Darwin-180B-RSI-R3 Ranks Second on SuperCLUE Math and Science Reasoning Test

·2 views

VIDRAFT, a South Korean AI startup, released Darwin-180B-RSI-R3, a 180-billion-parameter Mixture-of-Experts model built on Alibaba's Qwen3.8-Flash-Next. Independent evaluation by SuperCLUE on November 5, 2026 placed it second overall and first among open-weight models in math and science reasoning, with a score of 73.68. The model was produced by retraining only the attention, linear-attention, and shared expert layers while keeping the feed-forward expert blocks that store factual knowledge intact. Its weights are available under a Qwen community license permitting commercial use, and a 4-bit quantized edge variant can run on a 32 GB RAM laptop without a GPU. The reported gain over its base model in math and science reasoning was 15.8 points.

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)

Python Tool Exports Agoda Hotel Reviews and Sub-Scores to a Spreadsheet

A developer has published a tutorial on pulling Agoda hotel reviews, scores and sub-scores into a CSV file using Python. The guide notes that each Agoda hotel page combines Agoda's own guest reviews with Booking.com reviews republished on the site, each with separate scores and six sub-ratings. The approach uses a paid scraping tool, built by the author, that accepts hotel links, IDs or names and returns one row per review along with hotel-level data. The sample script also summarises average ratings by traveller type and flags hotels whose details may need checking. The post additionally suggests uses such as mining complaints from low-rated reviews.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

On-Device AI Hardware Faces Physical Limits: Heat, Power Droop, and Bandwidth

A PCB manufacturer warns that moving large language and vision models onto portable devices exposes physical constraints that software teams often overlook. Chips sustaining high compute loads can produce intense heat, and without adequate thermal copper, vias, or heatsinks, they throttle performance or shut down. Sudden surges in current demand can also cause voltage droop that crashes systems, a failure that code changes cannot fix. The article adds that memory bandwidth becomes a critical bottleneck on edge hardware, which is more constrained than in PCs or servers. The author argues that hardware design must be planned alongside software to make on-device AI reliable.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Byte Pair Encoding Builds the Token Vocabulary Behind Language Models

Before a language model processes text, a tokenizer splits it into integer IDs drawn from a fixed vocabulary, and the model never directly sees letters or whole words. Byte pair encoding builds that vocabulary by starting from single characters and repeatedly merging the most frequent adjacent pair until a target size is reached. The article notes that vocabulary size is a tradeoff: larger vocabularies pack text into fewer tokens but add parameters, and because the vocabulary is learned from data, languages like Urdu end up costing far more tokens than English. The method originated as a 1994 data compression technique and was adapted for neural machine translation in 2016 by Sennrich, Haddow, and Birch to handle rare words. The authors implemented the algorithm from scratch on the paper's example dictionary to demonstrate how frequent words become single tokens while rarer ones split into pieces.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineer outlines a step-by-step method for planning AWS infrastructure tasks

A DevOps lead with about 15 years of infrastructure experience describes how a vague ticket to deploy a web service on AWS can hide many small steps. He explains that engineers first gather project context, current infrastructure details, and product knowledge before planning. The work is then split into stages such as writing infrastructure code, building workflows, creating AWS resources, deploying the app, and verifying it, with attention to their dependencies. Each stage breaks down into simple actions, guided by team conventions and past experience. He presents this as a way to turn a one-line ticket into a clear, checkable plan with actions, verification steps, and rollback.

0 comments Read more at DEV Community