Ricursive's Ripple-65B, from AlphaChip's Anna Goldie and Azalia Mirhoseini, edges past 500B-class U.S. open models on Terminal-Bench for $32M, but trails Beam on SWE-bench

Ricursive Intelligence, the chip-design lab founded by AlphaChip co-creators Anna Goldie and Azalia Mirhoseini, has released its first model: Ripple-65B Preview, a 65-billion-parameter dense Transformer trained from scratch, with weights on Hugging Face under Apache 2.0. Its headline claim is a score of 80.52 on Terminal-Bench 2.1, which the company calls "U.S. open-weight state-of-the-art" in agentic coding, just ahead of Reflection Beam 501B at 80.1. Its other headline claim is the price: $32 million for the whole project.

Both claims hold up as stated, and both come with qualifiers worth reading.

One benchmark, one country

"U.S. open-weight" is a deliberate filter. The comparison set is Reflection Beam 501B, Meta's Muse Spark Open 816B, Thinking Machines' Inkling 975B and NVIDIA's Nemotron 3 Ultra 550B. Open-weight models from Chinese labs aren't in the table, so the card says nothing about where Ripple stands against them.

Within that set, the "state-of-the-art in agentic coding" rests mostly on Terminal-Bench. The model card's own table shows the rest of the coding column is mixed:

Several of the competitors' numbers are Ricursive's own runs, marked in italics on the card, because labs publish different subsets of benchmarks. The card is unusually frank about how shaky that is: "Replicating benchmark scores has been challenging," it says, and it found Terminal-Bench 2.1 scores "universally lower (with the exception of Muse Spark Open)" under the official Terminus 2 harness at 128K context with summarization. Beam drops from its self-reported 80.1 to 70.8 there. Ripple shows 80.5 in both rows. If those reruns are right, Ripple's lead on that benchmark is larger than the headline 0.4 points. If they aren't, the lead is a rounding error. That is a case for independent runs, not a verdict either way.

What $32 million buys

The $32 million, per the launch post, is the full research and development cost from scratch: the scaling ladder, data mixing, evals, debugging, synthetic data, post-training recipes and the final runs. The card adds that the team "scaled from 1 to 5 people during development."

The "one-tenth the cost of the previous open-weight SOTA" comparison is softer. Ricursive's footnote bases it on reports that Reflection committed to $150 million a month of compute at SpaceX Colossus, and on Reflection's statement that Beam took four weeks to pretrain and four to post-train. That is an estimate built from a compute commitment, not a disclosed training cost, and the company says it is "estimated to the best of our knowledge."

The more checkable efficiency numbers are in the training recipe. Ricursive says reinforcement learning reached its Terminal-Bench score with 1.4 million rollouts, "roughly 70x fewer" than Beam, using only synthetic environments it built itself and no purchased post-training data. Pretraining ran on 12 trillion tokens of mostly English text and code at 8K sequence length, followed by 200 billion tokens of midtraining and two long-context stages at 2 million tokens.

A dense model, on purpose

The architecture is the most interesting part for anyone who runs models. Ricursive chose a dense Transformer over mixture-of-experts specifically to keep training infrastructure simple, and says it held 40% model FLOPs utilization with "simple sharding, no custom kernels, and no quantization."

The config shows what "thin and deep" means here: a 5,120-wide residual stream across 80 layers, but 80 attention heads of dimension 128, so attention is twice as wide as the model itself. That is the "overprovisioned" attention the card mentions. To pay for it, four of every five layers use sliding-window attention with a 1,024-token window, and only every fifth layer attends globally, without positional encoding (NoPE). Gated attention, value residuals and an FFN activation clamp at 20 round it out. The card describes the base as moving from Olmo 3's Llama-like design to a Gemma 3n-like one for stability, and Ricursive credits the Olmo team at the Allen Institute for AI (AI2), whose model ladder was its starting point. It says its final ladder improved compute efficiency on code by up to 16x over Olmo 3's at 4B scale.

The practical consequence of a dense 65B model: it fits on one NVIDIA DGX Spark with the provided NVFP4 checkpoint, and Ricursive reports 250 tokens/s on two B200s at batch size 8 before quantization. Every token touches all 65B parameters, though, while Beam activates 23B and Muse Spark Open 42B. On datacenter hardware, sparse rivals can be cheaper per token despite being much larger on disk.

The model is a general-purpose variant of the proprietary one behind Ricursive's chip-design platform, and the release calls itself a preview with "relatively light posttraining." For an independent check, the Terminal-Bench number is the one to watch. A 65B model that really matches 500B-class open models on agentic coding, on someone else's harness, would be a story about training efficiency. Today it is Ricursive's measurement.