Ripple-65B Preview: open weights, evidence and auditability

Open Models collection · Documentary assessment

Model release: 4 November 2026 · Evidence reviewed: 12 November 2026 · Release status: Preview

Open weights make a model more accessible to examination. They do not settle whether its development can be reproduced or whether its explanations faithfully describe how it reaches a result. Ripple provides a useful case for examining those distinctions.

What the release contains

The official model card describes a dense Transformer with 65 billion parameters, 80 layers and 12 trillion pretraining tokens, concentrated on English and code. Its attention combines sliding windows with global attention without positional encoding. Light post-training emphasises coding and terminal use. The advertised context is 1,048,576 tokens; this is a capacity specification, not proof of equally reliable reasoning throughout that length.

The release family includes instruction-tuned BF16 and NVFP4 checkpoints, a base model and a DFlash2 auxiliary model for speculative decoding. These are different artefacts, not interchangeable performance measurements. The reported development team grew from one to five researchers; that statement does not establish the company’s total workforce. Release documentation.

The laboratory, its business and its resources

Ricursive Intelligence is an AI laboratory focused on systems for semiconductor design. Its company announcement identifies a Palo Alto headquarters and founders Anna Goldie and Azalia Mirhoseini. The initial financing was a US$35 million seed round led by Sequoia, announced on 2 December 2025. Company-issued launch announcement.

A US$300 million Series A, led by Lightspeed, was announced on 26 January 2026 at a US$4 billion post-money valuation. Named participants include Sequoia and NVIDIA’s venture arm, NVentures. Together, these two announcements describe US$335 million raised. They do not disclose current cash, revenue, runway or audited expenditure; valuation is not available funding. Company-issued Series A announcement.

Ricursive separately reports US$32 million in Ripple research and development, covering experiments, data work, evaluation, failed approaches and training, rather than only the final training run. This is the laboratory’s cost statement, not an independently audited account. Ripple announcement.

Chip design does not mean a proprietary processor trained Ripple

Ricursive presents the released model as a general-purpose variant of the model behind its proprietary chip-design platform. Its announcement says the architecture was designed for efficient training and serving on NVIDIA hardware. The reviewed material does not establish that Ripple was trained on a processor manufactured or developed by Ricursive, nor does it identify a complete training-cluster inventory or hosting contract. The private chip-design platform itself is not made open by releasing this checkpoint. Laboratory’s explanation.

The vLLM deployment recipe, updated on 10 November 2026, provides NVIDIA B200/B300 NVFP4 and AMD MI355X BF16 configurations and settings supporting up to 2,097,152 tokens. These are deployment recipes, not independent quality results at that context length. Memory requirements also depend on precision, context and concurrency; a checkpoint fitting a device does not guarantee practical long-context operation.

How open and auditable is it?

On small screens, scroll the tables sideways to read every column.

Release inspection: repository files · Legal terms: Apache 2.0 licence.

Open weights allow researchers to instrument and study a model more directly than an API alone. They do not automatically make a natural-language reasoning trace faithful. Research on chain-of-thought faithfulness shows why that property needs its own experiments. This is a general methodological caution, not a Ripple-specific experimental finding. Measuring faithfulness in chain-of-thought reasoning.

What has actually been measured?

Scroll the table sideways to read the result and its limitations together.

A substantive failure behind the aggregate score. Plotly reports that, after a query failed, Ripple presented twelve figures it had not calculated, without identifying them as estimates. This does not establish a general failure rate, but it prevents reading the aggregate score as assurance of trustworthy analysis. Reported failure.

These tests measure different things. Combining their scores, or comparing total parameter counts across dense and mixture-of-experts models, would not establish a general superiority ranking. An external evaluator can also have a commercial interest in its own test environment; independence from the model developer does not remove the need to inspect the protocol.

Scientific support and unanswered questions

The founders’ chip-design background includes AlphaChip, whose reinforcement-learning floorplanning work appeared in Nature in 2021. Google’s account connects that work to TPU layout design. This is relevant prior research, not a peer-reviewed validation of Ripple or evidence of a Ricursive-manufactured processor. Google DeepMind’s account and publication links.

The reviewed Ripple documentation provides a model card and launch report; no peer-reviewed Ripple study was identified in that material. The next useful evidence would be reproducible independent coding evaluations, fuller training-data provenance, long-context quality tests and model-specific investigation of reasoning faithfulness. Until then, open-weight coding model with partial auditability is a more defensible description than a fully transparent reasoning system.

Return to Open Models