Google’s AlphaChip creators say they built America’s best open coding AI for $32 million

Ricursive Intelligence, the chip-design start-up founded by the two researchers behind Google’s AlphaChip, has released Ripple-65B Preview, a free AI model it says is now the best American open-weight model at agentic coding. It says the whole thing cost $32 million to build, about a tenth of what the previous leader spent.

The weights are on Hugging Face under the Apache 2.0 licence, which allows commercial use, and the model is also available through Together AI, GMI Cloud and OpenRouter, with Fireworks to follow. Ricursive announced it on X on Wednesday. Every benchmark and cost figure below is the company’s own and hasn’t been checked independently.

A small model against giants

Ripple is a dense model with 65 billion parameters, trained from scratch on 12 trillion tokens, with a context window of up to 1 million tokens. Most big open models are mixture-of-experts designs that use only part of their parameters for each token. Ripple uses all of them every time, which Ricursive says gives strong reasoning and fast output, and it is small enough to run on Nvidia’s DGX Spark desktop machine. On two Nvidia B200 chips it reaches 250 tokens a second, the model card says. OpenRouter lists it at $0.20 per million input tokens and $0.60 per million output tokens, though with a 16,384-token context limit for now.

On TerminalBench 2.1, a test of AI agents doing coding work in a computer terminal, Ricursive puts Ripple at 80.52, just ahead of Reflection’s Beam (80.1) and above Meta’s Muse Spark Open (76.4), Thinking Machines’ Inkling (63.8) and Nvidia’s Nemotron 3 Ultra (56.4). Those rivals have between 501 billion and 975 billion parameters.

The lead isn’t across the board. In Ricursive’s own table, Beam beats Ripple on SWE-bench Verified (81.6 to 76.5) and SWE-bench Pro v1.0, and Muse Spark Open leads on most maths, science-coding and knowledge tests. Ricursive filled gaps in rivals’ published scores with its own runs and admits that “replicating benchmark scores has been challenging” for agentic tests. When it reran TerminalBench in the official Terminus 2 harness, Beam fell to 70.8 while Ripple held at 80.5.

$32 million, from scratch

Ricursive says the $32m is its full research and development bill, covering early experiments, data, evaluations, debugging and the final training runs, not just the last run. It sets that against Sam Altman’s remark that GPT-4 cost more than $100m to train, and against reports that Reflection committed to $150m a month for computing power on SpaceX’s Colossus.

It says it bought no post-training data, made its own synthetic data instead, and reached its TerminalBench score with 1.4 million reinforcement learning attempts, about 70 times fewer than Beam. It started from the training recipe of Olmo 3, the open model from the Allen Institute for AI, and credits open work from Prime Intellect, RadixArk and Nvidia.

The AlphaChip team

Ricursive was founded by Anna Goldie, its chief executive, and Azalia Mirhoseini, its chief technology officer, who created AlphaChip at Google, an AI system that lays out chips and helped design three generations of Google’s TPUs. Both were also early Anthropic employees. Within four months of launching, the company raised $335m at a $4bn valuation, including a $300m Series A led by Lightspeed, with Nvidia among its investors, TechCrunch reported in February. It is not the same company as Richard Socher’s Recursive, which released its Flint coding model last month.

Ripple is a general-purpose version of the private model behind Ricursive’s chip-design platform, and the company says it wants to “begin closing the loop between models and the hardware that fuels them”. Goldie described the payoff to TechCrunch:

We could design a computer architecture that’s uniquely suited to that model, and we could achieve almost a 10x improvement in performance per total cost of ownership.

Anna Goldie, CEO, Ricursive Intelligence, to TechCrunch

Still chasing China

The “US” in Ricursive’s claim matters. The strongest open-weight models still come from China, and by Reflection’s own figures Beam trailed the best of them at launch. On Wednesday Moonshot put the weights of Kimi K3.1 Pro on Hugging Face, and Tencent released Hy4. Both report the newer Terminal-Bench 4.0 rather than 2.1, so there is no like-for-like comparison: Moonshot says K3.1 Pro scores 65.7 on it, against 66.4 for Claude Opus 5.5. Ripple’s case is size and cost. K3.1 Pro has 2.8 trillion parameters and needs data-centre hardware, while Ripple fits on a desk.

Why it matters

If the figures hold up, a small team has matched American open models ten times its size for a fraction of the money, and given developers a permissively licensed coding model they can run locally. It also shows a chip-design company building its own frontier models, so that it can design the hardware they run on.

Sources: Ricursive Intelligence, “Introducing Ripple-65B Preview”; Hugging Face, ricursiveAI/Ripple-65B-Preview model card; Ricursive Intelligence on X; OpenRouter, Ripple-65B Preview; TechCrunch, “How Ricursive Intelligence raised $335M at a $4B valuation in 4 months”.