muse Spark Open 816B (76.4), but SWE-bench Verified 76.5 and DeepSWE 1.1 44.3 show it does not lead everywhere.
- ✓Ricursive says the core team grew from one to five people and code compute efficiency improved 16x versus the Olmo 3 ladder; it reaches about 250 tok/s on 2x B200.
- ✓Hugging Face hosts BF16, NVFP4, base and DFlash2 speculative-decoding checkpoints, all under Apache 2.0.
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the Industry Pain Point
US open-weight models are often seen as lagging on agentic coding, and most are huge MoE models that are hard to run locally. Ricursive Intelligence, led by CEO Anna Goldie (formerly Google DeepMind and Anthropic), released its first in-house model, Ripple-65B Preview, pitched as small-team, low-cost and runnable locally.
Architecture Highlights and Mechanism
According to the model card, Ripple uses a dense Transformer rather than MoE to keep training and serving simple: 80 layers, d_model 5120, more attention heads, a higher sliding-window-to-global layer ratio, plus gated attention, global NoPE and value residuals. It was pretrained on 12T tokens and post-trained with mid-training, long-context training, SFT and RL.
Benchmarks and Measured Results
Self-reported scores include TerminalBench 2.1 80.5, SWE-bench Pro v2.0 Hard 79.1, GPQA Diamond 85.2 and AIME 2025 98.3. DeepSWE 1.1 at 44.3 trails Muse Spark Open's 60.0, and SWE-bench Verified at 76.5 trails Reflection Beam's 81.6. Some competitor scores come from Ricursive's own eval runs, and the company notes agentic evals are hard to replicate.
Getting Started for Developers
Download ricursiveAI/Ripple-65B-Preview (BF16) or the NVFP4 version from Hugging Face; Ricursive provides a Colab to help run it locally on devices such as the DGX Spark, and the DFlash2 draft model supports speculative decoding. The license is Apache 2.0, so commercial use is allowed.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.