Ricursive's Ripple-65B Claims U.S. Open-Weight Coding Lead on $32M of R&D

TL;DR

Ricursive Intelligence, the chip-design lab built by the team behind AlphaChip, shipped its first model on November 4: Ripple-65B Preview, a 65B dense Transformer trained from scratch on 12 trillion tokens, with weights on Hugging Face under Apache 2.0. The launch post claims the U.S. open-weight lead on agentic coding: 80.52 on Terminal-Bench 2.1, ahead of Reflection Beam 501B at 80.1, Meta's Muse Spark Open 816B at 76.4, Thinking Machines' Inkling 975B at 63.8 and NVIDIA's Nemotron 3 Ultra 550B at 56.4. Full R&D cost from scratch, by a team that grew from one to five people: $32 million.

What shipped

Three repos went up with the announcement: the BF16 instruct model, an NVFP4 checkpoint aimed at local hardware, and a pretrain-only base model. License is Apache 2.0. Pretraining covered 12 trillion tokens of mostly English text and code, long-context stages ran at 2 million tokens, and the release supports up to 1M context. Hosted copies are live on Together AI and OpenRouter, with GMI Cloud alongside them and Fireworks promised for the same day.

The release calls itself a preview that went through "relatively light posttraining", focused on agentic coding. It is a general-purpose variant of the proprietary model that runs Ricursive's chip-design platform, which is the company's actual business.

Read the claim before you repeat it

Terminal-Bench 2.1 scores an agent doing real work in a shell.

Ricursive then reran the field itself under the official Terminus 2 harness at 128K context with summarization, and every rival's score moved. Beam's self-reported 80.1 fell to 70.8. Muse Spark Open rose from 76.4 to 78.7. Ripple reports the same 80.5 either way.

Under Ricursive's reruns the lead is ten points. Under the labs' self-reports it is 0.4, a rounding error. The model card admits the problem directly: "Replicating benchmark scores has been challenging, particularly for modern agentic evals."

The rest of the coding column is mixed. On SWE-bench Verified, Ripple's 76.5 sits behind Beam's 81.6. Ripple leads SWE-bench Pro v2.0 Hard at 79.1, where Ricursive's own runs put Muse Spark Open at 52.3 and Inkling at 2.0, a number low enough that it deserves its own investigation. On DeepSWE 1.1, Muse Spark Open leads at 60.0 to Ripple's 44.3. And "U.S. open-weight" is doing real work in the claim: no Chinese lab appears in the table, and the top objection in the Hacker News thread (224 points, 79 comments) was that Qwen models score higher on the same benchmark and were left out.

Dense on purpose

Every rival in that table is a mixture-of-experts. Ripple is dense, and the choice was infrastructure first: Ricursive says it held a 40% model FLOPs utilization through pretraining and posttraining with simple sharding, no custom kernels and no quantization. A five-person team cannot staff a custom-kernel MoE pipeline, so it did not build one.

The architecture is thin and deep: a 5,120-wide residual stream over 80 layers, with 80 attention heads of dimension 128, so attention is twice as wide as the model itself. Four of every five layers use sliding-window attention with a 1,024-token window; every fifth layer attends globally with no positional encoding. Gated attention, value residuals and an FFN activation clamp round it out, on a Gemma 3N-like base chosen for training stability at scale. Per token, every one of the 65B parameters computes: Beam activates 23B, Muse Spark Open 42B. On a memory-bound local box that is the whole game; in a datacenter, the sparse giants can still be cheaper per token despite their size on disk.

$32 million, and what it bought

The $32 million is the full R&D bill: pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipes, and the final hero runs. The "one-tenth the cost of the previous open-weight SOTA" comparison is Ricursive's own estimate, built from reporting that Reflection committed to $150 million a month of compute at SpaceX Colossus and Beam's published schedule of four weeks pretraining plus four weeks posttraining. The post flags it as "estimated to the best of our knowledge", which is the right amount of confidence for arithmetic on someone else's compute contract.

The checkable efficiency claims are in the recipe. Reinforcement learning reached the Terminal-Bench score on 1.4 million rollouts, which Ricursive says is roughly 70x fewer than Beam needed, using synthetic environments built in-house and no purchased posttraining data, with truncated importance sampling, score centering and a linear length penalty. Recipe work ran on a model ladder, small models testing each change before the big run, the way a kitchen scales a recipe by the teaspoon before betting a wedding cake on it. Starting from AI2's Olmo 3 ladder, Ricursive claims a 16x compute-efficiency gain on code at 4B scale, widening to 55x on the hero run. The pretrain-only base beats Poolside's Laguna M.1 225B on their evals and lands near DeepSeek V3.1 and V4 Flash on reasoning, math and code.

Who is Ricursive

The lab's site claims the team behind AlphaChip (Nature 2021) and best-paper awards at DAC 2023, DAC 2025 and ASP-DAC 2026, with hands-on time on Gemini, Claude, TPUs and Apple Silicon. TechCrunch reported in February that Ricursive raised $335 million at a $4 billion valuation within four months of founding. The business is an AI platform for chip design; the model that runs it got a general-purpose sibling, and the sibling is what you can download. A lab that designs custom silicon releasing a model sized for a desktop GPU box has an internal logic you can set your watch to.

Should you run it

• The NVFP4 checkpoint and its colab target a single NVIDIA DGX Spark, so local agentic coding on code that cannot leave the building is the obvious first use.
• Throughput is respectable for a dense 65B: Ricursive reports 250 tok/s on two B200s at batch 8 in BF16, before quantization or custom kernels.
• Hosted endpoints are on Together, OpenRouter and GMI Cloud if you would rather not own the GPUs.
• Calibrate expectations to the label. "Preview" with "relatively light posttraining" means a coding-first tool, and general assistant behavior may lag the benchmark column.
• Before you pick it over Beam on the strength of 80.52, wait for an independent Terminal-Bench run. The whole table is one lab's harness today.

Caveats

• Every benchmark number here is either a lab's self-report or Ricursive's own rerun, marked in italics on the model card where Ricursive supplied a missing score. No third party has published results yet.
• "U.S. open-weight" excludes the Chinese labs that dominate global open-model usage. The HN thread's most-upvoted point is exactly that omission.
• The one-tenth cost comparison is an estimate from a reported compute commitment, not a disclosed ledger from either company.
• It is a preview with relatively light posttraining, from a lab whose main product is chip design. Model behavior beyond agentic coding is unproven.
• Ricursive's own finding that Terminal-Bench scores move ten points with the harness cuts both ways, including for its own 80.5.

Key Takeaways

• Ricursive Intelligence released Ripple-65B Preview: 65B dense, pretrained from scratch on 12T tokens, 1M context, Apache 2.0, with BF16, NVFP4 and base checkpoints on Hugging Face.
• It claims the U.S. open-weight agentic coding lead at Terminal-Bench 2.1 80.52, ahead of Reflection Beam 501B (80.1), Muse Spark Open 816B (76.4), Inkling 975B (63.8) and Nemotron 3 Ultra 550B (56.4). Ricursive's Terminus 2 reruns put Beam at 70.8 and Ripple still at 80.5.
• The rest of the coding column is mixed: behind Beam on SWE-bench Verified (76.5 vs 81.6), ahead on SWE-bench Pro v2.0 Hard (79.1), behind Muse Spark Open on DeepSWE 1.1 (44.3 vs 60.0).
• Total R&D was $32M with a team that scaled from 1 to 5 people, 40% MFU with no custom kernels, and an RL climb of 1.4M rollouts, roughly 70x fewer than Beam. The one-tenth-cost claim is Ricursive's estimate.
• Dense means every token activates all 65B parameters, against 23B for Beam and 42B for Muse Spark Open: dearer per token in a datacenter, at home on a DGX Spark.
• The comparison set is U.S. open-weight only, with no Chinese labs in the table, and all numbers await independent replication.

Sources: Ricursive launch post, Ripple-65B Preview model card, Ricursive on X, Hacker News discussion, sysnote analysis, TechCrunch on Ricursive's raise, Reflection on Beam, Terminal-Bench, Allen Institute Olmo