Darwin-180B-RSI 如何只修改 0.02% 權重更新 180B 參數模型

Darwin-180B-RSI: How a 180B-Parameter Model Was Updated by Modifying Just 0.02% of Its Weights

TL;DR: VIDRAFT's Darwin-180B-RSI is a 180-billion-parameter model that was refreshed through a self-improvement (RSI) loop by updating a claimed 0.02% of its parameters — leaving the vast majority of the model frozen. Beyond the training methodology, the model is notable for running on consumer-grade laptop SSD storage via a streaming inference approach that achieves a reported 4.17 tokens/second. If you care about parameter-efficient self-improvement or commodity hardware deployment of frontier-scale models, this one is worth watching.

What it is

Darwin-180B-RSI is VIDRAFT's 180-billion-parameter language model, updated via a Recursive Self-Improvement (RSI) cycle. The key claim from the source is surgical: rather than retraining the full model, the RSI update touched approximately 0.02% of the model's total parameters. The remaining ~99.98% of weights were kept frozen during the update pass.

The second headline result is a hardware deployment story. Darwin-180B-RSI has been demonstrated running on a laptop with SSD-based weight streaming, achieving 4.17 tokens per second — a throughput figure that would have been implausible for a model of this scale on consumer hardware even a short time ago.

Together, these two results position Darwin-180B-RSI at an interesting intersection: extremely sparse self-modification during training, and extremely lean resource requirements at inference.

How it works

Recursive Self-Improvement (RSI) at a high level

The RSI framing means the model participates in its own update cycle in some capacity — generating signals, critiques, or candidates that feed back into a targeted weight-update step. The 0.02% figure implies the update mechanism is highly localized: rather than broad gradient descent across all layers, only a small subset of parameters are identified as candidates for modification and updated.

Conceptually, this is adjacent to techniques like sparse fine-tuning, parameter-efficient fine-tuning (PEFT), and selective layer updating, but the RSI label suggests the selection or update signal itself originates (at least in part) from the model's own outputs rather than purely from an external labeled dataset. The exact mechanism for identifying which 0.02% of parameters to update is not disclosed in the source.

SSD streaming inference

Running 180 billion parameters on a laptop requires moving weights from storage into compute memory on demand rather than holding the full model in VRAM or RAM simultaneously. The 4.17 tok/s figure is what the source reports for this streaming configuration. This approach trades peak throughput for accessibility — making a frontier-scale model runnable without a multi-GPU server.

This is conceptually similar to memory-mapped weight loading or quantized offloading strategies already present in the open-source ecosystem (think llama.cpp-style CPU/GPU offloading), applied here to a model an order of magnitude larger than typical open-weight community models.

Benchmarks & results

Based on what the source reports, the concrete figures are:

• Parameter update scope: ~0.02% of 180 billion parameters modified during the RSI update cycle
• SSD streaming inference throughput: 4.17 tokens/second on a laptop

The source does not report standard leaderboard benchmark scores (e.g., MMLU, HumanEval, MT-Bench, or similar) in the excerpted content available. Qualitatively, the framing from iTechGuides positions this as a technically novel result for the efficiency of the update method and the accessibility of the deployment target, rather than primarily a capability benchmark story.

No additional public benchmark numbers are available from this source, and none will be invented here.

How to try it

The source article does not include explicit public access instructions — no Hugging Face repository URL, GitHub link, or API endpoint is confirmed in the available content. Based on VIDRAFT's previously established developer channels (consistent with their other Darwin-series releases), you may want to monitor:

• Hugging Face: https://huggingface.co/vidraft — VIDRAFT has used this organization page for prior model releases
• VIDRAFT's official channels for announcements about Darwin-180B-RSI availability

If and when a public model card or API is announced, standard access patterns for OpenAI-compatible endpoints would apply (e.g., pointing an OpenAI SDK client at the announced base URL with your API key). No specific endpoint, model name string, or access token format is confirmed public at this time, so no shell commands are included here to avoid directing you to an endpoint that may not yet exist.

Check the source article directly for the most current access information: itechguides.com/darwin-180b-rsi-...

FAQ

Q: Does updating only 0.02% of parameters actually constitute meaningful self-improvement, or is this just a fine-tuning pass?
A: That's the technically interesting question. The RSI label implies the update signal has a self-referential component — the model's own outputs contribute to determining what gets updated — which distinguishes it conceptually from a standard supervised fine-tuning pass on external data. Whether the 0.02% threshold represents a deliberate architectural constraint or an emergent result of the update method isn't specified in the source.

Q: Is 4.17 tok/s on a laptop actually usable for development or just a benchmark curiosity?
A: For interactive use, 4.17 tok/s is on the slower end — roughly comparable to a fast human reading speed — but it's functional for batch inference, offline experimentation, or scenarios where GPU server access isn't available. For a 180B-parameter model on consumer hardware, it's a meaningful engineering result regardless of whether it matches server-grade throughput.

Originally reported by iTechGuides (글로벌) (2026-10-05) — source article.