Ripple-65B
A 65B-parameter dense open-weight language model released by Ricursive Intelligence under the Apache 2.0 license, focused on terminal agent tasks.
Timeline (last 90 days)
Ripple-65B is a 6.5-billion-parameter dense model released by a U.S. chip design company, its first publicly released large model.
The 4-bit quantized version of Ripple-65B can fit into a single NVIDIA DGX Spark personal device and run locally.
The releaser claims that Ripple-65B's entire development cost only 32 million USD and that the core team consists of just five people.
Across all 22 comparisons, Ripple-65B ranked first in only three, and lagged behind a 50-billion-parameter opponent on the bug-fixing benchmark test.
Ripple-65B's programming test scores are all self-reported by the releaser and have not yet been independently verified by a third party.
In terminal programming tests, Ripple-65B scored 80.52, surpassing an opponent model with 50.1 billion parameters.
Ripple-65B releases its weights open-source under the Apache 2.0 license, allowing free download, modification, and commercial use without paying licensing fees.
Ripple-65B was trained from scratch using 12 trillion tokens of text and code, with a context window reaching 1 million tokens.
The official team has simultaneously released preview weights and quantization files in the open-source community.
Ripple-65B is an open-source base model that topped the trending list in the open-source large-model community (the LocalLLaMA discussion forum).
13 more timeline events
All Facts (20)
Ripple-65B releases its weights open-source under the Apache 2.0 license, allowing free download, modification, and commercial use without paying licensing fees.
50%UnverifiedIn terminal programming tests, Ripple-65B scored 80.52, surpassing an opponent model with 50.1 billion parameters.
50%UnverifiedThe 4-bit quantized version of Ripple-65B can fit into a single NVIDIA DGX Spark personal device and run locally.
50%UnverifiedThe releaser claims that Ripple-65B's entire development cost only 32 million USD and that the core team consists of just five people.
50%UnverifiedAcross all 22 comparisons, Ripple-65B ranked first in only three, and lagged behind a 50-billion-parameter opponent on the bug-fixing benchmark test.
50%UnverifiedRipple-65B's programming test scores are all self-reported by the releaser and have not yet been independently verified by a third party.
50%UnverifiedRipple-65B is a 6.5-billion-parameter dense model released by a U.S. chip design company, its first publicly released large model.
50%UnverifiedRipple-65B was trained from scratch using 12 trillion tokens of text and code, with a context window reaching 1 million tokens.
50%UnverifiedRipple-65B adopts a "self-recursive architecture" and claims it can greatly reduce KV cache usage in long-range context scenarios.
50%UnverifiedThe specific technical details of Ripple-65B's "self-rule architecture" have not yet been elaborated in the original material, and its actual performance still needs community verification.
50%UnverifiedRipple-65B supports one-click seamless deployment with Olema and VLM.
50%UnverifiedThe video's author has compiled Ripple-65B's open-source weights and local running guide in the video description and pinned position.
50%UnverifiedRicursive did not provide Ripple-65B's throughput data in the model card.
50%UnverifiedA 65B dense model has higher inference-time VRAM usage and throughput costs than a MoE model of the same parameter count, and can serve fewer concurrent requests on the same hardware.
50%UnverifiedRipple-65B claims to achieve SOTA among US open-weight models on Terminal-Bench 2.1.
50%UnverifiedRipple-65B is a dense model that does not use a MoE architecture, and all parameters are activated on every inference.
50%UnverifiedRipple-65B has a parameter scale of 65B.
50%UnverifiedThe 65B in Ripple-65B points to a scale of roughly 65 billion parameters.
50%UnverifiedCompetition among open-weights models in the coding agent direction is heating up, and the emergence of Ripple-65B adds a US option to this track.
50%UnverifiedThe official team has simultaneously released preview weights and quantization files in the open-source community.
50%