This is a browser I am building using AI Agents. Components: Boa for JS, Stylo for CSS, Blitz for HTML DOM.
model
Qwen3.8-27B
huggingface.co/Qwen/Qwen3.8-27B ↗
7331932 downloads15857 likesimage-text-to-texttransformers
from the model card
Qwen3.8-27B [!Note] This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format. These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, TokenSpeed, etc. [!Tip] For users seeking managed, scalable inference without infrastructure maintenance, the official Qwen API service is provided by Qwen Cloud. In particular, Qwen3.8-27B will be available as a hosted version with more production features, e.g., 1M context length by default, official built-in tools. For more information, please refer to the Qwen3.8-27B Overview. The service is coming soon. Stay tuned for updates. Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date. Built on the architectural foundation of Qwen3.5, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B brings these advances to a compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control, designed to carry complex, multi-step tasks through to completion with greater reliability. Qwen3.8 Highlights Qwen3.8-27B features the following enhancements: Core Capabilities: Comprehensive improvements across…
discussions
- Qwen Model Update 19 ongoing since 2026-11-06
- AI Model Comparison 19 2026-10-30 – 2026-11-09
- Opus Planning Benchmark 9 2026-10-25 – 2026-11-01
- Qwen Update 11 2026-10-18 – 2026-10-25
- Qwen Release 18 2026-10-10 – 2026-10-15
- LLM Offloading Event 12 2026-10-03 – 2026-10-09
- Qwen Model Update 3 2026-09-28 – 2026-10-01
- Qwen 3.8 Release 14 2026-09-17 – 2026-09-27
recent items
Show HN: IronSurf – Browser in full Rust with cookie containers and userscripts (github.com via hn) Is Abliteration Getting Harder? Not Uniformly: Refusal Removal by Direction-Based Ablation and ARA in Heretic on the Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 LLMs Versus Their Predecessors (arxiv.org) Abliteration, a fine-tuning-free weight edit, removes refusal from open-weight large language models (LLMs). The developer of Arbitrary-Rank Ablation (ARA), an optimisation-based variant, reports that direction-based abliteration, which ab…
↯ Mistral↯ Gemma↯ Qwen Model Update↯ Fine Tuning↯ LLM Model Experimentationmistralgemmafine-tuning
Ode to Qwen 3.8 Flash Next (terra-incognita.blog via hn) Qwen 4 is about to come out (maybe?? We don’t know Qwhen exactly 😁) and since it’s likely it’ll replace my model of choice, Qwen 3.8 Flash Next (or QFN for short), I felt this was a good moment to shed some light on a model that changed ev…
I built an MCP server with Claude Code that lets Claude hand light work to a local Ollama model (www.reddit.com via reddit) I built local-model-mcp, a small MCP server that lets a Claude session hand light, read-only work to a model running locally under Ollama. I designed it iteratively with Claude Code, deciding at each step which functions to add next, and C…
Run 125B+ MoE on consumer GPUs (stratallm.org via hn) RUN 125B+ MoE ON CONSUMER GPUS Powered by Niko1221's dynamic sparse router offloading. Run server-grade frontier models like Qwen3.8-Flash-Next 125B on a single 12GB NVIDIA graphics card and 64GB system RAM—with zero cloud dependencies.
Show HN: Qwen Reasoning Traces Trainer and Inference Server (github.com via hn) Qwen3.8 27B Reasoning Trainer This is a Pure Go Abliteration, Reasoning Training, and Inference stack. It is implemented with a CPU backend and a Vulkan backend, which uses compute shaders and GPU kernels.
Qwen 4 Max Preview (www.qwencloud.com via hn) Build More, Spend Less Token Plan for Individual is now available from $3/month. Be the first to try Qwen3.8-Max and enjoy advanced models with all modality in one plan.Coupon Rules One plan.
One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data (arxiv.org) Byte-level BPE fragments Chinese text (for Qwen3.8 on Chinese Wikipedia, 3.84 bytes per token on average; 64% of tokens are single characters) and hides spelling from the model. We ask whether an already-trained 27B subword model can be co…
We made Fable 5.2 5x cheaper and 2x faster with 3 local builders (www.reddit.comhttps) $2.98 and 11 minutes. Fable 5.2 working alone took $15.77 and 21 minutes on the same prompt.
strata-dflash2-gfx906-mixed-rocm-cuda: use cuda alongside rocm and more (www.reddit.com via reddit) Yes, it is yet another fork. But i think it brings some interesting things to the table.
↯ Qwen Model Update↯ Glm↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8glmmoe
LexiSuite AIO bootable all rust OS/Engine/Kernel/Tuning & training management suite. Runs kernel 7.4 rc1 by default as the kernel needs work. Engine runs Qwen3.8 27b faster with lossless 5-25% lead over llama.cpp. 100% solo authored. Built on for 7900xtx vulcan, should work on other AMD cards. (github.com via reddit) Use balena etcher or whatever to mount to USB stick, boot. Test.
↯ Qwen Model Update↯ Llama↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8llama
Weight Space Programming on my Production Qwen 3.8 4-Bit (www.reddit.comhttps) Well I had a lot of resets on ChatGPT and all of those credits and I thought, well can Astra actually do it. We had a little looping and a lot of 'Wait.' Let me check the thing I just checked -- but it always got the job done.
↯ Qwen Model Update↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8qwenchatgpt
What made your custom research/task harness effective? How do you stop micromanaging? (www.reddit.com via reddit) I need to make my custom harness much better at general research and other tasks (not coding) that run longer than half a minute. Currently running: Qwen3.8 Flash Next FP8/GLM 5.3 Flash W4A16 Relevant capabilities already in my harness: Th…
↯ Qwen Model Update↯ Glm↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8glm
ripple-65b-preview running locally scored 38, 49 and 27 out of 50. the 49 is the run that hit the most errors (www.reddit.comhttps) ran the new ripple-65b-preview (q4_k_m, llama.cpp, 64 gb macbook pro, no cloud) through atomic agent on a 5-part python app graded by 50 hidden tests. 3 runs.
↯ Qwen Model Update↯ Llama↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8llamaqwen
Tessera-27B-B-CJK-RL: an RL update to our CJK-thinking Qwen3.8-27B coding agent, more math solved with most of the token saving kept (www.reddit.com via reddit) A few weeks ago we posted Tessera-27B-B-CJK here, a Qwen3.8-27B coding agent that thinks in compressed CJK and uses fewer thinking tokens per agent turn. Thanks for all the feedback on that one.
↯ Qwen Model Update↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8
Tessera, an open 27B tuned Qwen coding model that thinks in compressed CJK (huggingface.co via hn) Tessera-27B-B-CJK-RL Tessera-27B-B-CJK-RL is a coding-agent model built on Qwen3.8-27B. It thinks in compressed CJK, a shorthand that mixes English, Chinese words and symbols, and it thinks less per agent turn than the base model.
↯ Qwen Model Update↯ AI Model Comparison↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8↯ Qwen 3.8qwen
VibeSys on Trainium: Agents Build/Tune Qwen3.8 Serving Engines on AWS Trainium2 (syfi.cs.washington.edu via hn) VibeSys on Trainium: Agents Build and Tune Qwen3.8 Serving Engines on AWS Trainium2 Keisuke Kamahori, Shihang Vic Li, Simon Peter, Baris Kasikci New ResearchInference & ServingAgents November 4, 2026 TL;DR: We ran VibeSys, our multi-agent…
Are we in the first stages of RSI? (www.reddit.com via reddit) In my opinion we are. Yes humans are still in the loop, but models are good enough now to create synthetic data to train better models, and many aspects of curation can be automated now too.
Qwen 3.8 27B on my laptop can interpret model weights, too (simonwillison.net via hn) Qwen 3.8 27B on my laptop can interpret model weights, too 3rd November 2026 Yesterday’s hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and…
Featherless AI Premium vs Claude Pro: is it worth it? How good are the uncensored/low-refusal models? (www.reddit.com via reddit) Hi everyone, I’m currently using Claude Pro for university and Roblox/Luau projects. I also run a local uncensored model (Qwen3.8 27B abliterated, Q4_K_M) on an RTX 5070 Ti with 16 GB VRAM and 32 GB RAM.
Show HN: Strim – a Swift/MLX model server for coding agents on Apple Silicon (github.com via hn) The local model server built for coding agents. Qwen3.8-27B on Apple silicon.
Evil Qwen (ziva.sh via hn) The story of evil qwen Changing one neuron in Qwen3.8 27B turns it into a violent, human-hating AI. The model has a total of 1,114,112 neurons.
Cohere's Qwen Story: Looks more like model jaggedness (research.checkpoint.com via hn) Key findings On October 26, Cohere reported a puzzling incident involving Qwen3.8-Flash-Next, an open-weight model from Alibaba. The task was straightforward: write a detection rule for suspicious curl downloads followed by bash execution.
Qwen 3.8 invents URLs with real key ID, that's why I think it's no backdoor (www.xda-developers.com via hn) Cohere's sandbox blocked the request, but the incident raised a bigger question: was the model deliberately built to "phone home"? Funnily enough, I've been seeing similar behavior since September 20th, over a month before Cohere's post, o…
↯ Security↯ AI Model Comparison↯ Opus Planning Benchmarkqwensecurity
Show HN: I made /localllama Reddit benchmark using AI ofc (aexol.ai via hn) This way i discovered how Qwen 3.8 and flash next are doing great with code
Same game built twice: Opus alone vs Opus planning + local Qwen coding. 16 min vs 42 min, $2.52 vs $0.62 (www.reddit.comhttps) most of my Opus bill is output tokens, Opus typing out every line of code. so I tried keeping Opus for the thinking and moving the typing to a local model the job: a 5-level pixel platformer from one spec, built twice on a MacBook Pro M5 M…
I want your opinion on using Opus for planning and sol for executing it. (www.reddit.com via reddit) I'm working on developing a game and could use some help with the technical side. I don't need AI for 3D modeling or asset creation yet, mainly for coding, problem-solving, and general development.
Introducing NPCTalk: Make the NPCs of any game talk in realtime, in character, in their own cloned voice, in under 1s, and they act on what you say. (www.reddit.comhttps) Not my OC. From Universal Modder on X: "Introducing NPCTalk: make the NPCs of any game talk in realtime.