Ripple-65B Preview Claims the Open Coding Lead — but Only on Two Benchmarks, and Only by Its Own Count

A new open-weights coding model claims the top spot on two major benchmarks — and the claim deserves both attention and scepticism in equal measure. Ricursive's Ripple-65B Preview, released on 4 November 2026 under an Apache 2.0 licence, scores 80.52 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard, according to its own model card. The lead over its nearest rival is about 0.4 percentage points. No independent lab has reproduced the results yet, and in its own 22-row comparison table, Ripple is the top scorer on only three rows.

Ripple-65B Preview is a 65-billion-parameter dense Transformer, trained from scratch on 12 trillion tokens, with support for up to 1 million tokens of context. It comes from Ricursive, a startup founded by former Google researchers, and it is available now on Hugging Face and through Together AI, GMI Cloud, and OpenRouter, with Fireworks support expected soon. The model is genuinely open — Apache 2.0 permits commercial use, modification, and redistribution — and genuinely small for its claimed performance: 65 billion parameters against rivals with hundreds of billions.

One correction before the details: the Hugging Face repository lives under the organisation name "ricursiveAI", not "ricursive-ai". The lower-case URL fails to open; the live model card is at huggingface.co/ricursiveAI/Ripple-65B-Preview.

Illustration created for INORYVA

What Ripple-65B is

Ripple-65B Preview is a dense model, not a Mixture-of-Experts design. Ricursive says it deliberately chose a modernised dense Transformer rather than MoE — a "thin-and-deep" architecture with 80 layers. The trade-off is straightforward: dense models are simpler to run and reason about, but they cost more compute per token at the same scale. At 65 billion parameters, Ripple is small enough to run on a single high-end server, which is part of its appeal.

The training story is the other half of the pitch. Ricursive trained Ripple from scratch on 12 trillion tokens of predominantly English text and code, at an 8,000-token sequence length, with support for up to 1,048,576 tokens of context — one million tokens. The preview release includes an instruction-tuned model in BF16 and NVFP4 numerics, plus a BF16 base model; the Hugging Face repository contains the BF16 instruction-tuned model. That is the version most users will download.

The cost claim is striking: Ricursive says the total cost, including preliminary training experiments, dataset preparation, evaluations, troubleshooting, synthetic data production, later training stages, and the final runs, was $32 million — against the hundreds of millions or billions estimated for other frontier models.

The benchmark claims — and how narrow the lead is

The headline numbers are strong. On Terminal-Bench 2.1, Ripple scores 80.52, 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. On SWE-bench Pro v2.0 Hard, Ripple scores 79.1, ahead of Beam at 76.5, Muse Spark Open at 72.4, Inkling at 39.6, and Nemotron at 30.7. Both are agentic coding benchmarks — tests of whether a model can actually complete software engineering tasks in a terminal, not just answer questions about code.

But the lead on Terminal-Bench is 0.42 points — 80.52 against 80.1 — which VentureBeat states explicitly as approximately 0.4 percentage points. That is within the noise of benchmark measurement: a different harness, a different sampling temperature, or a different run could flip it. The SWE-bench Pro lead is wider, 2.6 points, and more meaningful. But the overall picture is not a dominant victory; it is a narrow win on two benchmarks against a much larger model.

Benchmark	Ripple-65B	Reflection Beam 501B	Muse Spark Open 816B	Inkling 975B
Terminal-Bench 2.1	80.52	80.1	76.4	63.8
SWE-bench Pro v2.0 Hard	79.1	76.5	72.4	39.6

The wider table tells a more cautious story. Counting the 22 rows of Ricursive's own instruction-tuned comparison table — spanning general knowledge, maths, coding, and safety benchmarks — Ripple is the top scorer on only three rows: Terminal-Bench 2.1 self-reported, Terminal-Bench 2.1 under the Terminus 2 harness, and SWE-bench Pro v2.0 Hard. That is three rows, but only two distinct benchmarks, since Terminal-Bench appears twice under two harnesses. On the other 19 rows, a rival scores higher. Ripple is a coding specialist that leads on two coding benchmarks — not an all-round leader.

The verification problem: whose numbers are these?

Every number in the launch comes from Ricursive, and no independent lab has reproduced the results yet. That is not an accusation — it is the state of play as of 6 November 2026, reported by multiple outlets, and Ricursive itself is candid about it. The model card states: "Our table below shows official model card scores where available, and we use our own in-house eval flow for missing evals (italized)." It adds that the results "represent a best-effort attempt at a fair comparison" and that "replicating benchmark scores has been challenging, particularly for modern agentic evals."

That candour is welcome, and it is also the reason to wait before treating the lead as established. Rival scores in the table are partly Ricursive's own runs, not the rivals' published figures, and agentic benchmarks are notoriously sensitive to harness details. No independent leaderboard — Artificial Analysis, Vals.ai, or any other — had listed Ripple at the time of checking. Until one does, the honest description is "Ripple claims the lead on two benchmarks, by its own measurement" — not "Ripple leads".

Illustration created for INORYVA

Price and availability — and a context-window catch

Ripple is available now for download on Hugging Face and through selected partners: Together AI, GMI Cloud, and OpenRouter, with Fireworks support expected soon. On Together AI, the model is available on serverless and dedicated infrastructure at $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.02 per million. The OpenRouter endpoint via GMI Cloud charges the same $0.20 in and $0.60 out, with cached input at $0.05 per million. That is cheap for a model claiming frontier coding performance — roughly the price of a mid-tier model, not a flagship.

One catch: the context window depends on where you run it. The model supports up to 1 million tokens, but Together AI's own page lists a 128K-token context window in its specification field, even while the same page says "Context Window 1M", and OpenRouter lists 16,384 tokens. Do not present 1 million tokens as the context you will actually get from a hosted API — that figure is the model's claimed capacity, and the served limit is set by the provider.

There is also a small date discrepancy to note. Launch coverage dates the release to 4 November 2026; Together AI's page lists a release date of 3 November 2026. This article uses 4 November, the date in the launch coverage, and flags the difference rather than harmonising it.

For readers who follow open models through the Impeccable 96,000-star GitHub skill story, the Claude SEO free 24,866-star plugin coverage, and the Hallmark 41,000-star GitHub skill story, Ripple fits a familiar pattern: a small team betting efficiency beats scale. Its adoption will depend on whether independent evaluations confirm the benchmarks.

Illustration created for INORYVA

Key Takeaways

• Ricursive's Ripple-65B Preview is a 65-billion-parameter dense Transformer trained from scratch on 12 trillion tokens, released under Apache 2.0, with up to 1 million tokens of context claimed.
• It claims 80.52 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard — ahead of Reflection Beam 501B (80.1 and 76.5), Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B.
• The Terminal-Bench lead is only about 0.4 points. In its own 22-row comparison table, Ripple tops just 3 rows — 2 distinct benchmarks. On the other 19 rows, a rival scores higher.
• Rival scores in the table are partly Ricursive's own in-house runs, and no independent evaluation had been found as of 6 November 2026. Treat the lead as a vendor claim until an independent leaderboard confirms it.
• On Together AI it costs $0.20 per million input tokens and $0.60 per million output tokens — but Together serves a 128K context window, not the claimed 1M. OpenRouter lists 16,384 tokens.
• Ricursive says total training cost was $32 million (per VentureBeat's reporting of the claim). The Hugging Face repository is under "ricursiveAI", not "ricursive-ai" — the lower-case URL does not open.

Conclusion

Ripple-65B Preview is the most interesting open-weights release of the week, and the most in need of independent verification. A 65-billion-parameter model claiming to beat 500- and 800-billion-parameter rivals on agentic coding benchmarks, trained for a reported $32 million, would be a genuine shift in what small teams can build — if the numbers hold. But the lead is 0.4 points on one benchmark, the model tops only 2 distinct benchmarks in its own table, and every figure comes from Ricursive itself, with rival scores partly run in-house. The weights are open and the price is low, so the community can now reproduce the benchmarks and find out whether the lead is real. Until then, Ripple is a claim worth watching, not a crown worth awarding.

Sources

• Hugging Face — Ricursive Ripple-65B-Preview model card — https://huggingface.co/ricursiveAI/Ripple-65B-Preview
• Together AI — Ripple-65B Preview model page — https://www.together.ai/models/ripple-65b-preview
• VentureBeat — "Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks" — https://venturebeat.com/technology/ex-googlers-new-apache-2-0-model-ripple-65b-preview-claims-the-u-s-open-coding-lead-on-two-major-benchmarks

Facts checked against the Hugging Face model card, Together AI, and VentureBeat on 6 November 2026. The $32 million cost figure is Ricursive's claim as reported by VentureBeat; the launch blog was not opened directly. No independent evaluation of Ripple had been found as of 6 November 2026.