Ricursive's Ripple-65B Preview Is a Dense 65B Open Model From a Chip-Design Startup — Explained Honestly

Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 5 November 2026. We have not tested Ripple-65B Preview hands-on, and nothing here is a hands-on finding.

A chip-design startup has just released a general-purpose AI model, and the headline numbers are loud enough to deserve a careful read. On 4 November 2026, Ricursive Intelligence released Ripple-65B Preview: a 65-billion-parameter open-weight model, released under the Apache 2.0 licence, pretrained from scratch on 12 trillion tokens, with up to 1 million tokens of context claimed.

That last sentence already contains three things worth separating. The model is dense, meaning all 65 billion parameters are active — not a mixture-of-experts system where only a fraction fires per token, so its parameter count should not be compared directly with a much larger MoE model's total. The weights are genuinely open, under a permissive licence. And this is a preview, not a final model — a distinction that matters for everything that follows.

What it is

Ripple-65B Preview is a general-purpose variant of the proprietary model behind Ricursive's chip-design platform. That origin story is the most interesting thing about it. Ricursive is a chip-design startup, and its founders, Anna Goldie and Azalia Mirhoseini, previously worked on Google's AlphaChip — the reinforcement-learning system for chip floorplanning. A model family built to reason about chip design, generalised into a coding and agentic model, is a plausible story: chip layout is, at heart, a long-horizon reasoning and optimisation problem, and the same machinery transfers to terminals, codebases and tool use.

The claimed research and development cost is around $32 million. That figure is Ricursive's own claim, reported in the coverage, and it is worth noting because it sits far below the headline training budgets of frontier labs — if it holds up, it is a data point in the ongoing argument that serious models no longer require frontier-lab-scale spending. But it is a claimed figure, not an audited one, and this article treats it as such.

The headline claim — labelled as self-reported

The headline claim, self-reported by Ricursive and not yet independently verified, is a score of 80.52 on Terminal-Bench 2.1, claimed to be the top score among US open-weight models, plus a claimed lead on SWE-bench Pro v2.0 Hard among US open models.

Read that sentence again, slowly, because the qualifiers do all the work. "Self-reported" means Ricursive ran the benchmarks and published the numbers; no independent lab has reproduced them yet. "US open-weight models" is a narrowed field — it excludes closed models and non-US open models, so it is not a claim to be the best model, or even the best open model, overall. And Terminal-Bench 2.1 is a specific version of a benchmark: do not compare this 80.52 directly against scores on Terminal-Bench 4.0, which is a different version of the test, scored differently. Cross-version comparisons are one of the easiest ways to mislead yourself with benchmark numbers, and this article will not do it.

What the claim does suggest, if it survives independent testing, is that Ripple-65B Preview is genuinely strong at terminal-based agentic work — the kind of long, messy, tool-using tasks that developers actually hand to coding agents: navigating a codebase, running commands, fixing what breaks, and finishing the job rather than just writing a plausible snippet.

What you can actually use today

The weights are on Hugging Face, under Apache 2.0 — that part is real and downloadable, not a promise. For hosted access, the model is available via API through Together AI, GMI Cloud and OpenRouter, with Fireworks support expected soon. The OpenRouter listing prices it at $0.20 per million input tokens and $0.60 per million output tokens. No other pricing is stated here, because no other pricing has been reported in the sources for this article — and inventing a price for Together AI or GMI Cloud would be exactly the kind of gap-filling this series refuses to do.

One context-window caveat, flagged by VentureBeat and worth repeating plainly: the advertised context is up to 1 million tokens, but the limit currently listed by OpenRouter is 16,384 tokens. Those two numbers describe different things — an advertised capability and a gateway's current listing — and until the hosted providers align on the full window, treat the 1-million-token figure as claimed, not as what you will get through every API today.

Why it matters

Three things make Ripple-65B Preview worth attention even before independent verification. First, density at this scale with a permissive licence is genuinely useful: a 65-billion-parameter dense model is large, but it is a size that well-resourced teams and labs can actually run and fine-tune, unlike frontier-scale MoE systems that exist mostly as API endpoints. Second, the chip-design lineage is a real differentiator — if the reasoning machinery that designs chips transfers to agentic coding, this model family could punch above its parameter count on exactly the tasks developers care about. Third, the open-weight field keeps getting more crowded and more credible: each release like this narrows the gap between what you can download and run yourself and what only a closed lab can offer.

Who it matters for

• Developers building coding agents. Terminal-Bench is your benchmark family, and a claimed top US open-weight score there is worth watching — but wait for independent runs before rebuilding your stack around it.
• Teams that self-host or fine-tune. Apache 2.0 weights on Hugging Face are the practical starting point. Check the actual download, the licence file and the model card before assuming anything about commercial use.
• Anyone tracking the open-weight race. Ripple-65B Preview is another data point that the race is no longer just about parameter count: training data scale (12 trillion tokens, claimed), architecture choice (dense, not MoE) and task-specific lineage (chip design) all matter as much as size.
• Budget-conscious API users. The OpenRouter listing of $0.20 / $0.60 per million tokens is competitive if the quality claims hold — but price per token means nothing without quality per token, and quality here is still unverified.

The limits — stated plainly

• This is a preview, not a final model. Expect rough edges, expect changes before any final release, and do not treat preview behaviour as a settled product.
• All benchmark claims are self-reported, with no independent verification yet. The 80.52 on Terminal-Bench 2.1 and the SWE-bench Pro v2.0 Hard lead are Ricursive's numbers, in a narrowed field (US open-weight models).
• Do not compare the Terminal-Bench 2.1 score directly against Terminal-Bench 4.0 scores. Different versions, different tests.
• The advertised 1M-token context exceeds the 16,384-token limit currently listed by OpenRouter, as VentureBeat noted. Hosted reality may lag the advertised figure.
• Ripple trails larger open models on some other benchmarks, also noted by VentureBeat. The headline is a terminal/agentic story, not an across-the-board lead.
• Fireworks support is expected soon, not live. No date has been given in the sources for this article.
• No pricing beyond the OpenRouter listing above is stated, because no other pricing was reported. No figures have been invented.
• No hands-on testing was performed for this article, and no benchmarks or prices have been invented.

The bottom line is a measured one. Ripple-65B Preview is a genuinely interesting release: a dense, permissively licensed, downloadable 65-billion-parameter model from an unexpected corner — chip design — with a loud, self-reported agentic claim. The claim is worth excitement and worth scepticism in equal measure. Try the weights or the API if you are curious, run your own tasks against it, and judge it when independent testers have published their numbers. Until then, treat Ripple-65B Preview as exactly what its name says: a preview with a promise, not yet a proven result.

Sources: MarkTechPost, 4 November 2026; VentureBeat; CryptoBriefing, 5 November 2026; OpenRouter model page for Ripple-65B Preview. Benchmark and cost figures are labelled as self-reported or claimed in the article. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.