Your central hub for AI news and updates on Qwen. We're tracking the latest articles, discussions, tools, and videos from the last 7 days.
Fyra's Brief
NVIDIA has released Qwen3.5, an open-source vision-language model with 397B parameters for native multimodal agents, ideal for tasks like coding, visual reasoning, and complex search.
Qwen3.5 is a significant milestone in multimodal AI, offering improved capabilities for native multimodal agents with wide-ranging applications.
Fyra's Brief
Alibaba's Qwen AI project has lost a key technical leader just a day after unveiling its new Qwen 3.5 open-weight small models. Junyang Lin's departure has left the community shocked and expressing admiration for his contributions.
Junyang Lin's departure highlights the intense competition in the AI landscape, where companies must adapt and innovate to stay ahead.
0.8B-9B native multimodal w/ more intelligence, less compute Discussion | Link...
Check back soon for video content
Community talk
Qwen 3.5 35B A3B is better than free-tier Chatgpt and Gemini
Qwen3.5 27B scores 42 on Intelligence Index and is the most intelligent model under 230B. Nearest model GLM-4.7-Flash 31B-A3B, Scores 30
Qwen3.5 is dominating the charts on HF
Alibaba CEO: Qwen will remain open-source
Qwen3.5 9B is the first local model that I tried, that can make adequate flappy bird version
Qwen3.5-9B abliterated — 0% refusals + vision
Qwen3.5-4B Uncensored Aggressive Release (GGUF)
Finished a Qwen 3.5 9B Opus 4.5 Distill!
Unsloth fixed version of Qwen3.5-35B-A3B is incredible at research tasks.
Qwen 3.5 27b: a testament to the transformer architecture
Breaking : Today Qwen 3.5 small
Qwen3.5 Small Dense model release seems imminent.
What if LLM agents passed KV-cache to each other instead of text? I tried it -- 73-78% token savings across Qwen, Llama, and DeepSeek
Qwen 3.5-35B-A3B is beyond expectations. It's replaced GPT-OSS-120B as my daily driver and it's 1/3 the size.
Qwen 3.5-27B punches waaaaay above its weight (with a slightly different prompt) -- very impressed
New Qwen3.5-35B-A3B Unsloth Dynamic GGUFs + Benchmarks
PewDiePie fine-tuned Qwen2.5-Coder-32B to beat ChatGPT 4o on coding benchmarks.
Qwen3.5 feels ready for production use - Never been this excited
Qwen3 9B can run fine on android phones at q4_0
Qwen3.5-0.8B - Who needs GPUs?
Qwen3.5-35B-A3B hits 37.8% on SWE-bench Verified Hard — nearly matching Claude Opus 4.6 (40%) with the right verification strategy
Qwen3.5-27B Q4 Quantization Comparison
Running Qwen 3.5 0.8B locally in the browser on WebGPU w/ Transformers.js
Visualizing All Qwen 3.5 vs Qwen 3 Benchmarks
Qwen 3.5 2B on Android
PSA: Qwen 3.5 requires bf16 KV cache, NOT f16!!
Running Qwen3.5 27b dense with 170k context at 100+t/s decode and ~1500t/s prefill on 2x3090 (with 585t/s throughput for 8 simultaneous requests)
The last AMD GPU firmware update, together with the latest Llama build, significantly accelerated Vulkan! Strix Halo, GNU/Linux Debian, Qwen3.5-35-A3B CTX<=131k, llama.cpp@Vulkan&ROCm, Power & Efficiency
Qwen3.5 35b a3b first small model to not hallucinate summarising 50k token text
Dense (non-thinking) > MoE? Qwen-3.5-27B is blowing me away in coding
Qwen 3.5 27B is the best Chinese translation model under 70B
How to switch Qwen 3.5 thinking on/off without reloading the model
Qwen3 Coder Next | Qwen3.5 27B | Devstral Small 2 | Rust & Next.js Benchmark
Qwen3.5 35B-A3B replaced my 2-model agentic setup on M1 64GB
Little Qwen 3.5 27B and Qwen 35B-A3B models did very well in my logical reasoning benchmark
Qwen3.5-35B-A3B running on a Raspberry Pi 5 (16GB and 8GB variants)
Follow-up: Qwen3.5-35B-A3B — 7 community-requested experiments on RTX 5080 16GB
Qwen 3.5 Architecture Analysis: Parameter Distribution in the Dense 27B vs. 122B/35B MoE Models
Qwen3.5 27B vs Devstral Small 2 - Next.js & Solidity (Hardhat)
Qwen 3.5 Family Comparison by ArtificialAnalysis.ai
Qwen3.5-35B-A3B Q4 Quantization Comparison
Introducing FasterQwenTTS
speed of GLM-4.7-Flash vs Qwen3.5-35B-A3B
Strix Halo, GNU/Linux Debian, Qwen3.5-(27,35,122B) CTX<=131k, llama.cpp@ROCm, Power & Efficiency
Qwen3.5-27B-heretic-gguf
Qwen3.5-35B-A3B is awesome
Qwen3.5 122B in 72GB VRAM (3x3090) is the best model available at this time — also it nails the “car wash test”
Qwen/Qwen3.5-35B-A3B creates FlappyBird
[P] I trained Qwen2.5-1.5b with RLVR (GRPO) vs SFT and compared benchmark performance
Running Qwen3.5-0.8B on my 7-year-old Samsung S10E
Qwen 3.5 small , soon
qwen3.5 35b-a3b evaded the zero-reasoning budget by doing its thinking in the comments
Is Qwen3.5 a coding game changer for anyone else?
System prompt for Qwen3.5 (27B/35BA3B) to reduce overthinking?
What's Next for Qwen After Junyang Lin's Departure?