Ricursive releases Ripple-65B Preview open-weight coding model

On November 4, 2026, AI startup Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter dense open-weight coding model published under the Apache 2.0 license. Founded by former Google AlphaChip creators Anna Goldie and Azalia Mirhoseini, Ricursive built the model for $32 million, claiming top U.S. open-weight performance on Terminal-Bench 2.1 (80.5) and SWE-bench Pro v2.0 Hard (79.1) while using 1.4 million RL rollouts.

Model release and deployment details

• ▪Ricursive Intelligence released Ripple-65B Preview in three checkpoints: an instruction-tuned BF16 version, an NVFP4 version, and a BF16 base model
• ▪Ricursive Intelligence released Ripple-65B Preview, an open-weight 65-billion-parameter coding model, under an Apache 2.0 license
• ▪Ripple-65B Preview is hosted via OpenRouter's GMICloud endpoint at $0.20 per million input tokens and $0.60 per million output tokens

Architecture and technical specifications

• ▪Ricursive Intelligence's Ripple-65B Preview features a dense Transformer architecture with 65 billion parameters across 80 layers and a model dimension of 5120
• ▪The experimental NVFP4 checkpoint of Ricursive Intelligence's Ripple-65B Preview quantizes MLP layers while keeping attention in BF16 to target local hardware such as NVIDIA DGX Spark
• ▪Ricursive Intelligence built Ripple-65B Preview using gated attention, global NoPE, value residuals, and feed-forward network activation clipping

Serving performance and context limits

• ▪OpenRouter's hosting endpoint for Ripple-65B Preview currently restricts context length to 16,384 tokens
• ▪On datacenter GPUs, the unquantized Ripple-65B Preview model from Ricursive Intelligence reaches 250 tokens per second on two NVIDIA B200 GPUs at batch size 8

Pretraining process and efficiency

• ▪Ricursive Intelligence pretrained Ripple-65B Preview from scratch on 12 trillion text and code tokens at an 8,000 sequence length
• ▪Pretraining of Ripple-65B Preview achieved roughly 40% model FLOPs utilization on NVIDIA hardware without custom kernels or quantization during training

Model architecture and pretraining baselines

• ▪Small-scale pretraining experiments by Ricursive Intelligence demonstrated up to a 16-fold compute efficiency improvement in code-prediction loss compared to an Olmo 3 baseline
• ▪Ricursive Intelligence shifted from Olmo 3's Llama-like layout to a Gemma 3N-like design for Ripple-65B Preview to improve model stability

Post-training and reinforcement learning

• ▪Post-training for Ricursive Intelligence's Ripple-65B Preview incorporated synthetic data pipelines, long-context stages at 2 million sequence length, and reinforcement learning using truncated importance sampling
• ▪Ricursive Intelligence reached target reinforcement learning performance for Ripple-65B Preview using 1.4 million rollouts

Benchmark performance across coding and math

• ▪Ripple-65B Preview scored 79.1 on SWE-bench Pro v2.0 Hard and 76.5 on SWE-bench Verified
• ▪On mathematics evaluations, Ripple-65B Preview scored 98.3 on AIME 2025 and 90.5 on the February 2026 HMMT benchmark
• ▪Ripple-65B Preview achieved a self-reported score of 80.5 on Terminal-Bench 2.1, topping reported U.S. open-weight coding models as of November 2026

Cost and efficiency comparisons to Reflection Beam

• ▪Ricursive Intelligence reported requiring roughly 70 times fewer reinforcement learning rollouts for Ripple-65B Preview than the 100 million rollouts reported for Reflection Beam
• ▪Ricursive Intelligence reported total research and development costs of $32 million to produce Ripple-65B Preview
• ▪Ricursive Intelligence estimated that producing Ripple-65B Preview cost approximately one-tenth as much as the previous open-weight coding leader

Data discrepancies and re-evaluations

• ▪Evaluation tables in Ricursive Intelligence's launch materials contained discrepancies with its published model card, such as listing Meta's SWE-bench Pro score as 72.4 in Ricursive Intelligence's launch graphics versus 52.3 in the model card
• ▪Re-evaluating Terminal-Bench 2.1 using the official Terminus 2 harness at 128K context lowered Reflection Beam's score from 80.1 to 70.8 while Meta Muse Spark Open reached 78.7

Company background and funding

• ▪Ricursive Intelligence announced a $300 million Series A funding round led by Lightspeed at a $4 billion valuation
• ▪Startup Ricursive Intelligence was founded by CEO Anna Goldie and CTO Azalia Mirhoseini, who previously helped create Google's AlphaChip chip design software