Playbooks · Coding tasks
Ripple-65B: the $32M open model topping coding tests
Written November 5, 2026
Ricursive's 65B open-weight model claims the US open coding lead on two benchmarks. The claims, the price, and the context limit that isn't 1M.
What it is
Ricursive Intelligence released Ripple-65B Preview on November 4. It is a 65-billion-parameter dense Transformer — all 65 billion parameters active on every token, not a mixture of experts — pretrained from scratch on 12 trillion tokens, with support for up to 1 million tokens of context, under the Apache 2.0 licence. The weights are on Hugging Face as ricursiveAI/Ripple-65B-Preview, in three checkpoints: an instruction-tuned BF16 model, an NVFP4 version, and a BF16 base model. The model is a general-purpose variant of the proprietary model behind Ricursive's chip-design platform, built by a team whose founders previously helped create Google's AlphaChip. Ricursive says the whole research and development effort cost $32 million, covering early experiments, data, evaluations, and the final training runs, and estimates that at about a tenth of what the previous open-weight coding leader spent. That comparison rests on Ricursive's own estimate, not on comparable accounting from both companies.
The benchmark claims, and their limits
Every score here is Ricursive's own, from the model card, and none has been independently rerun. On Terminal-Bench 2.1 the card reports 80.52, just ahead of Reflection Beam at 80.1 and Meta Muse Spark Open at 76.4. On SWE-bench Pro v2.0 Hard it reports 79.1, ahead of Beam at 76.5. Those are the two tests behind the "US open coding lead" claim, and the claim covers selected US competitors and selected tests only — Reflection's own published comparison lists several Chinese-developed models above Ripple's Terminal-Bench result.
The same card shows Ripple trailing on other tests: DeepSWE 1.1 at 44.3 against Muse Spark Open at 60.0, and lower scores than Meta on the displayed automation and adversarial safety tests. Ricursive's own materials also disagree with each other on some rival scores: the card lists 52.3 for Meta and 2.0 for Inkling on SWE-bench Pro v2.0 Hard, while the launch graphic lists 72.4 and 39.6. The card says its tables combine published results with its own internal evaluations and acknowledges that reproducing agentic benchmark scores is difficult. Read the lead claim as a candidate for evaluation, not as proof the model will do a given team's coding tasks better.
The price
The weights are free to download under Apache 2.0, so there is no vendor list price. The clearest hosted rate is on OpenRouter, where the GMICloud endpoint charges $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.05. Together AI lists the model as coming soon, with no Ripple-specific rate published, and Fireworks has no confirmed Ripple-specific rate either — its general tier for unlisted dense models is not a Ripple price. The OpenRouter rate is the cost of that one access route, not a vendor rate card.
The context limit that is not 1M
Ricursive advertises a 1-million-token context window. The hosted reality is different, and OpenRouter's own pages disagree with each other. At launch, the OpenRouter listing showed a 16,384-token limit for the endpoint, while its FAQ states a 1,048,576-token context window. Do not assume the hosted price buys access to the advertised window: check the live listing before planning long-context work on the hosted route. Self-hosting the weights is the only route where the full window is in the reader's own hands, and even there it needs measuring — the card's long-context training used 2M-token sequences, but that is a training detail, not a measured serving result.
What to check before adopting
- Run your own tasks. No independent run of Ripple exists yet. Compare it against the current model on real repositories, recording completed tasks, errors, time, token use, and human intervention — not benchmark scores.
- Check the hosted context limit live. The OpenRouter listing and FAQ disagree. Confirm the limit on the endpoint before relying on long context.
- Price the whole job. Token prices are one part of the bill. Measure how much text each task generates, how often the agent retries, and the cost of the environment it runs tools in.
- Pick the right checkpoint. The release ships BF16 and NVFP4 instruction-tuned weights plus a BF16 base model. The base model is a pretrain-only checkpoint without the post-training, so it is not a drop-in for the tuned model.
Limits
- All benchmark scores are Ricursive's own, from the model card and launch materials. No independent rerun exists, and the card itself notes that changing the evaluation framework moved scores.
- No model run is described here. The 250 tokens per second on two B200 GPUs is the model card's figure, not a measured result, and local deployment memory and speed need separate measurement.
- The $32 million build cost is Ricursive's own figure for its full research and development bill. It is not a directly comparable accounting disclosure, and it does not establish the cost of running the model.
- For how to read benchmark claims, see BrokenML: a benchmark of broken machine-learning jobs. For the other open-weight launch of the week, see Kimi K3.1 Pro weights are out: what 1.56 TB buys.
Related playbooks
- TwIL-8B: a local model for math and coding
- DeepSeek's cheap coding model changed under the same name: retest Flash 1017 before you rely on it
- BotPsy 0.8B: an agent model that fits on-device
- Coding-agent traces in advisory fixes
Sources
- Ripple-65B Preview model card — Hugging Face
- Ripple-65B Preview pricing and providers — OpenRouter
- Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks — VentureBeat
Back to: Coding playbooks hub · Also: BrokenML: a benchmark of broken machine-learning jobs · Kimi K3.1 Pro weights are out: what 1.56 TB buys