A $4 billion chip-design startup gave away its AI model. Most downloads are of someone else's copy

Ricursive Intelligence, a Palo Alto startup valued at $4 billion for using AI to design computer chips, has given away a general-purpose AI model anyone can use commercially. Four days after release, the strongest sign of demand is a copy the company did not make.

Hugging Face, the main site for sharing AI models, counts downloads over the past 30 days, which here covers only the days since launch. On November 8 it showed a repackaged version of the model, Ripple-65B Preview, at 58,957. That is up 79% from 32,913 a day earlier, which was itself up from 12,544 the day before. Ricursive's own main repository showed 3,803, about one-fifteenth as many. Its other official versions add a few thousand more.

The copy comes from AtomicChat, the team behind an app for running AI models on a user's own computer. It compresses the model so it fits on laptops and desktop machines, trading some precision for size. Ricursive does not test community conversions like it, according to a review by eesel. Many people trying Ripple on their own hardware are likely running a version its maker has not vouched for.

Why a chip company is giving away a model

Ricursive released Ripple on November 4 under Apache 2.0, a standard license that allows commercial use, changes and redistribution. It carries no royalties and no cap on a user's revenue or size.

The company was founded by CEO Anna Goldie and CTO Azalia Mirhoseini, former Google researchers who created AlphaChip, an AI method for laying out chips. The startup says that work has been used in four generations of Google's custom AI chips. In January it raised a $300 million Series A at a $4 billion valuation, led by Lightspeed, with Nvidia's venture arm among the investors.

Ricursive calls Ripple "a general-purpose variant of the proprietary model" behind its chip-design platform. In other words, the free model is a showcase; the paid product stays closed. Ricursive publishes no price list of its own for Ripple and points developers to outside hosts. VentureBeat notes the coding scores say nothing about how well the chip-design platform works.

The more telling claim is cost. Ricursive says Ripple cost about $32 million in total, including experiments, data work and the final training runs. That is roughly a tenth of the $335 million it has reportedly raised. It also estimates the figure is a tenth of what Reflection AI spent on its larger rival model. That comparison rests on press reports of Reflection's computing commitments, and VentureBeat cautions that it is not like-for-like accounting.

What the benchmarks do and don't show

Ricursive's headline number is 80.5 on Terminal-Bench 2.1, a test of whether an AI can finish multi-step tasks on a computer's command line. Reflection's Beam, with 501 billion parameters, scored 80.1, by Ricursive's count. Meta's 816-billion-parameter Muse Spark Open scored 76.4. Parameters are the internal settings a model learns; more usually means costlier to run.

That 0.4-point lead comes with heavy caveats. Every score is self-reported, and no outside lab had rerun them as of November 5. BenchLM, which tracks model scores, keeps Ripple out of its rankings because all its rows are company-run. Ricursive's own model card says agent tests have been hard to replicate and calls its comparison "best-effort." VentureBeat found figures that differ between the card and the launch graphic.

The "state of the art" claim also covers only U.S. open models. Reflection's own published comparison places several Chinese-built models above Ripple on the same test, VentureBeat reported. Ripple trails Meta's model on a separate coding test, 44.3 to 60.0, and on an adversarial safety test, 83.8 to 94.2. By eesel's count, it leads on just two of the 22 benchmarks in Ricursive's own table.

Rivals are not standing still. OpenRouter's public usage rankings showed Reflection's Beam up 33% on November 7, to 100.4 billion tokens (word fragments, the unit AI services bill by) processed.

What it costs to run

Ripple is a "dense" model: all 65 billion parameters work on every word it produces. Beam activates only 23 billion of its 501 billion at a time. Dense models are simpler to deploy, but each answer can demand more computing than a selectively activated rival.

The one hosted option at launch, on OpenRouter via GMI Cloud, charges $0.20 per million tokens of input and $0.60 per million for output. That is not the cheapest choice. Together AI lists OpenAI's open GPT-OSS-120B at $0.15 and $0.60. At launch, the hosted route also capped how much text the model could consider at once at 16,384 tokens, less than 2% of the advertised one million; OpenRouter's listing now shows the full million, but buyers should test before relying on it. Together and Fireworks have promised support but published no Ripple price.

For running it in-house, the full model is about 130 gigabytes. An experimental compressed version, about 57 gigabytes, fits on Nvidia's DGX Spark, a desktop machine with 128 gigabytes of memory. Ricursive's script for that box handles one request at a time and caps working memory at about 128,000 tokens. Its full million-token setup uses two Nvidia B200 data-center chips. Loading the official files with the popular Transformers software also runs code shipped with the model, which security teams will want to review first.

For buyers, the permissive license and modest hardware needs make Ripple cheap to trial. But the case for it rests on company numbers, a preview label and fast-spreading copies its maker has not checked. The next tests are independent benchmark runs, a promised technical report and confirmed pricing from the larger hosts.