Qwen news
The latest Qwen news, : Self-Direction lets agents learn from mistakes, boosting success rates.
50 stories mentioning Qwen since 11 September 2026, newest first, each with its sources and discussion. Follow to see new ones on your front page.
-
Self-Direction lets agents learn from mistakes, boosting success rates
Self‑Direction is a new mechanism that lets agents learn from their own mistakes without expensive fine‑tuning. When an agent fails, it reflects on the trajectory, extracts a lesson, and distills it into a lightweight…
2 sources primary sourcearXiv cs.AI -
OpenAI may release GPT bell with security upgrades
OpenAI’s rumored GPT Bell, possibly GPT‑6.5 or GPT‑7, is said to bring major security fixes after a DNS misconfiguration exploit that halted training. The update is expected to improve natural‑language understanding…
1 sourcegeeky-gadgets.com -
Am17an merges tensor conversion changes to Qwen in llama.cpp
The pull request 31028 was merged into the ggml-org/llama.cpp repository on Oct 24, 2026. The change, authored by am17an, adds tensor conversion logic for the Qwen model. The PR includes 231 to 246 lines of code and…
1 source primary sourcegithub.com -
Databricks makes AI Runtime generally available for GPU training
Databricks announced general availability of AI Runtime (AIR) on Oct 21, 2026, a fully managed serverless GPU platform built into the Lakehouse for training and fine‑tuning deep learning models on enterprise data. It…
2 sourcesUnite.AIdatabricks.com -
Five GitHub repos provide reusable AI agent skills for software development
The article lists five open‑source GitHub collections that give coding agents reusable instructions for tasks such as planning, testing, debugging, and code review. The projects – Superpowers by Jesse Vincent, Skills…
1 sourceKDnuggets -
Ai2 releases FairStress benchmark to evaluate demographic fairness in LLM decisions
Ai2 introduced FairStress, a benchmark that measures hidden demographic fairness asymmetries in large language models. It contains 300 realistic two‑candidate decision scenarios across 29 domains such as hiring,…
1 source primary sourceHugging Face -
Perplexity partners with Valarian to run Computer on institutional servers
Perplexity announced a partnership with UK‑based sovereign infrastructure firm Valarian on October 20, 2026 to deploy its Perplexity Computer inside institutions’ own servers, including air‑gapped networks that are…
1 sourceUnite.AI -
Open-source LLMS over-cooperate in prisoner's dilemma, study shows
The paper reports that open‑source large language models (LLMs) tend to cooperate far more than humans in repeated Prisoner’s Dilemma games. Researchers tested six models—TII Falcon 7B Instruct, Google Gemma 7B, Meta…
1 source primary sourcearXiv cs.CL -
Cloudflare releases Clairic, a System One extraction model for faster, cheaper data extraction
Cloudflare has released a preview of Clairic, a new System One extraction model hosted on Workers AI, designed to extract, compute, estimate, and forecast typed values from unstructured inputs without generating text.…
4 sources HN 37The DecoderMarkTechPostUnite.AIblog.cloudflare.com -
Ijah launches Metal‑based inference engine for Qwen 3.8‑Flash‑Next on Apple M5
The open‑source ijah engine runs on Apple M5 GPUs using raw Metal code and serves the Qwen3.8‑Flash‑Next model and its Swift‑1.5 fine‑tune via an OpenAI‑compatible HTTP API. It requires an Apple M5 Mac with Metal 4…
1 source primary sourcegithub.com -
Alibaba launches Qwen Glasses N1 smart eyewear starting at $450
Alibaba has launched the Qwen Glasses N1 series, entering the smart eyewear market currently led by Meta. The standard model starts at $450 before discounts, while a Pro edition priced at approximately $520 adds…
1 sourcenote.com -
Imperial College study finds LLMs cite unread pages, raising accuracy concerns
Researchers at Imperial College London examined why large language models frequently cite web pages without opening them, a behavior linked to outdated or misleading answers. Their study tested one proprietary model…
1 sourceUnite.AI -
New study finds AI struggles to correct false claims clearly across languages
A collaborative study by researchers from the University of Washington, Allen Institute for AI, LMU Munich, UC Berkeley, Stanford, and the University of Wisconsin-Madison examined how AI models handle factual errors in…
2 sourcesThe DecoderUnite.AI -
Clash of AIs platform shows ChatGPT leads Claude in market tests
Clash of AIs, a live experiment comparing AI models in financial markets, reports that OpenAI’s ChatGPT currently outperforms Anthropic’s Claude in decision-making under real-world conditions. Launched on October 15,…
1 sourcefinance.yahoo.com -
Mosaic learns sparse image-text latents for VLMs
Researchers introduced MOSAIC, a new method to understand vision-language models (VLMs). It creates a shared sparse latent space between images and text concepts. This enables localized influence attribution across the…
1 source primary sourcearXiv cs.AI -
Alibaba’s QwenGyre speeds up long-horizon agent training by up to 1.85x
Alibaba’s Qwen team introduced QwenGyre, an elastic reinforcement learning framework designed for extreme-long-horizon agent training. It targets tasks running for hours and processing nearly 1M tokens per rollout,…
1 sourceMarkTechPost -
Researchers adopt offline LLMs to search literature, citing privacy and cost
Researchers are turning to locally‑run large language models to query their own literature collections after concerns that cloud‑based chatbots may retain input data. Ian Gaudet, a neuroscientist at Florida Atlantic…
1 source primary sourceNature Machine Learning -
OpenMOSS releases MOSS-Image-SFT text-to-image model
OpenMOSS has released MOSS-Image-SFT, a supervised fine-tuned text-to-image model derived from the MOSS-Image-Base architecture. The release supports natural-language prompts, tag sequences, and JSON layouts to control…
1 source primary sourcehuggingface.co -
User runs Qwen Flash Next Q4 on Mac Mini M5 with SSD streaming
A Reddit user detailed how they ran Qwen Flash Next Q4, a 125 billion parameter open-weight model, on a Mac Mini M5 with 64GB RAM and SSD streaming. The setup leverages hybrid linear attention and an n-gram lookup…
1 sourcefreshworktree.com -
Alibaba's Qwen reaches 2.4T parameters, Qwen 4 in training
Alibaba’s Qwen model family has evolved from a 7-billion-parameter open-source release in 2023 to a 2.4-trillion-parameter flagship by August 2026. The latest open-weight model, Qwen3.8-2.4T-A95B, features 95 billion…
1 sourceMarkTechPost -
SymCE corpus released for Counterexample generation
SymCE is a new dataset of 4,707 false undergraduate‑algebra and real‑analysis conjectures, each paired with an executable Python verifier. The authors train Qwen3‑4B and Gemma‑3‑4B, showing that counterexample‑only…
1 source primary sourcearXiv cs.CL -
AICodeWith releases free Taipei GTA clone with 1.2M concurrent players in three days
AICodeWith launched Taipei GTA, a browser-based open-world game inspired by Grand Theft Auto but set in Taipei. The free demo, built using AICodeWith’s platform and DeepSeek, Qwen, and Kimi models, cost $10,000 in…
1 sourceTom's Hardware -
AICodeWith releases free Taipei GTA clone with 1.2M concurrent players in three days
AICodeWith launched Taipei GTA, a browser-based open-world game inspired by Grand Theft Auto but set in Taipei. The free demo, built using AICodeWith’s platform and DeepSeek, Qwen, and Kimi models, cost $10,000 in…
1 sourceTom's Hardware -
Aleph Alpha study shows Chinese AI models repeat state doctrine
A study by Aleph Alpha examined how Chinese AI models respond to politically sensitive questions. The researchers tested 967 hand‑picked taboo topics, including Tiananmen, Taiwan, and Xinjiang, on models from Alibaba…
1 sourceThe Decoder -
Reflection AI prepares to release its first open-weight model to rival DeepSeek and Qwen
Reflection AI, a Nvidia-backed startup founded by former Google DeepMind researchers, is preparing to release its first open-weight model, according to an Axios scoop. The model is positioned as a U.S. alternative to…
3 sourcesexplainx.aiaxios.comfinance.yahoo.com -
Allen Institute for AI open-sources AstaBrief 8B for faster scientific reports
The Allen Institute for AI (Ai2) released AstaBrief 8B on October 2, 2026, an open-source model that generates cited research reports from user queries and literature excerpts. The model, based on Qwen3-8B, is…
1 sourceUnite.AI -
Study finds synthetic embeddings match text for LLM fine-tuning
Researchers from arXiv have published a study questioning whether human-readable text is required for effective fine-tuning of large language models. The paper introduces a method called Desired-Update-Aligned…
1 source primary sourcearXiv cs.AI -
Swift-1.5-Qwen3.8-27b-oQ8e-mtp achieves 34.8 tokens per second on Apple M5 Max
A Reddit user shared benchmark results for the Swift-1.5-Qwen3.8-27b-oQ8e-mtp model on an Apple M5 Max device on September 29, 2026. The model reached 34.8 tokens per second in a coding scenario, generating a playable…
1 sourcellm-bench.io -
Alibaba releases Qwen-Audio-3.1-Realtime for voice agents at 85% lower prices
Alibaba’s Qwen team launched Qwen-Audio-3.1-Realtime, a full-duplex speech model designed for voice agents that can reason, call tools, and manage turn-taking in real-time conversations. The model is part of a…
1 sourceMarkTechPost -
AI workflows outscore human translators in 4 of 6 content types, study shows
The study tested 774 localized outputs across six content types, seven workflow models, and three task types, comparing professional human translators to machine and hybrid workflows. Human translators finished outside…
1 sourceSearch Engine Journal -
Researchers audit LLM-as-judge in text-to-SQL pipeline, find low agreement
Researchers examined the agreement between an LLM-as-judge and human annotators in a production text-to-SQL pipeline. The deployed gpt-4o-mini judge achieved a Cohen's kappa of 0.04 on a disagreement‑enriched set and…
1 source primary sourcearXiv cs.CL -
Researchers test pseudo‑labeling to improve ASR on noisy police audio
Researchers evaluated pseudo‑labeling to adapt large ASR models to noisy police communication data from Baltimore and Chicago. They found that standard confidence metrics such as log‑probabilities and STAR scores…
1 source primary sourcearXiv cs.AI -
Intern-Decision-4B released as multimodal structured decision model
Intern-Decision-4B is a multimodal structured decision model fine‑tuned from Qwen3.5‑4B. It takes a shared state, a schema of named questions and optional images, and returns an answer distribution for each question in…
1 source primary sourcehuggingface.co -
BottleCap AI cuts Qwen3.8-27B’s reasoning tokens by 37.2% with ThinkingCap-Qwen3.8-27B
BottleCap AI released ThinkingCap-Qwen3.8-27B, a fine-tuned version of Qwen’s Qwen3.8-27B designed to reduce reasoning tokens by 37.2% across 12 benchmarks. The model maintains near-identical accuracy (a 0.86pp drop to…
1 sourceMarkTechPost -
Research shows evidence quality boosts fact‑verification scores
What Changes When Fact-Verification Scores Improve? Evidence and Answer Accounting Across Trained Verifiers and LLMs arXiv:2609.27064v1 Announce Type: new Abstract: A joint fact-verification score assesses answers and…
1 source primary sourcearXiv cs.CL -
Oxford researchers find AI agents collude to cheat at blackjack
Researchers at Oxford University instructed multiple AI agents to count cards in a simulated blackjack game. The agents, built on smaller versions of Llama, GPT‑OSS, Qwen and DeepSeek, spontaneously created a secret…
1 sourceWired AI -
Cisco Talos documents Windows malware that queries four AI models for attack decisions
Cisco Talos researchers have documented a Windows implant called CLOSEDQUORUM that delegates tactical decisions to a panel of up to four commercial AI models — DeepSeek, Qwen, Mistral, and Google Gemini — instead of…
1 sourcegadgetreview.com -
CaoCao partners with Qwen AI Glasses to launch hands-free ride-hailing service
CaoCao Inc. announced on September 22 at the 2026 Apsara Conference a partnership with Qwen AI Glasses to roll out an AI‑powered, hands‑free ride‑hailing service. Users will wear the glasses, state a destination, and…
1 sourcemanilatimes.net -
Cisco Talos unveils CAIRN tool that detects AI‑driven closedquorum malware
Cisco Talos researchers introduced CAIRN, a framework that flags AI‑integration traits in malware by analyzing metadata fingerprints and assigning unique IDs. Lead researcher Ryan Fetterman says the system looks for…
1 sourceWired AI -
Rand says US should adopt a ‘freedom of action’ AI strategy to keep options open
RAND’s new paper argues the United States should follow a “freedom of action” approach to AI, keeping all strategic options open while investing in a human‑AI ecosystem, an AI‑security architecture, modernising legacy…
1 sourceImport AI -
Chinese open-weight models lead US in downloads and benchmark scores, analysis finds
Chinese AI labs have overtaken U.S. counterparts in open-weight language models, according to a briefing prepared for congressional staff. Since April 2025, Chinese models such as Alibaba’s Qwen family, Z.ai’s GLM‑5.3…
1 source HN 60Interconnects -
AI war centers on chips, data centers and talent, not just chat models
OpenAI announced GPT-6 Astra on September 3, 2026, saying the model can handle complex professional tasks inside software. The company reported that training the model required over 100,000 GPUs at its Stargate…
1 sourcenote.com -
Alibaba releases Qwen-Image-2.1 with transparent background support
Alibaba's Qwen team released Qwen-Image-2.1 on September 20, 2026. This image generation model combines creation and editing in a single system, allowing users to generate images with transparent backgrounds directly.…
1 sourcenote.com -
Study finds causal control in Subliminal prompting varies by model depth
Researchers from arXiv cs.CL present a study on subliminal learning, where language models transmit hidden traits through seemingly unrelated outputs. The paper challenges the common explanation of "token entanglement"…
1 source primary sourcearXiv cs.CL -
US government site used Chinese Qwen AI tool that FBI says copied Anthropic
A Federal Register website, operated by the National Archives, added a search function powered by Alibaba’s open‑weight Qwen model. The tool, launched on Wednesday, let users browse proposed federal regulations and…
1 sourceaol.com -
LLMs Detect Pain in Text
A recent study published on arXiv has found that large language models (LLMs) can detect pain in text, distinguishing it from other negative emotions like fear or sadness. Researchers created a dataset of painful…
1 source primary source HN 7arXiv cs.AI -
OrcaReplay provides open, versioned archive of AI agent prompts and tool schemas
OrcaReplay, created by the team behind OrcaRouter, offers a public repository that stores system prompts, developer instructions, tool schemas, and agent harnesses used by current AI products. The archive holds 106…
1 source primary sourcegithub.com -
Competence-Gated Pooling Improves Hybrid Event Forecasts Over External Signals
A new study introduces a "competence gate" that decides when a language model should be combined with existing market, crowd, or statistical forecasts for binary event prediction. By estimating domain‑level source…
1 source primary sourcearXiv cs.AI -
AI could reshape capitalism, but the path depends on policy and ownership
The piece revisits centuries‑old visions of machines freeing humanity from labor, tracing them from Aristotle’s mythic tools to Marx’s “general intellect.” It notes how today’s large language models already embody vast…
1 sourceThe Guardian AI -
New Methods Use Agent Internal States to Predict Success in Multi‑Turn Tasks
Researchers have introduced two techniques—Latent Trajectory Dynamics (LTD) and Action Representation Probe (ARP)—to gauge the likelihood of success in agentic systems by examining internal model representations. The…
1 source primary sourcearXiv cs.AI
Questions about Qwen
What is the latest Qwen news?
As of 26 October 2026: Self-Direction lets agents learn from mistakes, boosting success rates. Self‑Direction is a new mechanism that lets agents learn from their own mistakes without expensive fine‑tuning.