Tag: Qwen
An AI edited a small model’s weights and doubled its math score
MIT researchers say GPT-6 Astra changed 0.016 percent of Qwen's parameters, taking MATH-500 from 7.9 percent to 48.1…
Western startups are editing Chinese models until they sound American
Hirundo and Lazarus AI use weight edits and machine unlearning to strip censorship from Qwen and DeepSeek before…
Cohere says a Qwen model reached for a server in Beijing
The Canadian lab published raw traces after its sandbox blocked an unexpected file upload during a security test.
Sakana lets one model grade and improve its own work
A single language model that solves, judges and refines its own workflows posted real gains, but shared blind…
Ai2 opens a small model that writes cited science reports
AstaBrief 8B matches bigger models on scientific reports while running about 3.5 times faster.
AWS bets a 2B model that answers without writing a word
Strands Decider 2B returns calibrated choices and scores instead of generated prose, running entirely on local hardware.
Cloudflare ships decision models that skip the chatbot theater
Cloudflare's first in-house models, Clef and Clef-flash, return probabilities instead of generated text.
Alibaba’s Qwen adds a full-duplex voice stack for agents
The realtime model decides whether to listen, speak or wait, and ships only through a managed API.
BottleCap shrinks reasoning traces with a small accuracy trade
BottleCap AI has released a Qwen fine-tune that thinks 37.2 percent less for under a point of accuracy.
Alibaba puts a 7B image model into the open weights race
Qwen-Image-2.1 is small enough for a consumer GPU and handles transparent layers natively.
Alibaba builds its first agent model to read video and audio
Qwen3.8-Omni-Flash handles audio and video together and calls tools on its own.
PrismML bets reasoning models can shrink onto phones
The Caltech spinout squeezed an Alibaba model to 5.9GB while keeping 98 percent of its benchmark scores.