Ripple-65B opens its coding model weights, with a smaller local version

Ricursive Intelligence released downloadable weights for Ripple-65B Preview on November 4, giving developers a coding-focused language model they can run on their own infrastructure. The 65-billion-parameter release includes a BF16 instruction-tuned checkpoint and a smaller NVFP4 version, alongside access through hosted model services.

A check of the public file listings found 16 BF16 weight shards totaling about 129.6GB and six NVFP4 shards totaling about 57.2GB, in decimal units. Those are download sizes: running memory also has to accommodate context state and other overhead. The repositories include configuration, tokenizer and conversation-format files as well as the weights. Both are ungated, and their file listings opened without signing in.

The NVFP4 checkpoint quantizes the feed-forward layers while retaining attention in BF16. Its model card labels the release experimental. Ricursive names NVIDIA DGX Spark as a local deployment target, but its published coding comparison uses BF16 models. The mixed-precision download therefore needs to be assessed separately from the headline benchmark result.

Its configuration also excludes the input embeddings and output head from quantization. The 57.2GB total therefore measures the actual mixed-precision files. Multiplying every parameter by four bits would understate the storage required for this particular distribution.

OpenRouter’s live listing, checked on November 5, showed a 1,048,576-token context window and a maximum completion length of 131,072 tokens. The context window covers the material a request and its continuation can accommodate; the completion limit caps generated output.

The coding lead has a defined comparison group

Ricursive reports 80.52 on Terminal-Bench 2.1, compared with 80.1 for Reflection’s Beam 501B. Its leading-model claim covers a selected group of US open-weight systems. The ordering changes on SWE-bench Verified, where the table lists Ripple at 76.5 and Beam at 81.6.

The company says comparisons use BF16 and maximum effort settings. The table combines developers’ published scores with internal measurements for missing results. Ricursive notes that changing the agent harness can shift scores. These conditions limit what the ranking establishes about other deployments.

Ricursive puts the full research and development bill at $32 million. It includes small-model experiments, data mixing, evaluations, debugging, synthetic data generation and development of the post-training recipe, as well as the final training runs. Ripple is a general-purpose variant of the proprietary model behind its chip-design platform.

The company also calls that roughly one-tenth the cost of the previous open-weight coding leader. That estimate uses reported monthly compute commitments and published training duration, rather than audited expense categories.

OpenRouter listed GMICloud as the provider at $0.20 per million input tokens, $0.60 per million output tokens and $0.05 for cache reads. It supports tool calling and JSON output, without JSON-schema enforcement. These are hosted API rates; self-hosting expenses depend on the hardware and operation.

The repository includes the Apache 2.0 license, allowing use, modification, redistribution and commercial use under its terms. Redistribution requires a license copy, notices identifying modified files, and preservation of applicable copyright and attribution notices. The license does not grant trademark rights. The main release retains the Preview name, with Experimental displayed on the NVFP4 checkpoint.