Darwin-180B-RSI Leads MMLU-Pro and GPQA Diamond on AIhoster Open-Model Board

·1 views

VIDRAFT's Darwin-180B-RSI, a 180-billion-parameter open-weight model, ranks first on the MMLU-Pro and GPQA Diamond leaderboards tracked by the independent benchmark aggregator AIhoster. It scored 88.1% on MMLU-Pro and 94.4% on GPQA Diamond, ahead of several models with substantially more parameters. AIhoster, a Russian site, compiles publicly reported results from model cards and papers without recalculating or combining them into a composite score, as of its October 19, 2026 update. The model was evaluated with a thinking budget of up to 131,072 tokens, and GPQA results used majority voting over up to 16 samples. VIDRAFT, a Korean Pre-AGI startup, also has other Darwin-family models placed highly on the GPQA Diamond list.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Stripe to acquire OpenRouter for $7.5 billion

Payments company Stripe has reportedly agreed to acquire OpenRouter, a platform that routes traffic between AI models, in a deal valued at $7.5 billion. The information comes from a Startup Fortune report, which describes OpenRouter as a kind of traffic cop for AI model requests. The source material available here is limited to the headline and link, so details on deal terms, timing, and strategic rationale could not be independently confirmed. Discussion of the report on Hacker News drew modest attention at the time of sourcing.

0 comments Read more at Hacker News

ProgrammingHacker News ·

Qwen3-4B fine-tuned to suppress common AI writing patterns, shared on Hacker News

A developer has released a version of the Qwen3-4B language model that has been trained to reduce the stylistic habits typical of AI-generated text. The project was posted to Hacker News as a Show HN submission, linking to a write-up on the 100toolkit.com blog. The post drew 14 points and 3 comments from the community. The available source does not detail the training method, evaluation results, or the specific writing tics targeted.

0 comments Read more at Hacker News

ProgrammingHacker News ·

Poll Finds Most Americans Oppose Live-Streaming Executions

A Washington Post poll reports that roughly eight in ten Americans oppose broadcasting executions live. The question was raised in connection with the case of the Fort Hood shooter, whose potential execution has drawn public attention. The survey indicates broad public discomfort with turning capital punishment into a live spectacle. The findings were first highlighted on Hacker News, where the article gained modest attention. Details beyond the headline poll figure are limited in the available source material.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Persona Lines in AI Instruction Files Correlate With Lower Quality Scores

A study of the 10,000 most-starred repositories found 285 main AI instruction files, such as CLAUDE.md and AGENTS.md, that open with a 'You are a...' persona line. On the ails check 0 to 10 scale, these files had a median score of 6.6, compared with 7.7 for the other 20,551 main files. The gap persisted across all four file-size groups, ranging from 0.9 to 1.25 points. The analysis attributes the difference to the rest of the file: persona files devote more lines to instructions, and their prohibitions are more often written as strong 'never' statements without matching permissions, hedged, or weakly ordered. The author suggests replacing persona lines with specific, tool-named directives.

0 comments Read more at DEV Community