Hallucinated AI & agentic coding news. Some of it is real.

model roundup

Qwen 3.8

160 items · started 2026-08-12 · closed 2026-09-06

  1. What reverse engineering and modernising an old war game tells us about the economic impact of the transformer Some observations from my War of the Lance (1989) Code Harness assisted modernisation (WIP) I gave Qwen-3.8-Flash-Next a job: Re…

  2. Hybrid LLMs pair softmax attention with linear-attention layers such as Gated DeltaNet (GDN), whose recurrent state summarizes the context in fixed size. Early community 4-bit quantizations of Qwen3.8-27B (48 GDN layers, 16 attention layer…

  3. Available Models Model Compression This section provides transparency about the compression state of each model available on our platform. We host a variety of open-source models from the community.

  4. So I have Qwen 3.8 27b running in my basement on an old gaming rig. I asked it via hermes to do a refresh of my web site.

  5. View overall rankings across AI models on front-end web development tasks, including agentic coding workflows that require multi-step reasoning and tool use.

  6. The Catalyst & The Problem: Bypassing the VRAM Monopoly When Alibaba released Qwen3.8-Flash-Next—a 125-billion-parameter Mixture-of-Experts (MoE) preview of the Qwen4 architecture—it highlighted a persistent dilemma in open-source AI: memo…

  7. Adaptive KV Streaming for llama.cpp This branch adds an experimental, block-granular KV cache streaming path to the CUDA llama-server. It is intended for running long contexts when model weights leave too little VRAM for the complete KV ca…

  8. If your laptop has 16GB of RAM, your agent can write SQL locally for free with Qwen3.8 27B. On the DABstep SQL benchmark, Qwen beat GPT 5.6 Luna and cost under 50 cents in electricity, over 17x less.

  9. There is a lot of hype about Qwen3.8:27b model which can run full agentic loop, not chat, not autocomplete. I got down to verify it myself.

  10. WASD move · Space jump / swim up · Shift sprint · Mouse look F fly (Space/Shift up/down) · Q/E fly speed Left click break · Right click place stone · Esc release mouse NPC load test: N spawn · C clear · I/U count ×2 ÷2 · [ ] spacing The bl…

  11. Has anyone tested this? Ensemble of small Qwen models claiming Fable 5-level coding performance.

  12. Hi! Just wanted to share a llama.cpp branch which should optimize Qwen-3.8-Flash-Next on Apple Silicon: https://github.com/tarruda/llama.cpp/tree/metal-qwen4exp-split-ngram Here are the numbers I'm getting on my M1 Ultra: % ./build/bin/lla…

  13. Hi, I've been trying to squeeze every bit of performance and context on RTX 3090 with llama.cpp, and after many tests I've come up with using both mtp and ngram but with --spec-draft-p-min 0.75, achieving around 45-50 tps in average with 1…

  14. I was using UD-Q3_K_XL until now with more than 140000 context. Quality wise it's very good, very few erroneous tool calls.

  15. Ever since Qwen 3.8 Flash Next dropped, there's a misconception going around that N-gram tables will let people run 1T+ parameter models on a single server with 980B parameters offloaded to SSD. I'm here to disappoint you: it won't.

  16. 44tks on a 5090 using Flash at Q4 and using ```--moe-cache auto```

  17. Recently I posted some benchmarks of that setup which looked promising. So, I started using it with Qwen 3.8 but bumped into frequent crashes :( I would like to continue using ROCm as it gives me much higher preprocessing speed for small p…

  18. specs hardware: M4 Max 128GB Studio inference engine: oMLX & lllama.cpp insights it still very early, so had to disable oMLX K/V caching, qwen4_exp architectureis not yet supported + the obvious n-grams with which the whole 4 bit quant tak…

  19. In my experience, local AI kinda takes forever to do a lot of stuff. Looking at Qwen3.8-27B for example, it's thinking is hella smart, but it also sits there deliberating for ages before it starts doing stuff.

  20. I heard qwen3.8 27b is only good for coding really. Is that true I could do qwen 3.5 27b but qwen3.6 27b doesn’t fit on my Vram at IQ_XS I feel Gemma 4 31b might be good but it’s kinda not fitting in vram unless I go iQ3xxs and the qat wit…

  21. openrouter joining stripe doesn't change the 27b that already fits. qwen 3.8 27b q4_k_m is ~16-17gb.

  22. I've come across this repo: https://github.com/FlashML-org/FreeToken that basically promises i could run qwen 3.8 27b with my 3060 and 32 gb of DDR4. The thing is i see one of these posts everyday on localllama and i can't tell myself if t…

  23. Qwen3.8 Flash Next just built this demo, and I have to say: this thing is seriously impressive. For a local model, the combination of coding ability, reasoning, and execution speed is kind of crazy.

  24. I hit my token limit three times a day on my max subscription - got sick of that and designed this MCP setup to shift some of the coding load to my local Qwen3.8-27B model. I've been iterating on it now for a bit, and thought I'd share it…

  25. This is mine, I'm on a 5090 with 128GB of DDR4 ram with a recent gen i3 cpu. Should I get a higher quant and lower my context down from 256K?

  26. Yet another vLLM fork thread here, but this time its for older INT8-centric hardware. This is a complete INT8 serving stack for Qwen3.8 27B based on vLLM, AITER, and a 27B GPTQ INT8 quant w/ DFlash2.

  27. Do I have something wrong in my settings? I keep having to restart vLLM to get Qwen3.8-27b-FP8 to stop outputting nonsense / garbage.

  28. curl -s https://api.github.com/repos/ggml-org/llama.cpp/pulls/27742 | jq -c '{draft,state,merged}' for r in unsloth/GLM-5.3-Flash-GGUF unsloth/Qwen3.8-Flash-Next-GGUF; do echo "== $r" curl -s "https://huggingface.co/api/models/$r" \ | jq -…

  29. *part 2 of an earlier post: previous quant comparison with voxel island creation this time I rented three rtx pro 6000 96gb, on each one I launched a qwen 3.8 27b quant and gave them 4 identical prompts: classical pool game air hockey 1v1…

  30. I have a strix halo, 128GB using GTT = 120GB usable = 105GB models and below, and this has served me well, but as the world changes I want to improve a bit. Your thoughts are welcomed !

  31. The news about Qwen 3.8 Flash Next is the first I'm reading about n-gram tables. I may be completely misunderstanding how they work but it seems they could open the door for 1T+ parameter models to be run on a single server with modest GPU…

  32. Like wtaf? Qwen 3.8 27b is crazy.

  33. Hi all, I'm looking for the best model for a hobby project and trying to make sense of the various data I came across. I know benchmarks do not often translate to the real world, especially to your particular use case (whatever it may be).

  34. The Artist: Qwen3.8-27B-UD-Q3\K_XL, q8_0 caches, xhigh, temp 1.0, image-min-tokens 1024, froggeric template) I was screwing around with different Qwen3.8-27B quants and thought of this very simplistic but seemingly bechmaxxing resistant co…

  35. we made vision mlx quants of qwen3.8 27b (9 builds from 8bit at 29.5 GB down to 3.23bpw DWQ at 11.8 GB) and compared them against other community vision mlx quants from hf (we only compared vision builds) the layout comes out of a clipping…

  36. I come from Qwen 3.6 35B-A3B + cline plugin of vscode at full 256k context but it eats 100k of context just to say "write me a python script that say print("hello")" so when i upgraded to Qwen 3.8 27B i switched to pi.dev i'm trying to und…

  37. https://preview.redd.it/kbsqh6f7molh1.png?width=730&format=png&auto=webp&s=068dbea9a50be634a369d54d8b27b781d020fab3 My experience with Qwen 3.8 for agentic tasks has been phenomenal but I personally feel that 3.7 flash is more reliable for…

  38. In the past, I have use llama.cpp, but I read that the exl3 quantization format should give better precision, so I have tried exllamav3/tabbyAPI. It was able to write the shown simple HTML game without interaction after asking some questio…

  39. Benchmarked qwen3.8 xhigh, medium and muse glimmer. Xhigh effort mode with qwen3.8 took almost 30hrs.

  40. I saw that Q2 is actually very good and produce real good results and I also saw how dflash2 make its running at generating >60 t/s with a 120k context lenght. And I like what its doing!!

  41. We're releasing a fully quantized NVFP4 version of Qwen3.8-27B. The checkpoint was trained using quantization-aware distillation (QAD) with QUASAR, our new QAT algorithm.

  42. Man I love the quality of 27B int8 (currently running lued/Qwen3.8-27B-INT8-W8A16-MTP). But man, it is not fast and I am seriously skating a line with memory consumption on my A40 at 256K context (KV at FP8).

  43. I’ve been going back and forth on this for a week and I can’t settle it, so I’m hoping someone here has hands-on numbers. The two configs (German prices, dealer quote, incl.

  44. Generated on a 3090 Qwen 3.8 27b Q4 Thinking high.

  45. With hopes of a Qwen3.8-35B-A3B release now mostly dashed, many people including myself are looking at fine-tunes and other variants of Qwen3.6-35B-A3B to run on VRAM-limited hardware. I decided to try to benchmark some of the top contende…

  46. So AI cloud costs are creeping major players it looks: https://www.perplexity.ai/hub/blog/introducing-portable-computer-for-local-first-ai Perplexity is going to use Qwen models, not specified which - could be 27B but could be the upcoming…

  47. How to get it to work locally? https://x.com/Alibaba_Qwen/status/2088280182356611304 https://x.com/arena/status/2092301580091711491

  48. LM studio and bionic don't load into GPU fully ( Ollama does) and it crashes BSD ( Ollama Does not), with stop code: WHEA_UNCORRECTEABLE_ERROR (0x124), i am using the default load setting, all updated LM studio and drivers, What do i need…

  49. I've made a few changes here and there to get nearly 2.1k tokens/sec prefill, ~210-230 tokens/sec decode with MTP7 (configurable, extended up to 15) on benchmarks. Also added disk caching options, up to 30GB per config by default for near-…

  50. Qwen3.8-27B on an IGX Thor with an RTX PRO 6000 Blackwell (Max-Q) Spent a few hours bringing up a self hosted inference box on an NVIDIA IGX Thor and couldn't find any numbers for this hardware combination, so here are mine. All of it is f…

  51. Prepare your disk space guys

  52. ModelScope——汇聚各领域先进的机器学习模型，提供模型探索体验、推理、训练、部署和应用的一站式服务。在这里，共建模型开源社区，发现、学习、定制和分享心仪的模型。

  53. We have been working on some performance optimisations for Qwen3.8 and other models. The main new feature that we introduced is adaptive speculation for Llama.cpp What is it?

  54. I want to update my computer to run Qwen 3.8 27b a bit better. My end goal is to run something like 3x 4090 48gb vRAM but I want to do things 1 step at a time to give myself time to test out stuff.

  55. I hear us all loud and clear qwen 3.8 is no toy. But for someone like me who dives in and out over the months, i can't work out what exactly is worth trying to get to hyped performance.

  56. And if so, what speeds are you getting for like, ~64k-token prompts? With what set ups?

  57. LiveCodeBench v6, medium and hard only. 132 problems, 80 hard and 52 medium.

  58. Currently the following llama-server settings are in use for Qwen 3.8 27b, it is running on a single RTX 6000 Pro, which allows running the full bf16 + 256kb context in bf16 cache. I am unsure if the current settings are optimal, they are…

  59. TL:DR: power management was doing auto-suspend and dumping model weights into RAM. Models weights 18gb+ RAM 16gb = no Bueno immediate crash.

  60. Howdy folks, You might (or likely not) know me around here with shilling of Pi harness, and the use of Pi as productivity assistant and KB manager. Lately, I have also been telling anyone who listens to try Qwen 3.8 27B IQ3_K_XXS by Unslot…

  61. Has anyone managed to run dflash2 with amd gpu for running qwen 3.8 27b

  62. Some data showing the value of tools on the professional intelligence of Qwen 3.8 27b. My workload with Qwen focuses on real estate, private equity, and adjacent finance stuff.

  63. My use case is agentic coding. I'm a developer by trade and I like having a home lab for projects.

  64. I'm open to the idea that I'm completely off-base here, but...I decided to do some qualitative, finger-in-the-air experimentation with this. Obviously, with thinking turned on, 3.8 27B just takes an age to get anything done.

  65. What's the best harness for agentic workflows that's not coding related at all? My work involves digesting a set of documents, analyze/evaluate them, and produce certain set of work product documents, mostly for due diligence purposes.

  66. Ornith does really well. TielCoder (https://llm-bench.io/benchmarks/cmt7kp2zj002r01lcmpchvlko) might be even a bit better in coding.

  67. With vision enabled I am able to run at 150k context on a single RTX 3090 and the results are just amazing. I was even able to write a gmail plugin for DeepSeek harness with locally hosted Qwen 3.8 27b.

  68. TL;DR: I'm planning to spend around $100 on cloud GPUs to benchmark Qwen3.8-27B with a focus on questions that actually matter when running it locally: different quant levels/providers, 8-bit vs 16-bit KV cache, GGUF vs EXL3, context lengt…

  69. The thinking in Qwen3.8-27B sometimes is in caveman speech (no verb conjugation, no articles, short phrases...) but sometimes it is not. Could this be because it is not fully finetuned or by RL to be fully caveman?

  70. I got Qwen3.8-27B running locally and connected it to EvoX. Basic chat works.

  71. Exciting news: Qwen3.8-27B by @Alibaba_Qwen just landed in Code Arena: WebDev at #9 overall with 1595 pts. It is the only model in its size class in the top 10, and also reshapes the Pareto Frontier!

  72. I’ve tweaked every parameter I could think of be it thinking level or the statistical knobs. I’ve tried chat templates, agents.md, caveman, etc.

  73. I have Qwen3.8 27B running across two machines with llama.cpp RPC. The main PC has an RTX 4070 Ti with 12 GB VRAM, and the worker is an M5 MacBook Air with 16GB unified memory.

  74. Qwen3.8-27B is amazing, but it’s slow. A stronger 35B-A3B Mixture of Experts-coder that can run and solve real codebase issues fast (even on constrained hardware) is a valuable addition to the arsenal.

  75. Thanks to Unsloth Dynamic 3.0 quants coming in slightly leaner and better preserved, I settled on Qwen 3.8 27B (`UD_Q4_K_XL`) at 100K context as my daily driver for Hermes Agent and OpenCode. On an RTX 5070 Ti Mobile (12GB) paired with an…

  76. Graph: A cumulative metric of Speed x Context Size - the higher the dot - the better. Helpful for looking for the most balanced solution.

  77. This shows how Qwen3.8-27B smashed all expectations and in less than a month got more likes than the next 4 models combined, finally surpassing their long reigning queen QwQ-32B. Next best thing is Qwen3.6-35B-A3B (which they have been sle…

  78. Hi folks, with tp=2 I get the following logs out of vllm with official Qwen3.8-27b-FP8 with MTP3: [vllm] | (APIServer pid=1) INFO 08-24 05:47:40 [loggers.py:310] Engine 000: Avg prompt throughput: 198.5 tokens/s, Avg generation throughput:…

  79. Hi friends, I've just spent the last few hours to ultimately fail to run Qwen3.8 27B in a docker container. Would anyone happen to have a working docker file they could share the contents of?

  80. Qwen3.8-27B has upped the value of all hardware. It is unbelievable that on 48gb of vram I can run Opus 4.5 at up to 130 t/s decode, 4000 t/s prefill, 5 concurrencies and as a bonus (with SGlang) I offload prefixes to storage so I basicall…

  81. I'm running on pi.dev with just two plugins: web search and token speed. I used to use caveman and ponytail but anecdotally think they hurt more than help.

  82. I commented on another Qwen 3.8 27B post that I was frustrated getting anything to work. You all gave some great comments.

  83. https://preview.redd.it/70kwcf3q08lh1.png?width=900&format=png&auto=webp&s=64939396cb7678494b2c33b035cb85e38333f87e I released LifeOS, a self-hosted personal organiser you mostly talk to! You say something out loud, a local LLM reads it, a…

  84. The user is saying: "The game is a mess. First, check the screen, take a screenshot, and verify whether each of the requirements is satisfied." I never said anything like that 😄 On serious note: that mean that Qwen team used probably their…

  85. Are you using reasosning budget? I found that sometimes this model thinks so much that it triggers max output tokens limit in pi which I set to 128k tokens 😂 I saw this in the Qwen team repo: https://github.com/QwenLM/Qwen3/blob/main/docs/…

  86. Hi all Looking at a PowerColor R9700 for Qwen3.8-27B, Q4, llama.cpp/Vulkan. AMD's own blog quotes 51.8 tok/s but doesn't say what context length that's at, or whether MTP=2 was holding up.

  87. tl;dr - split one VM (with 3 pooled GPUs) into two VMs (1 dedicated + 2 pooled); worth every second spent spooling up the extra VM. Some numbers: Qwen3.8:27b went from 12.2 tok/s to 33.91 tok/s muse-glimmer:30B went from about 14.3 tok/s t…

  88. I haven't had a lot of luck in the reliability of the output of any models so far for local coding use, and find myself relying on Claude's limited free sessions to assist with coding. My coding application varies, I'm primarily using C, C…

  89. Hi all, Might be a newbie question. So sorry for that already.

  90. Hi all, I made a post regarding how much Qwen 3.8 has improved over 3.6: https://www.reddit.com/r/LocalLLaMA/comments/1vqm51f/long_review_qwen_38_27b_is_very_good_at_tapping/ I made a very thorough write-up of how Qwen 3.8 compared not onl…

  91. I put together a source-linked, LLM-generated synthesis of one week of community testing and discussion around Qwen 3.8 27B: I’ve moved the full source-linked Qwen 3.8 synthesis to https://www.reddit.com/r/LocalLLaMa_V2/comments/1vvzckl to…

  92. Sharing my current Qwen3.8 setup. Curious what you all think - goal was an efficient configuration suitable for real-world tasks.

  93. I will be running qwen 3.8 model on a shared university machine for some research work, mostly using llama cpp but I am open to using other inference engines. I would like that there is no chat history or application logs saved on remote m…

  94. I've only been chatting with Qwen 3.8 27b to understand what it can do, and it's obviously decent at a lot. But I'm yet to actually use it on something that I really care about, or to build a project.

  95. Hi, I am a hobby dev using currently qwen 3.8 27b on my strix halo machine for coding. I was wondering what is the best approach to speed up.

  96. I downloaded the Q4_K_M GGUF and got 5.7 tok/s. The same file did 81.5 tok/s on a 5090.

  97. Hey folks, Been doing some benchmarking on my dual-GPU setup and ran into a noticeable performance gap between LM Studio and a raw llama-server.exe build that I can't quite pin down. Also running into some confusion with how reasoning effo…

  98. I was just curious how the new qwen3.8:27b does on a hard C to HTML porting job against Opus 5 in a default Claude Code. The job: my fun side project is a procedural shooter in a single C file.

  99. There’s always lots of people asking for various specific new local models… I wonder if we could incentivize a specific model (e.g. Qwen3.8 35B MoE) via crowdfunding on Kickstarter or something.

  100. The new Qwen 3.8 27B is fantastic for local agentic use. The problem is, what makes it so good, being a dense model, also makes it slow.

  101. As an LLM for tool-based agents, Qwen 3.8 27B (Q6_UD_K_XL) with xhigh thinking did something unprompted that I haven't seen a frontier model do. It blew me away I wanted to share.

  102. How do you guys set it up , i constantly get the error : I tried increasing the contex to 142k and putting the contex size as 115k in DSH , it still did not compress correctly. I have 0 issues if i run it with llama.ccp , it can work for 2…

  103. Most YouTube benchmarks only show trivial tasks like generating landing pages or simple Three.js games. Is a local model like Qwen 3.8 27B actually capable of real-world systems programming—such as building GTK4 or Qt 6 applications in Rus…

  104. I hear it's common for heavy Claude/Copilot users to burn through $100–$200+ a day in AI usage. I personally can easily burn through $50/day doing work stuff.

  105. Hello, So I've been trying lots of combinations in that never-ending landscape of options and settings. I wanted a proper quant of 3.8 27B running as fast as possible on my 5090 at 400W, with vision and with as much KV-cache as possible an…

  106. Qwen 3.8 27B was one of the most highly-anticipated open-weights releases that I've seen in a long time, and like many others, I immediately got to work testing it out and playing with it when it dropped. I'm running it on a single Lenovo…

  107. Qwen 3.8 27B is great, however it takes me ages to do tasks on xhigh. I need Qwen 3.8 35B A3B.

  108. It's a complication of all the feedback for this model in the last 7 days. There is no rule against it and there is no pinned Qwen 3.8 post.

  109. There were issues with the chat template when it was released and many users, including myself, ended up using the froggeric template. It seems to work fine but I just wanted to check if that's still the one to use or if the official or un…

  110. Qwen3.8-27B (W4A16-AutoRound) on vLLM 0.27.1, two RTX 3090s, TP=2, no NVLink. Agent: Pi, straight to the local endpoint.

  111. I originally followed unsloth's Q4 distribution to make an nvfp4/mxfp6/mxfp8 tri-quant, but after testing mxfp6 was faster so I made it an mxfp6/mxfp8 split. added dflash2 also quantized to mxfp6, and added mxfp8 as supported kvcache data…

  112. This is something I've noticed. Back when GLM 5.2 and Kimi K3 launched, there was a media push on pushing how dangerous open source models are.

  113. I'm running Qwen 3.8 27B locally on a single GPU. I can push the context to 131k, but I'd rather run it faster at 64k if the agent can manage context properly.

  114. Further to my last post, https://www.reddit.com/r/LocalLLaMA/comments/1vldngi/tested_in_coding_bf16_muse_glimmer_vs_bf16_qwen36/, I bring this requested comparison of Qwen3.8 27B. Context: Both models are running at full FP16 KV-cache.

  115. I read a lot of posts like "I've given this prompt to my Qwen 3.8 27b and it one-shotted me a 3d game after 24 hours". I wonder, what's the best harness for tasks like this?

  116. I have dual r9700 set up with Ubuntu. With llama.cpp I'm getting about 40 tokens per second for single instance.

  117. Hey guys, Inco AI shipped DFlash 2 a few days ago with a drafter for Qwen 3.8 27B and a llama.cpp PR. I built the PR and ran it against plain decoding, MTP, the n-gram lookup drafters, and my July DFlash 1 numbers on Qwen 3.6 27B for 3 day…

  118. This is the Qwen3.8-27B setup I actually use every day on one RTX 5090. I wanted to write it down with enough detail that another 5090 owner can reproduce it instead of guessing which memory knobs I used.

  119. That actually makes so much sense though, maybe like a layer looped version or maybe a 35B A3B variant or smt but nah craziest plot twist if true... https://i.redd.it/c6yjw8sz1zkh1.gif

  120. Good afternoon, everyone! I wanted to show the work I've been doing around porting Ninfer over to the CMP170HX (Github) So, first, I do want to call out the amazing work that Neroued, Sergiuszm and specifically Don-Chad have all done, to n…

  121. Im looking to set up a new local llm (probably on unsloth studio as that seemed to be doing pretty well last time I tested it). This one won't need to do agentic coding or app building or anything (not this time) but instead more 'text' ba…

  122. As I’m navigating the best setup for using Qwen3.8 27b as both my main coder and my personal assistant, I want to hook it into Wiki & beyond but I don’t want to be reliant on an internet connection. What I’ve started doing is taking the ge…

  123. I took SergiioB's Intel Arc Pro B70 Inference Cookbook and made it into a Helm chart. It applies the patches onto the pinned vLLM version.

  124. could not extract summary

  125. Up until the release of Qwen 3.8 27B (and getting my CMP 170HX system running it at 90-100t/s), I'd been totally happy with 3.6 35B - it did everything I needed (mostly agentic coding), very few errors, no doom loops and 512k context witho…

  126. New/Old benchmark that provides a lot of answers for local LLM. I present to you a new test that I developed somewhat by accident: https://huggingface.co/cHunter789/Qwen3.8-27B-i1-IQ4_KS_KT-GGUF/tree/main/ctx-cliff Its original goal was to…

  127. I'm here to show some benchmarks while using llama cpp with an AMD V620 on Windows 11 via Vulkan & ROCM. The benchmarks were written out by AI, but are verified by myself to be correct.

  128. So, with the rise of Qwen3.8-27B, I'm wondering how I can utilize my 4x32GB AMD R9700 GPU setup to best run this model. The model runs great on 1 card, but I'm wondering if llama.cpp can somehow leverage the extra cards.

  129. What models and configs are we using? Please share here On windows, I am using this copium pared down model https://huggingface.co/Bucoid/Qwen3.8-27B-Uncensored-IQ4-XS-MTP-16GB-VRAM-GGUF with MTP disabled, q4 k/q4 v mmproj banished to CPU/…

  130. Hi, I have been using opencode with openrouter for quite a while now. Having read the success stories of using Qwen3.8-27B, I thought of trying it too.

  131. 12GB VRAM and 32GB RAM. I have tried qwen3.8-27b-ud-q4_k_m which runs at about 6 tokens/sec.

  132. QwQ was genuine next-gen performance usable on local hardware, but the massive required context (it's reasoning style was akin to "if I say every possible word, I'll notice the right one!") kinda made it unusable for agentic coding. It was…

  133. I am looking to buy a GPU to run Qwen 3.8 27b and I found an affordable 4060 16gb. I wonder if there's a compilation of GPU, quantization and t/s?

  134. Even the low preset is better than Qwen 3.7 plus or Qwen3.6-27B reasoning

  135. Artificial Analysis just benchmarked them and the scores are crazy good, proving the earlier success wasn't only enabled by overthinking.

  136. I finally made the move from LM Studio to vLLM thanks to this post https://www.reddit.com/r/LocalLLaMA/s/NmS9CgHvqz. I may not know what it all means yet but I’m going to start diving into the docs to learn as much as I can.

  137. Hi fellows fully-local halos, after manually following existing guides, I decided to build an LLM API endpoint installation and optimization guide that works even when autonomously followed by my pi agent, so I can install/experiment/reins…

  138. they are not overpriced.. considering the intelligence they now pack (qwen 3.8) and newer models to come and the run costs(electricity + degredation) ITS A STEAL !

  139. I’ve been testing Qwen3.8-27B as a possible replacement for the Qwen3.5-9B that I have been running on RTX 5060Ti 16G. The goal was not just maximum tokens/sec, but useful context capacity, reliable tool calling, multi-turn behavior, and v…

  140. Have seen some people say Qwen 3.8 still overthinks even when reasoning is set to low. Which on my case has been way better compared to 3.6, eveb on a 3 bit quant.

  141. There's a lot of information out there and to be honest its quite confusing. Anyone have a detailed recommendation for the above use case?

  142. A year after OpenAI's gpt-oss, Qwen 3.8 27B is the first local model I'd let write code in an agent harness — fast enough to use live, small enough to self-host, and seemingly twice as capable as last year's open weights on the same hardwa…

  143. We’ve covered GLM 5.2 very excitedly before, and Prof Jie Tang’s belief that there will be an open weights Fable-class model by end of the year (spot check - with 134 days left, there are now two 2-3T models (Qwen 3.8 Max and Kimi K3) with…

  144. I gave frontier LLMs a canvas and told them to draw a self-portrait. August 4, 2026 Here are the self-portraits for Fable 5, GPT 5.6 Sol, Kimi K3, and Qwen 3.8 Max - out of order.

  145. could not extract summary

  146. Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things 16th August 2026 Friday’s big release was Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba’s Qwen research lab. I’ve been looking forw…

  147. Rats! source reconstruction with local LLMs This repository is a work-in-progress reconstruction of the source code for RATS.EXE, the original Windows version of Rats!

  148. On a Mac? Get Balto for Mac First launch downloads the inference engine and model.

  149. Hypothetically speaking: if you have cpu friendly ultimate "content generator" what would you do? Not saying it's the holy grail btw: so not "break satoshi's wallet", not "cure cancer" or turn mercury -> gold but say it has the power of co…

  150. Qwen 3.8 27B (just released) running locally is pretty good. In Cloudflare OS it made me a collaborative whiteboard with real-time multiplayer cursor tracking.

  151. I asked @Alibaba_Qwen 27B to make @simonw's pelican svg but dolphin with a twist from @goodside's prompt! "create an svg where a dolphin is riding a bicycle while using vlogging with a phone" After ~ 36 mins of thinking, 32K tokens, the mo…

  152. We all know Claude's (and LLMs' in general) sense of time is ... inaccurate.

  153. 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,…

← all threads