Ex-Googlers Release Ripple-65B, a 65B Open Coding Model That Beats Models 10x Its Size

Quick Facts

• Ricursive Intelligence released Ripple-65B Preview on November 4, 2026, under an Apache 2.0 license, pretrained from scratch on 12 trillion tokens.
• The model scores 80.52 on TerminalBench 2.1 and 79.1 on SWE-Bench Pro v2.0 Hard, beating U.S. open-weight rivals hundreds of billions of parameters larger.
• Ricursive built and trained the model for $32 million, one-tenth the cost of the previous open-weight state-of-the-art on agentic coding.

Ricursive Intelligence, an AI lab founded by former Google researchers, released Ripple-65B Preview on November 4, 2026. The 65-billion-parameter model claims the top spot among U.S. open-weight models on two agentic coding benchmarks, outperforming models that are 8 to 15 times larger.

The company was founded in 2025 by Anna Goldie and Azalia Mirhoseini, who previously co-led AlphaChip at Google. That system generated chip layouts in hours rather than the weeks required by human designers and helped design multiple generations of Google's Tensor Processing Units.

What the Model Does

Ripple-65B Preview is a dense Transformer model with 5,120-dimensional embeddings across 80 layers. Unlike sparse mixture-of-experts models, it activates all 65 billion parameters on every request. The team built the architecture to run efficiently on NVIDIA hardware, achieving 40% mean floating point utilization without custom kernels.

The model supports up to 1 million tokens of context and was trained with roughly 1.4 million reinforcement learning rollouts during post-training. That rollout count is about 70 times fewer than the number used to train Reflection Beam, the previous open-weight leader on agentic coding tasks.

Ricursive shipped four model variants: the BF16 instruct model, an NVFP4 quantized version labeled experimental, the BF16 base model, and a draft model for speculative decoding. All are available on Hugging Face and through API partners including Together AI, GMI Cloud, and OpenRouter.

Benchmark Results

On TerminalBench 2.1, an execution-based benchmark where agents navigate real Bash environments, write code, and manage system administration tasks, Ripple-65B Preview scored 80.52. That places it ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4.

On SWE-Bench Pro v2.0 Hard, the model scored 79.1, compared with Beam's 72.4 and Muse Spark Open's 39.6. Additional scores include 90.5 on Math HMMT February 2026, 83.8 on FORTRESS Adversarial, 44.3 on DeepSWE 1.1, and 36.4 on AutomationBench Public.

In its launch post, Ricursive described the result as surpassing "significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B."

Cost and Capital

Ricursive states the $32 million figure covers the full research and development cost, including pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, post-training recipe development, and the final training run. The company called this "one-tenth the cost of the previous open-weight SOTA on agentic coding."

Ricursive raised $335 million in total, split between a $35 million seed round and a $300 million Series A led by Lightspeed Venture Partners in January 2026. That round valued the company at $4 billion when it had roughly 10 employees. DST Global, Nvidia's NVentures, Felicis Ventures, 49 Palms Ventures, and Radical AI also participated.

Strategic Context

Ricursive's stated mission is to use AI to design better chips, then use those chips to build better AI, creating a self-reinforcing loop. CEO Goldie said at the Series A that "the pace of AI progress is dictated by hardware." CTO Mirhoseini, who is also an assistant professor of computer science at Stanford, described the company's goal as operating "at the Pareto frontier of intelligence and computational efficiency."

The team includes researchers from Google DeepMind, Anthropic, NVIDIA, Cadence, Apple, Stanford, MIT, and Harvard. Their published work spans AlphaChip in Nature in 2021, two DAC Best Paper awards in 2023 and 2025, and an ASP-DAC Best Paper in 2026.

Releasing a competitive coding model under an open Apache 2.0 license positions Ricursive to attract developer adoption before its chip products reach the market, giving the company a base of users and feedback ahead of its core hardware business.