Models
Ricursive releases Ripple-65B Preview, an Apache 2.0 open-weight model for agentic coding derived from its chip-design model
Ricursive released Ripple-65B Preview on Nov. 4: 65B dense, Apache 2.0, aimed at agentic coding, not chip design. Benchmarks are self-reported.
On November 4, 2026, Ricursive Intelligence, a lab that builds AI for chip design, released Ripple-65B Preview: a 65-billion-parameter dense model with open weights under the Apache 2.0 license. The model is not itself a chip-design tool. Ricursive describes it as a general-purpose variant of the proprietary model behind its chip-design platform, with post-training focused on agentic coding. Its headline benchmark result is self-reported, and "open-weight" is not the same as "open source": the training data was not released.
The essentials
- What it is: 65B dense, pretrained from scratch on 12 trillion tokens, mostly English and code, text only, with a claimed context of up to 1 million tokens.
- The license: Apache 2.0 on the weights, with BF16 instruct, experimental NVFP4, base and draft-model checkpoints on Hugging Face.
- The claim: 80.5 on Terminal-Bench 2.1, presented as the best among US open-weight models. It is Ricursive's own result.
- The correction: it is a general-purpose, coding-oriented model, not a model for designing chips.
- The cost claim: $32 million of full R&D, per Ricursive, which is not audited.
What is Ripple-65B Preview?
It is the first public model from Ricursive, a Palo Alto lab co-founded by Anna Goldie and Azalia Mirhoseini, who previously worked on Google's AlphaChip. Ricursive chose a dense architecture over mixture-of-experts to simplify training and serving. The model always reasons in think tags, with no switch to turn that off. Post-training is described as relatively light, and the release is a preview.
How strong are the benchmark claims?
On Terminal-Bench 2.1, Ricursive reports 80.5, about a tie with the 501B Reflection Beam model at 80.1, and ahead of Meta's Muse Spark Open. The claim is limited to US open-weight models: some Chinese open models are reported to score higher, though figures differ between trackers. Every result is self-reported, and the model card itself says Terminal-Bench scores were generally lower on the official harness. Ricursive's own comparison shows Beam ahead on several tests, including SWE-bench Verified (81.6 against 76.5). The launch chart and model card also disagree on some competitors' scores.
Is it really open source?
The weights are open under Apache 2.0, which permits commercial use. The coverage we read does not say that the training data, training code or full recipe were released, so "open-weight" is the accurate term. Check the license file and model card before building on it.
Can you run it, and what does it cost?
In BF16, the model needs about two B200 GPUs. The NVFP4 checkpoint, around 57 GB, is reported to run on a DGX Spark, one request at a time at 128K context. Hosted prices of $0.20 per million input tokens and $0.60 per million output tokens appear on Together AI and OpenRouter. Context is claimed up to 1M tokens, but one hosted endpoint was reported capped at 16,384.
What is not established?
- Independent verification of the 80.5 score, the harness and the other benchmark claims.
- The cost: the $32 million figure and the "one-tenth" comparison are Ricursive's own estimates.
- Whether the 1M-token context holds in practice.
- Whether the chip-design lineage actually helps coding performance.
- What was released beyond the weights: training data and code were not reported as open.
Summary table
| Point | What is stated as of November 6, 2026 | Source |
|---|---|---|
| Release | November 4, 2026, preview | Together AI, DEV, eesel |
| Size | 65B dense, 12T tokens, text only | BenchLM, Together AI |
| License | Apache 2.0 weights | Together AI, BenchLM |
| Headline score | 80.5 on Terminal-Bench 2.1, self-reported | eesel, DEV |
| Cost | $32M R&D, per Ricursive, unaudited | eesel, DEV |
| Chip design | Variant of the model behind its platform; released model is for coding | Together AI, DEV |
Frequently asked questions
Is Ripple-65B a chip-design model?
No. It is a general-purpose variant of the model behind Ricursive's chip-design platform, post-trained for agentic coding.
Is Ripple-65B open source?
Its weights are under Apache 2.0, so it is open-weight. Training data and code were not reported as released.
Is it the best open model?
Only among US open-weight models, per Ricursive's own benchmark. Some Chinese open models are reported to score higher.
Method and limits
Sources as of November 6, 2026. We read the Together AI model page, BenchLM's profile and reviews from eesel and DEV Community, plus a launch explainer from Digital Applied. We could not open MarkTechPost's article and did not read Ricursive's launch post or model card directly. We downloaded and ran nothing.
Sources
- Together AI: Ripple-65B Preview (verified 2026-11-06)
- BenchLM: Ripple-65B Preview benchmarks and context (verified 2026-11-06)
- eesel: Ripple-65B review (verified 2026-11-06)
- DEV Community: Ricursive's Ripple-65B Preview, explained honestly (verified 2026-11-06)
- Digital Applied: Ripple-65B Preview, a new US open-weight model (verified 2026-11-06)
Spotted an error in this article? Report it · Our corrections policy
