{"slug": "qwen3-8-27b-and-radeon-r9700-and-13-year-old-pentium-actual-good-performance", "title": "Qwen3.8-27B and Radeon R9700 and 13-year-old Pentium = actual good performance", "summary": "A developer running Qwen3.8-27B on an ASUS Turbo Radeon AI Pro R9700 with 32 GB of GDDR6 VRAM in a 13-year-old Pentium G3420 system reports good performance, achieving usable speeds for code analysis tasks. The setup uses Ubuntu 26.04 on a USB SSD, llama.cpp Vulkan build b10586, and Unsloth Dynamic 3.0 GGUF quants, with the model outperforming expectations despite the outdated CPU and motherboard.", "body_md": "Updated August 23, 2026: new Unsloth GGUF files and updated benchmark results.\n\nIt’s time to go big or go home! I’ve bought ASUS Turbo Radeon AI Pro R9700, an RDNA4 graphics card with a whopping 32 GB of GDDR6 VRAM, right before yet another price increase.\n\nSadly, the only desktop computer I have is an ancient custom build with a 4th-gen Intel Pentium G3420, 8 GB DDR3 RAM, and a single PCIe 3.0 slot. The CPU socket is damaged; the PC works only when I push down on the CPU cooler while booting (so that POST won’t fail). I don’t have any SATA disks anymore, and of course NVMe wasn’t a thing back in the Haswell days.\n\nYesterday, a new Qwen3.8 model was released. It’s a dense model with 27 billion active parameters, vision support, and a built-in MTP (multi-token prediction) layer. If benchmarks are to be believed, it’s better than Claude Opus 4.6.\n\nI have some real-world experience with Qwen3.6‑35B‑A3B, the sparse / MoE version that runs at acceptable speeds on a laptop iGPU, so expectations are really high. Let’s try it out.\n\n## First impressions\n\nI’ve installed Ubuntu 26.04 on a USB SSD and downloaded all the updates. This cursed setup works just fine, even if some apps report that there may be compatibility issues:\n\n`Warning: radv is not a`\n\n`conformant Vulkan implementation`\n\nBy the way, a USB‑C cable matters A LOT. I have two such cables, both with no logos or markings. The first one is too slow, at only 42 MB/s. The other one is 10x faster, but every time you look at it wrong, it would lose a connection and the OS dies.\n\nI had to update VBIOS to make the graphics card quiet. This update changes the\nminimum fan speed from 30% to 12% (about 1000 RPM). Unfortunately, VBIOS\ncan be updated only on Windows, but you can temporarily\n[install Microsoft malware directly on SSD](https://gist.github.com/Alee14/e8ce6306a038902df6e7a6d667544ac9)\nwithout losing any data.\n\nThere is no official AMD control panel for Linux, but you can use a wonderful\n[LACT app](https://flathub.org/en/apps/io.github.ilya_zlobintsev.LACT) instead.\nI’ve set the fan curves to start at 12% and never exceed 40%. You can also\nchange power limits. The default is 300 W, but I haven’t noticed any\nperformance issues at 210 W, and it helps keep the temperatures below\n85°C without thermal throttling.\n\nQwen3.8 has multiple reasoning modes, with `xhigh`\n\nbeing the default. It behaves\nlike a caveman, which is a downgrade for me. With Qwen3.6 I could easily analyze\nthe reasoning and refine the prompt early if the model hallucinated. So it’s\nhelpful to switch to a `medium`\n\nreasoning mode.\n\nWe need answer user question. Need analyze code, explain functionality, rate quality. We need think in English same language. Need inspect code carefully. (…)\n\n## Testing\n\nFlathub’s\n[moderation code](https://github.com/flathub-infra/website/blob/1430248c92c81b485f8cbf5fc57340924ae55ff0/backend/app/moderation.py)\nis approximately 2350 lines of text (including empty lines) in a 90 kB\nfile, which is about 19700 input tokens. I don’t think the model has been\nbenchmaxxed against this file, so it should show the real performance.\n\nHow long would it take to churn through a large chunk of code and answer these two questions:\n\n*What does this code do?**How would you rate it?*\n\nI’ve limited the output to 2048 tokens and applied other `llama-cli`\n\nparameters:\n\n```\n--load-mode none\n--n-gpu-layers all\n--flash-attn on\n--fit off\n\n--ctx-size 131072\n--cache-type-k q8_0\n--cache-type-v q8_0\n--temperature 1.0\n--top-k 20\n--top-p 0.95\n--min-p 0\n--repeat-penalty 1\n--presence-penalty 0\n--reasoning on\n--reasoning-effort medium\n\n--spec-type draft-mtp\n--spec-draft-type-k q8_0\n--spec-draft-type-v q8_0\n--spec-draft-n-max 2\n--n-gpu-layers-draft all\n```\n\nTested with the official\n[llama.cpp Vulkan build b10586](https://github.com/ggml-org/llama.cpp/releases/tag/b10586),\nAMD Lemonade\n[llamacpp-rocm build b1314](https://github.com/lemonade-sdk/llamacpp-rocm/releases/tag/b1314),\nand\n[Unsloth Dynamic 3.0 GGUF quants](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF),\nwith 128k context size and an average MTP acceptance rate of 0.65 at 2 tokens\nahead. Sampler parameters are from the model card.\n\nNote that Q8_K_XL quant won’t fit in 32 GB VRAM, so Q8_K_L is used instead.\n\n## Results\n\n*(Wide tables can be scrolled on mobile.)*\n\n| Quant | Vulkan pp/s |\nVulkan tg/s |\nVulkan VRAM GB |\nROCm pp/s |\nROCm tg/s |\nROCm VRAM GB |\n||\n|---|---|---|---|---|---|---|---|---|\nWithout MTP |\n||||||||\n| Q4_K_M | 755 |\n26 | 19.4 | 809 |\n24 | 19.9 | ||\nWith MTP |\n||||||||\n| Q4_K_M | 648 | 39 |\n20.1 | 766 | 36 |\n20.6 | ||\n| Q4_K_XL | 645 | 37 | 21.1 | 743 | 34 | 21.6 | ||\n| Q6_K_L | 621 | 32 | 28.2 | 619 | 29 | 28.6 | ||\n| Q8_K_L | 650 | 30 | 30.5 | 726 | 29 | 31.0 |\n\n## Observations\n\nPeople report they can also generate about 30-40 tokens per second on a modern\nhardware. **Looks like you can buy a powerful graphics card, pair it with\nmediocre components, and still be happy.** No need to pay extra for a modern CPU\nand expensive DDR5 RAM, yet.\n\nQwen3.8 is based on a battle-tested `qwen35`\n\narchitecture and requires no custom\npatches for llama.cpp. It doesn’t crash, it has valid reasoning (but way too\nlong and with primitive language), and responses are correct and coherent. MTP\nworks, resulting in huge performance improvements on Vulkan. ROCm has faster\nprompt processing, but token generation speed is slightly slower, and it\nrequires a little more VRAM.\n\nI think it’s not worth trying to use 6-bit quants. Their performance is similar to 8-bit quants, if not worse, and it’s not like I’m going to run out of RAM.\n\n## What didn’t work\n\nAMD claims that\n[Qwen3.8‑27B is capable of up to 50tg/s](https://www.amd.com/en/blogs/2026/run-qwen-3-8-27b-on-amd-ryzen-ai-max-and-radeon-graphics-cards-day-0.html)\non a single R9700 on Windows. But it doesn’t work for me. Both Vulkan and ROCm\nfail to allocate memory when loading a model.\n\nI’ve tried hipfire, ROCmFPX, and vLLM with various container runtimes, but they do not work since Pentium G3420 doesn’t support AVX. Tough luck.", "url": "https://wpnews.pro/news/qwen3-8-27b-and-radeon-r9700-and-13-year-old-pentium-actual-good-performance", "canonical_source": "https://mateusznowak.dev/articles/running-qwen38-legacy-hardware/", "published_at": "2026-08-28 21:58:56+00:00", "updated_at": "2026-08-28 22:19:07.669312+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-infrastructure", "ai-tools"], "entities": ["Qwen3.8-27B", "ASUS Turbo Radeon AI Pro R9700", "Intel Pentium G3420", "Ubuntu 26.04", "llama.cpp", "Unsloth", "AMD Lemonade", "Flathub"], "alternates": {"html": "https://wpnews.pro/news/qwen3-8-27b-and-radeon-r9700-and-13-year-old-pentium-actual-good-performance", "markdown": "https://wpnews.pro/news/qwen3-8-27b-and-radeon-r9700-and-13-year-old-pentium-actual-good-performance.md", "text": "https://wpnews.pro/news/qwen3-8-27b-and-radeon-r9700-and-13-year-old-pentium-actual-good-performance.txt", "jsonld": "https://wpnews.pro/news/qwen3-8-27b-and-radeon-r9700-and-13-year-old-pentium-actual-good-performance.jsonld"}}