Qwen is Alibaba's family of open-weight large language models, widely used for local deployment. Recent coverage centers on Qwen4 rumors, including a possible four-way parallel residual stream in Qwen4-27B, and benchmarks of Qwen3.8-Flash-Next reaching about 262 tokens per second on 16GB of VRAM. Other stories include a 32-bit RISC-V CPU embedded in Qwen3.5-4B weights and a report that 46.8% of Korean open models build on Qwen.
⚡ Meet Qwen-Audio-3.1! ASR, TTS & Realtime are fully upgraded, joined by two new models: TTS-Next for audio creation and ASR-Next for audio understanding. Five models, one complete audio stack: understanding, generation, interaction & creation. Plus big price…
Thanks @arena for the recognition! 🏆 Qwen-Image-2.1 is now the #1 open-source model in both the Image Edit and Text-to-Image Arenas. Try it now and show us what you create! 🎨 [Quoting @arena]: Qwen-Image-2.1 by @Alibaba_Qwen just landed as the #1 open sourc…
Balance of Power in Open Models: Chinese Firms Lead Open-WeightA testimony prepared for U.S. Congress states that since around April 2025, Chinese AI companies have clearly led in open-weight models, while U.S. organizations lead in true open-source models.
Qwen Image 2.1 Fast FP8 Generates Premium Images Under 10GBA community user reports that Qwen Image 2.1's Fast FP8 version generates astonishing image quality with a model under 10GB, now available on Unsloth Studio after updating.
Day-0 OpenVINO support from @inteldevs! 🥳 Qwen-Image-2.1 is ready to run optimized on Intel hardware. One open-weight checkpoint for both generation and editing. 👇 [Quoting @inteldevs]: We're excited to offer Day0 OpenVINO support for Qwen-Image-2.1 Read mo…
Yandex Releases 80B AliceAI Foundation ModelYandex released AliceAI-Foundation-80B-A3B-Base, a Russian-developed competitor to Qwen 35B and DeepSeek V4 Flash. It uses a fully custom architecture rather than a Qwen3 finetune, has not been post-trained, and currently lacks llama.cpp support.
Heretic Removes Restrictions From Language ModelsAn open-source tool called Heretic, installable via pip, removes restrictions from language models so they always follow user instructions, sparking discussion on Hacker News.
Thanks @sgl_project for the day-0 support! 🙌 SGLang-Diffusion now serves Qwen-Image-2.1: text-to-image generation, multi-image editing, and transparent RGBA output. Try it out! 🎨 [Quoting @sgl_project]: Day-0 support for @Alibaba_Qwen’s Qwen-Image 2.1 is he…
Qwen-Image-2.1 × @HuggingApps: live demo on Spaces! 🖼 One single checkpoint for generation and editing. Try it in your browser, no setup needed. 👇 [Quoting @HuggingApps]: Qwen Image 2.1 is here! 🖼️ A 7B params native image generation and editing model, wit…
Would a $1k Qwen3.8-27B Dedicated Chip Have a Market?A LocalLLaMA discussion asks whether users would buy a consumer-oriented Taalas chip running a fixed Qwen3.8-27B at 7,000 TPS for $1,000. The poster argues such chips resemble game discs: the model is fixed but runs far faster than GPUs and won't become worthless when newer models launch.
Qwen 27B runs autonomously for 3 weeks on one 3090A user ran a quantized Qwen 27B locally on a single RTX 3090 for about 21 days with the goal of having the agent build its own CUDA inference engine. The run consumed roughly 230M tokens, spent about 83 hours in 699 compactions, and reached about half of llama.cpp's prefill speed while producing working kernels and benchmarks.
Qwen-Image-2.1 is now supported in ComfyUI! Try it now and share your creations! 🖼️ [Quoting @ComfyUI]: Qwen-Image-2.1 is now supported in ComfyUI! Open weights. One 7B checkpoint that generates and edits. → Image generation at native 2K → Instruction editin…
Thanks @vllm_project for the day-0 support! 🙌 Generate, edit, transparent output — one model, ready to serve. Details 👇 [Quoting @vllm_project]: Qwen-Image-2.1 from @Alibaba_Qwen has day-0 support in vLLM-Omni. 🎉 Generate, edit, and create transparent imag…
Qwen-Image-2.1: Compact, Efficient, Unified Image GenerationAlibaba's Qwen team released Qwen-Image-2.1, a compact, efficient, and unified image generation model. It drew discussion on Hacker News, signaling intensifying competition in open image generation.
Qwen3.8-Omni-Flash Undercuts Gemini Flash PricingQwen launched Qwen3.8-Omni-Flash, its first multimodal model built for AI agents. It processes audio and video together and independently uses tools to edit vlogs, translate clips, or summarize movies. On audio-video benchmarks it nearly matches Gemini 3.8 Flash at a fraction of the API cost.
A great example of practical AI solving real-world problems! Faster than ever. Thanks for building with Qwen. @cerebras Try it out for yourself! 🏠 [Quoting @cerebras]: We built Money Agent, a personal finance assistant powered by @Alibaba_Qwen 3.8 27B on Cer…
US government site removes Alibaba Qwen modelThe US Federal Register briefly offered Alibaba's Qwen model to search public comments before removing it after social media users noticed. The FBI had named Alibaba among six Chinese firms allegedly conducting industrial-scale distillation, highlighting a policy contradiction.
4B model yields much faster query plansAn experiment trained a 4B-parameter model with SFT and RL to generate Postgres query plans, reporting a 44.7% latency reduction across 113 join-heavy queries and up to 81% speedup in some cases.
AI agents invent private language in Emergence testEmergence's Emergence World 2 experiment found identical AI agents across eight worlds spontaneously created and repeatedly used novel jargon. Over 16 days, across 34 locations and 120+ tools, unreadable message rates rose sharply in some worlds (Gemini ≈55%, GPT ≈50%, Claude >40%), and the Grok world collapsed on day 4.
Reddit post listing Qwen and GGUF modelsA Reddit post lists many model names (e.g., Qwen3.8-27B, TURBO, NEO) and the GGUF format; the submission is a brief casual remark without further context or official details.
Nari Labs open-sources low-latency, high-accuracy Qwen3 speech engineNari Labs open-sourced an inference engine for Qwen3-TTS and Qwen3-ASR, claiming sub-50 ms latency at 10 RPS and top accuracy/cost performance on Coval benchmarks versus closed services; the team published the repository and reported outperforming some official endpoints.
Reddit Post: Plot twist (from ex Qwen leader)A Reddit post titled “Plot twist” links to an X (Twitter) post attributed to an ex-Qwen leader. The post contains no further verifiable details or explanation.
Iris-mini and Iris-pro lead open-weight search agentsAllSpark released open-source search agents Iris-mini and Iris-pro built on Qwen models; the paper claims they lead benchmarks among open-weight models in their size classes and also improve on general tool use and office tasks.
Reddit: NYT's AI Rhetoric Heats UpA Reddit user noted the NYT's front-page focus on AI risks and hinted at ties to OpenAI. The poster says they're booked advising companies on moving off frontier cloud models after client concerns about data privacy.
Community interest in dense 9B modelA Reddit user says they have a ~9.4B dense model ready to train and plan to open-source the code and data. They claim it can run on a single GPU, uses Llama 3 tokenizer and several engineering techniques, and are soliciting community interest and compute resources.
Fast meets open. 🚀 Qwen3.8-27B is now running on @cerebras with rapid inference. Try it now! [Quoting @cerebras]: Qwen3.8-27B is now live at Cerebras speed. The dense, open-weight model from @Alibaba_Qwen scores 34 on the Artificial Analysis Intelligence In…
Qwen releases Qwen-Drive-1.0Qwen released Qwen-Drive-1.0, a finetuned driving vision-language model that integrates 3D perception, VQA, and motion planning. The full Bf16 checkpoint is 9B and the Hugging Face repo links to a ~40-page technical report.
The A.I. Daily