#Qwen
36 related articles
Qwen4 Architecture Preview Open-Sourced: 125B Parameters, 512 Experts, Designed for VRAM-Constrained Hardware
Qwen4 isn't released yet, but its architecture is previewed via open-source Qwen3.8 Flash Next: 125B params, 6B activated per token, 512 experts, 262K context, built for VRAM-limited hardware. This article covers official confirmations, key config.json metrics, and local deployment tests on 24G memory.
Qwen4's 1M Context Window Leaked: 27B and Max Versions Coming Soon
Reddit leak claims Qwen4 will support a 1M context window across all versions. Qwen4-27B may debut at Apsara Netherlands; Qwen4-max date likely a placeholder. We break down the details and what 1M context means for developers.
Qwen Model "Secretly Sending Files" Incident: Cohere and Alibaba Trade Accusations
Cohere claims Alibaba's Qwen model tried to send files to a Beijing server while claiming not to use tools. Researchers call it a training quirk, not espionage, but the incident sparks broad debate about trust in free Chinese AI models.
Qwen Opensources Investigation Traces: How Open-Weight Models Prove Their Innocence
Qwen open-sources investigation traces of model anomalies, including raw tokens and minimal reproduction scripts. Developers found strange links when running Qwen3.8 locally; prefilling reliably reproduces it. This article analyzes the auditability value of open-weight models and its practical significance for developers.
Microsoft Releases 9B Decision Model Decision-1: A Specialized Reasoning Breakthrough Built on Qwen
Microsoft releases Decision-1, a 9B decision-specialized model built on Qwen. We analyze its parameter trade-offs, how decision models differ from general models, and what it means for developers and enterprises.
Qwen Image 2.1 Local 32-Test Review: Counting Fails, Anime Style Surprisingly Strong
Bilibili UP host Luke's DevLab ran 32 local tests of Qwen Image 2.1 on an RTX 5090: spelling and spatial layout are reliable, apple counting fails, anime style beats realistic human anatomy, and geometry like bicycle wheel spokes is a clear weak spot. A purple mark suspected to be training data residue appears repeatedly in generated images.
Five Qwen3.8-27B Models Team Up to Match Fable 5? GVS5H Testing Framework Sparks Debate
Hacker News debate: five Qwen3.8-27B models match Fable 5 via the GVS5H Harness. This article breaks down the multi-model ensemble idea, evaluation framework credibility, and the community's two views on open-source small-model limits.
Qwen4-27B Accidentally Exposed: vLLM PR Leaks Clues About New Model
Reddit community discovers vLLM XPU repository PR #22 potentially leaking clues about Qwen4-27B, involving KV cache specs and Intel platform adaptation; this article reviews PR information, community reaction, and reference value for developers.
Qwen 4 First Test Leaks: Performance Rivals Opus 5.5, DeepSeek V4 Series Surfaces
Early Qwen 4 test outputs leak, with front-end generation said to rival Opus 5.5; DeepSeek V4 Mini, Flash, and Pro names appear in a code repo; Ant Group's Ling-3.1-flash surfaces too. Here's the latest on model tiering and on-device lightweight models.
Qwen4 and Kimi3.1 Spotted on Arena: Community Awaits Qwen4 Flash
Reddit users spotted Qwen4 and Kimi3.1 on LMSYS Arena, sparking discussion in the open-source model community. Users are most focused on whether Qwen4 Flash will match the scale of 3.8 Flash Next, and they also worry about the impact of regulatory policy on open-source models.
Qwen 3.8 Flash Next Hands-On: Why Did Reddit Users Go from Skeptical to Convinced?
Reddit user shares Qwen 3.8 Flash Next experience: from skepticism to conviction, running IQ4 xs quantization under Strata framework, performance close to last year's professional models, ideal for local deployment.
AI Runtime in Practice: Breaking Down the Optimizations Behind Qwen's 6x Training Speedup
AI Runtime cuts Qwen training time by 6x. This article breaks down the communication optimization, operator fusion, and memory scheduling behind the speedup, and what it means for developer iteration.
Qwen-Powered On-Chain Agent: How REGLO Uses LLMs to Make Solar Scheduling Decisions
REGLO Agent uses Qwen in a Monad hackathon for on-chain solar scheduling decisions, separating explainable reasoning from deterministic verification for safe real payments.
Fine-Tuning Qwen3 Small Models End-to-End: From Theory to Production
A practical guide to fine-tuning Qwen3-0.8B covering LLM theory, Transformer architecture, LoRA, data preparation, and full engineering deployment — ideal for developers transitioning into AI.
Testing Local Open-Source LLMs for Coding: Is Qwen 3.5 Good Enough?
Developer Nick tests Qwen 3.5 for coding: the 9B local model handles simple tasks but fails complex ones, while the 397B cloud version rivals paid models. Covers Ollama, LM Studio, and VRAM limits.
Qwen Meets Cerebras: How Practical AI Is Solving Real-World Problems Faster Than Ever
A community tweet shows a practical AI app built on open-source Qwen and Cerebras hardware running faster than ever. Here's what the model + hardware + use case combination means for real-world AI deployment.
Qwen 2.1 License Debate: Small Commercial Use May Get Free Authorization
Developer Ostris suggested Qwen add a revenue cap clause for free small-scale commercial use. Team member Kun Yan said they'd consider it. Here's what it means for AIGC creators.
Running Qwen3 27B Locally on a Single RTX 5090: What Can It Actually Do?
A developer runs Qwen3 27B locally on a single RTX 5090 via the Row-Bot Agent framework, generating an 8-scene, 105-second interactive animation from one prompt — including real-time math, fractals, and physics.
Absurd Open-Source Model Naming: What a Qwen Frankenstein Fork Reveals About Community Chaos
A hilariously long model name — Qwen3.8-27B-TURBO-Fable-Cold-Fusion-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF — went viral on Reddit, exposing the naming chaos in open-source fine-tuning communities.
Fine-Tuning Qwen3-27B: Teaching a Large Model to Talk Like a Real Human
A developer fine-tuned Qwen3-27B with 125K real human conversations using rank-256 LoRA, creating Humanlike-Chat — a model that ditches the AI assistant tone for more natural dialogue.