VIDRAFT open-releases Darwin-180B-RSI-R3, a recursively self-improved MoE reasoning model

What was released

Korean AI startup VIDRAFT has published the weights of Darwin-180B-RSI-R3, a large language model tuned for mathematical and scientific reasoning. According to a dev.to write-up on the release, the model is a Mixture-of-Experts (MoE) design with roughly 180 billion total parameters, of which only about 3 billion activate per token at inference time. It started from Alibaba's Qwen3.8-Flash-Next checkpoint — listed simply as Qwen3.8-Flash in the accompanying benchmark table — and was then pushed through three rounds of what VIDRAFT calls Recursive Self-Improvement (RSI), which is what the R3 suffix denotes.

The weights are on Hugging Face, along with a 4-bit GGUF quantization that dev.to puts at roughly 111 GB. That rules out most consumer GPUs but makes CPU-only inference plausible on a workstation with plenty of RAM, and because the MoE routing activates so few parameters per token, throughput should be far better than a dense model of the same nominal size. The write-up is blunt that CPU generation is for experimentation rather than production throughput, and it advises confirming the exact license on the model card before any commercial deployment.

How the self-improvement loop works

The training method is the unusual part of the story. Rather than fine-tuning on human-written solutions or expert reasoning traces, the model attempts problems whose answers can be verified programmatically — formal mathematics and science questions with checkable ground truth. Only the solutions the model itself gets right are kept, and those verified outputs become the fine-tuning data for the next iteration. The improved model then seeds the following round, and the cycle repeats.

According to dev.to, no human annotation enters this loop; the entire training signal is the model's own verified output. The reported effect is large: the base Qwen3.8-Flash scored 57.89 on the SuperCLUE math and science evaluation, while Darwin-180B-RSI-R3 — three RSI rounds later, with no change in scale or architecture — reaches 73.68.

Where it ranks

On the SuperCLUE November 2026 math and science reasoning test, dev.to reports the following standings:

• GPT-6.1-Sol (OpenAI, closed): 76.32
• Darwin-180B-RSI-R3 (VIDRAFT, open weights): 73.68
• Gemini-4-Argon (Google, closed): 71.05
• DeepSeek-V4.1-Flash (DeepSeek, open weights): 71.05
• Grok-4.7 (xAI, closed): 68.42
• Qwen3.8-Flash (Alibaba, open weights, the base model): 57.89

That places Darwin second out of 19 models overall and first among all open-weight entries. It is also, per the post, the only Korean model in the rankings.

The specialization cuts the other way too. On SuperCLUE's five-domain composite index the model manages 43.08 — ninth overall and fifth among open-weight models — with dev.to explicitly flagging long-context handling and instruction following as weak spots relative to closed competitors. This is a reasoning specialist, not a general-purpose assistant.

How solid are the numbers

The benchmark figures come from VIDRAFT's model card and SuperCLUE results as relayed by dev.to, which credits a 7 November article by Japanese outlet ラボメモ as its origin. SuperCLUE's test set is described as non-public, and the scores have not been independently verified.

Why it matters

Two things make this release notable. First, an open-weight model now sits within roughly three points of the best closed system on a reasoning benchmark and ahead of the Google and xAI entries as listed — significant for anyone building reasoning-heavy applications who needs a deployable model rather than an API dependency.

Second, the RSI methodology matters beyond this single checkpoint. If a model can generate its own training data from verifiably correct work, improving it no longer depends on armies of expert annotators, who are scarce and expensive. A 16-point benchmark jump attributed to self-generated data alone, rather than added scale, is exactly the kind of result that invites replication. Until independent evaluations land, the prudent read is promising-but-self-reported — but since the weights are public, anyone with suitable hardware can check.