Ripple-65B Preview: The 65B Model That Beats 500B Rivals for $0.20

A five-person team says it built a frontier coding model for $32 million, then watched it outscore models eight to fifteen times its size. That is the claim behind Ripple-65B Preview, released November 4, 2026 by Ricursive Intelligence, a chip-design startup whose founders helped build Google's AlphaChip. The model is free to download under Apache 2.0, and the only confirmed hosted price is $0.20 per million input tokens. This review checks what the numbers actually say, what a session costs, and where the headline stops holding up.

Every benchmark in this article comes from Ricursive itself. No independent evaluation had surfaced as of November 6, 2026, and the model is not yet listed on Artificial Analysis. That does not make the numbers wrong, but it does mean you should read them as the company's case for its model, not as a settled verdict.

What Ripple-65B Preview actually is

Ripple-65B Preview is a 65-billion-parameter dense Transformer, trained from scratch on 12 trillion tokens of mostly English text and code. Dense means all 65 billion parameters run on every token, unlike the mixture-of-experts rivals it is compared against, which activate only a fraction of their weights per token. The model has 80 layers, a narrow model dimension of 5,120, and a context window of up to 1,048,576 tokens, according to its Hugging Face model card. Its knowledge cutoff is December 17, 2025.

The release ships as three checkpoints: an instruction-tuned model in BF16, an experimental NVFP4 build aimed at local hardware, and a BF16 base model that is pretrain-only. A DFlash2 draft model is also listed for faster serving. The license is Apache 2.0, which permits commercial use, modification, and redistribution with attribution. There is no revenue threshold or monthly-user cap that triggers a separate commercial agreement, which is the catch that has tripped up other "open" releases this year.

Ricursive is a chip-design lab first and a model lab second. The public model is a general-purpose variant of the proprietary model that powers the company's AI chip-design platform, not the platform itself. Its founders, CEO Anna Goldie and CTO Azalia Mirhoseini, previously helped create Google's reinforcement-learning chip design software, AlphaChip. The company announced a $300 million Series A in January 2026, led by Lightspeed at a $4 billion valuation, with participation from Nvidia's NVentures, DST Global, and Sequoia, according to VentureBeat.

One naming trap to avoid: Ricursive Intelligence is a different company from Recursive, the lab behind the Flint coding model. The two names look alike and both build small coding models, but they are unrelated.

The benchmark headline, and what it leaves out

The headline number is 80.52 on Terminal-Bench 2.1, which Ricursive says is the highest score among US open-weight models. Its table puts Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and Nvidia Nemotron 3 Ultra 550B at 56.4. On SWE-bench Pro v2.0 Hard, Ripple leads again at 79.1 against 76.5 for Beam. Those two rows are the whole case for the launch, and they are real entries in Ricursive's own table.

The lead is narrow. On Terminal-Bench 2.1, Ripple beats Beam by 0.4 points, which is a tie in any honest reading. The more interesting number is in the same table's second Terminal-Bench row, run on the official Terminus 2 harness: Ripple holds at 80.5 while Beam drops to 70.8. That is a near-10-point gap, but the Beam figure there is Ricursive's own evaluation run, not Reflection's published score, so it needs independent confirmation before you lean on it.

The rest of the table cuts the other way. Ripple trails Beam on SWE-bench Verified (76.5 against 81.6), trails Muse Spark Open on DeepSWE 1.1 (44.3 against 60.0), SciCode (39.4 against 58.5), and MCP Atlas (71.67 against 89.2). On GPQA Diamond it scores 85.2, behind all four rivals. On competition math it is strong, with 98.3 on AIME 2025 and 95.0 on AIME 2026.

Source: Ricursive Intelligence model card, November 4, 2026. All figures are vendor-reported; rival scores in the card are a mix of official model-card scores and Ricursive's own in-house evaluation runs.

Three caveats deserve space. First, the comparison covers US open-weight models only. It leaves out Chinese models entirely, and VentureBeat notes that Reflection's own published comparison lists several Chinese-developed models above Ripple's Terminal-Bench result. Second, Ricursive's launch graphic and its model card disagree on some rival scores: on SWE-bench Pro v2.0 Hard the launch page shows Muse Spark Open at 72.4 and Inkling at 39.6, while the model card has 52.3 and 2.0. Ripple's own numbers match on both pages, but the rival columns should be treated as approximate until the company reconciles them. Third, the model card itself warns that Terminal-Bench 2.1 scores came out lower for nearly every model when run on the official Terminus 2 harness at 128K context with summarization, so reproducing the headline number depends on the harness. The newer Terminal-Bench 4.0 leaderboard had not listed Ripple as of November 5.

The fair summary is narrower than the launch framing. Ripple is a 65B model that matches or beats 500B-plus rivals on two agentic coding benchmarks in the vendor's own testing, and trails them on several others. That is still a strong result for a model this size. It is not proof that it is the best coding model you can run.

What it costs

The weights are free. The hosted price is where the fine print starts, because only one provider has a confirmed public rate. OpenRouter, served by GMI Cloud, lists Ripple-65B Preview at $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.05 per million. At November 6, 2026 market rates (1 USD = 0.747 GBP = 1.408 CAD), that is about 15p input and 45p output per million tokens in the UK, or about C$0.28 and C$0.84 in Canada.

Together AI lists the model as coming soon, with no published Ripple-specific rate as of November 4. Fireworks had no Ripple-specific rate either; its pricing page publishes a general $0.90-per-million-token tier for unlisted dense models above 16 billion parameters, which should not be treated as a Ripple price. GMI Cloud is named as a launch partner, but its public pricing page focuses on rented GPU capacity and no direct Ripple rate was independently verified. Ricursive itself publishes no price and runs no first-party API, so the OpenRouter rate is the cost of that one access route, not a universal price for the model.

The context gap in that table is the real pricing catch. The model card advertises a 1-million-token window, but OpenRouter's endpoint currently supports 16,384 tokens. A coding agent that reads a whole repository can burn through 16,384 tokens before it has finished its first pass, so the $0.20 price does not buy the agentic coding the model was built for, at least not through this route today. Self-hosting is the only way to reach the full window right now.

Here is how a session of 10 million input tokens and 1 million output tokens prices out on OpenRouter's list prices, checked November 5, 2026:

Source: OpenRouter listings, checked November 5, 2026, via eesel.ai. Ripple's own rates were re-verified directly on its OpenRouter model page on November 6.

Ripple is $0.90 cheaper than Beam on that session, but with 64 times less context. For most API buyers that trade is not worth it. DeepSeek V4.1 Flash is the budget pick on this table at $0.95, though note its $0.045 input rate is OpenRouter's listing; DeepSeek's own rate card lists $0.15 per million input tokens off-peak, double at peak hours. Against proprietary models the gap is wider: Claude Opus 5.5 lists at $4 input and $20 output per million tokens, so a Ripple session costs roughly a twentieth of an Opus session on input, before you count the quality difference.

How to use it

The fastest route is the OpenRouter API. The endpoint is ricursiveai/ripple-65b-preview, served by GMI Cloud, and the API is OpenAI-compatible, so most SDKs work by swapping the base URL and model name. OpenRouter lists tool calling and structured JSON output support, with a 131,072-token completion cap. Two limits apply today: tool choice is restricted to none or auto, and the chat template has no reasoning on/off switch or effort levels, so every reply pays for a reasoning pass in <think> tags.

Self-hosting is where Ripple earns its place. The BF16 instruct weights are about 130 GB, which fits on a multi-GPU server; the model card's server guide uses SGLang with the DFlash2 draft model and a 1M context default, quoting 250 tokens per second on two B200 GPUs at batch size 8 before quantization. The experimental NVFP4 checkpoint is about 57 GB and is aimed at single-workstation hardware such as the Nvidia DGX Spark, running one request at a time at 128K context. Ricursive also links a Colab notebook for local experiments.

Three setup notes before you start. The Hugging Face repository carries a custom-code tag, so loading it may require trusting remote code. A community Q4 GGUF quant exists at about 39 GB, but stock llama.cpp, Ollama, and LM Studio cannot load it until an unmerged llama.cpp patch lands. And there is no official Apple Silicon build yet, despite a co-founder's post saying the model can run on a MacBook Pro.

What it cannot do yet

The Preview label is doing real work. Ricursive says the release has gone through relatively light post-training, focused on agentic coding, so the instruct model should be read as a first cut. That shows in the numbers: it is a coding-first model that trails its rivals on tool-use and science benchmarks, and its base model trails DeepSeek V4 Flash and Nemotron 3 Ultra on several traditional pretraining evaluations in Ricursive's own base table.

It is text-only and mostly English. The 12-trillion-token pretraining corpus is text and code, so there is no image or audio input. The hosted endpoint's 16,384-token context is the practical limit for API users today, and the one provider serving it showed 70% availability over the three days to November 6, which is normal for a day-two launch but not something to put in production yet. The $32 million training-cost claim is also worth reading carefully: it is Ricursive's full research and development bill from scratch, and the "one-tenth the cost" comparison is the company's own estimate built from a press report of Reflection's compute commitments, not an audited figure from both sides.

Ripple vs the rivals: who should pick what

Against the other US open-weight models, Ripple's case is size and license. It is the only one of the five that fits on a single workstation, and Apache 2.0 lets you ship commercial products on it without a royalty or a revenue threshold. Beam is the better pure API choice today: it costs $0.25 per million input tokens on OpenRouter, barely more than Ripple, and gives you the full 1M context window. Muse Spark Open leads on DeepSWE, MCP Atlas, and SciCode, but costs $1.25 per million input tokens, more than six times Ripple's rate. Inkling, the 975B model from Thinking Machines, is the most expensive of the group at $1.00 per million input tokens and scores lowest on Terminal-Bench 2.1 in Ricursive's table.

Against Chinese models the picture changes, because Ricursive's chart excludes them. Qwen3.8 27B costs $0.15 per million input tokens on OpenRouter and, by one community comparison, scores higher on Terminal-Bench 2.1 than Ripple. If your team can use Chinese-lab weights, Ripple's headline lead disappears. If your security team has ruled out models from companies outside US jurisdiction, Ripple is now the strongest US open model that fits on a single workstation, which is the market it is really aimed at.

For a coding agent on a budget, DeepSeek V4.1 Flash remains cheaper per token on OpenRouter and carries the full context window. For the biggest open models with the strongest tool-use scores, Meta's Muse Spark Open is the comparison to beat, and Thinking Machines' Inkling 2 Small shows what the same lab charges for a smaller, cheaper sibling. Ripple's niche is the middle: a permissively licensed model you can actually own and run, at a price that makes experimentation cheap.

Common mistakes

Do not quote 80.52 as a settled fact. It is Ricursive's number, on a benchmark where the company filled in rival scores itself and where the official harness moves most scores. Wait for an independent run before treating the lead as proven.

Do not plan around a 1-million-token context through the API. The model supports it, but the only confirmed hosted endpoint caps at 16,384 tokens. If your agent needs long context, budget for self-hosting or pick a rival with the full window on OpenRouter.

Do not call it "open source" without the qualifier. The weights are open under Apache 2.0, but the training data, the full post-training recipe, and the proprietary chip-design model it derives from are not public. Open-weight is the accurate term.

Do not confuse Ricursive with Recursive. Same-sounding names, both small coding models, different companies. And do not read the $32 million figure as a running cost: it is what Ricursive says it spent developing the model, not what it costs you to run it. Your cost is hardware or API tokens, and the BF16 weights need about 130 GB of GPU memory.

Is Ripple-65B Preview really free?

The weights are free to download under Apache 2.0, including for commercial use, with attribution. Running them is not free: you pay for GPU hardware, or for API tokens at $0.20 per million input and $0.60 per million output on OpenRouter.

Can I run Ripple-65B on a laptop?

Not on an ordinary laptop. The BF16 weights need about 130 GB of GPU memory. The experimental NVFP4 checkpoint at about 57 GB is aimed at high-memory workstations such as the Nvidia DGX Spark, and there is no official Apple Silicon build yet.

How does Ripple-65B compare to Claude Opus 5.5?

On price, Ripple is far cheaper: $0.20 per million input tokens against $4 for Opus 5.5. On capability, they are different classes of model. Opus 5.5 is a proprietary frontier model with much stronger general reasoning and tool use; Ripple is an open-weight coding model whose best results come on terminal-based coding tasks.

Is the 80.52 Terminal-Bench score trustworthy?

It is Ricursive's own reported score, and the company is upfront that rival scores in its table mix official results with its own evaluation runs. The lead over Beam is 0.4 points, effectively a tie, and no independent evaluation had appeared as of November 6, 2026. Treat it as a promising claim awaiting confirmation.

Where can I try Ripple-65B Preview?

The fastest way is the OpenRouter API at $0.20 per million input tokens, with a 16,384-token context limit. For the full 1-million-token window, download the weights from Hugging Face and self-host, or wait for Together AI and Fireworks to publish their rates.

Conclusion

Ripple-65B Preview is an interesting release: a small, dense, permissively licensed model that punches far above its size on two coding benchmarks, built by a five-person team for a claimed $32 million. The API price is low, but the 16,384-token hosted context and the vendor-only benchmarks are real limits. Try it on your own repositories before you switch anything to it, and run it locally if you want the model it actually is.

Sources

• Ricursive Intelligence, "Ripple-65B Preview" model card, Hugging Face, November 4, 2026
• Ricursive Intelligence, "Introducing Ripple-65B Preview, a Frontier Open-Weight Model," November 4, 2026
• OpenRouter, "ricursiveAI/Ripple-65B-Preview" API pricing and providers, accessed November 6, 2026
• VentureBeat, "Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks," November 2026
• eesel.ai, "Ripple-65B review: benchmarks, price, and who should run it," November 2026
• Depth Grid, "Ripple-65B Preview: 65B Dense Open Model, Benchmarks, Apache 2.0," November 6, 2026