Ricursive Launches Ripple-65B Preview: 65B Open Model Claims Lead Among US Peers in Coding

What happened

The model was released on Nov 4 on X and Hugging Face at the same time. Ricursive Intelligence was founded in late 2025 by Anna Goldie and Azalia Mirhoseini, the two researchers behind Google's AlphaChip; Goldie is chief executive and Mirhoseini is chief technology officer. Ripple-65B is a general-purpose version of the model behind the company's chip-design platform, launching with instruction-tuned BF16 and NVFP4 checkpoints, a BF16 base model and a DFlash2 draft model.

The architecture is dense rather than mixture-of-experts: all 65 billion parameters activate on every token, in an 80-layer, d_model-5120 thin-and-deep Transformer. Starting from an Olmo 3 baseline, the team iterated on the training recipe, lifting compute efficiency on code 16x over the Olmo 3 ladder and up to 55x for the full hero run; the team grew from 1 to 5 people during development and held 40% MFU through pretraining and post-training. Pretraining used 12 trillion text-only tokens, mostly English and code, at 8K sequence length, with context up to 1M tokens.

On results, the company reports 80.52 on Terminal-Bench 2.1, which it calls the top score among US open-weight models, ahead of Reflection Beam 501B at 80.1 and Meta Muse Spark Open 816B at 76.4; it reports 79.1 on SWE-bench Pro v2.0 Hard. The model card adds that the Preview release went through relatively light post-training focused on agentic coding, that the scores are company-reported, and that no independent re-run exists yet. At 65B the model fits on consumer hardware such as the NVIDIA DGX Spark, with the NVFP4 checkpoint at about 57.2 GB; on two B200s it reaches 250 tokens per second at batch size 8, before quantization or custom kernels. The company puts total R&D at about $32 million, covering the scaling ladder, data work and the two final runs.

Key facts

1. Publisher and date: Ricursive Intelligence, Nov 4, 2026; the company's first model
2. Founders: Anna Goldie and Azalia Mirhoseini, the researchers behind Google's AlphaChip; the company was founded in late 2025
3. Scale: a 65B-parameter dense Transformer, all parameters active on every token; 80 layers, d_model 5120
4. Training: pretrained from scratch on 12 trillion tokens; context up to 1M tokens
5. License: Apache 2.0, with no usage thresholds
6. Benchmarks (company-reported): 80.52 on Terminal-Bench 2.1; 79.1 on SWE-bench Pro v2.0 Hard; 76.5 on SWE-bench Verified
7. Cost (company figure): total R&D of about $32 million
8. Runnability: the NVFP4 checkpoint is about 57.2 GB and fits on a DGX Spark; a BF16 base model and DFlash2 draft model are also available; Together AI, OpenRouter and GMI Cloud offer hosted API access
9. Caveat: the model card says the Preview release went through relatively light post-training; scores are company-reported with no independent re-run yet