Ricursive releases Ripple-65B Preview with Apache 2.0 weights and a low-cost API

Ricursive Intelligence released Ripple-65B Preview on November 4, adding a downloadable coding model to a market crowded with hosted agents. The 65-billion-parameter dense model comes with Apache 2.0 licensing and an API listing starting at $0.20 per million input tokens. The practical question is which version to use: public weights for a deployment you control, or a hosted endpoint whose limits and availability depend on its provider.

The launch announcement describes Ripple as a general-purpose relative of Ricursive’s proprietary chip-design model. Its coding results and development-cost comparisons come from the company, not independent ByteForward testing.

What is available to download?

The official Hugging Face repository contains 16 downloadable safetensors weight files and shows approximately 130 GB of repository storage. It also includes configuration, tokenizer, custom modeling code and inference files. These are public artifacts, rather than a promise to release weights later. We inspected the listing without downloading or running the model.

The included Apache 2.0 license permits use, modification and redistribution subject to its conditions, including preserving required notices and providing the license when redistributing. It is a permissive license suitable for commercial use, not a grant of trademark rights or a warranty that the model is fit for a particular application.

Ricursive’s model card identifies this repository as the BF16 instruction-tuned release. It links separate NVFP4 and base checkpoints. The lower-precision option is relevant to local deployment, but a smaller download does not by itself establish usable speed or memory requirements on your hardware.

For a contrasting access model, see our SWE-2 Flash guide. Cognition describes a future weight-access program for selected partners, while Ripple already has publicly listed files and a redistribution license.

API pricing, context and provider availability

At our November 5 check shortly after 00:00 UTC, OpenRouter’s live listing showed GMICloud as the sole provider, serving BF16 weights. Listed prices were $0.20 per million input tokens, $0.60 per million output tokens and $0.05 per million cached input tokens.

The live page specified a 1,048,576-token context window and up to 131,072 completion tokens. It listed native tool calling and JSON output, while noting that JSON-schema enforcement was unavailable. EU in-region routing was also unavailable. Those are the hosted listing’s advertised capabilities, not a guarantee that every long request will succeed.

Provider pages need checking separately. Although Ricursive names Together in its announcement, Together’s own model page still said the model was coming soon to its Serverless API when checked. Its specification listed 1M context, but that should not be mistaken for a live serverless endpoint.

How to read the coding benchmarks

Ricursive reports 80.52 on TerminalBench 2.1. Its model card combines other labs’ published results with Ricursive’s own evaluations where scores were missing, and explicitly warns that agentic results shift with the harness and configuration. Treat the comparison as the developer’s evidence, not a controlled independent leaderboard.

The same card reports 76.5 on SWE-bench Verified, below several comparison models in that table. Its headline coding result therefore does not establish superiority across software tasks. Parameter comparisons also need care: Ripple is dense, while the larger comparison models use mixture-of-experts architectures. Total parameter count alone is not a measure of inference cost.

The $32 million development claim

Ricursive says its full research and development bill was $32 million, including experiments, data work, evaluation, debugging and the final training runs. Its estimate that this was one-tenth the cost of a previous leading open-weight coding model relies on an inferred comparison, not audited accounts from both labs.

The company’s broader argument is that choosing an architecture around available hardware can simplify training and serving. That makes Ripple an interesting deployment candidate. It does not turn a reported development budget into a prediction of what your own inference will cost.

Who should try it?

Ripple is worth evaluating if you want permissively licensed weights or a low-priced coding endpoint. Start with representative repository tasks, check tool-call reliability and validate generated changes with tests. For local use, account for runtime overhead and context memory as well as the weights themselves. For hosted use, confirm current provider limits before planning a long-context workload.

The release remains a Preview, with Ricursive describing relatively light post-training focused on agentic coding. Public weights and attractive token prices make evaluation easier. Your own task results should decide whether it belongs in a production workflow.

Featured image: AI-generated editorial illustration by ByteForward, composed with the official Ricursive Intelligence wordmark from its model card. Conceptual artwork, not a photograph of Ricursive hardware. No endorsement implied.