{"slug": "qwen-3-8-vs-ornith-vs-nemotron-vs-muse-glimmer", "title": "Qwen 3.8 vs Ornith vs Nemotron vs Muse-Glimmer", "summary": "A developer benchmarked four open-source models—Qwen 3.8 27B, Ornith 1.5 35B-A3B, Nemotron 3.5 Lightning 30B-A3B, and Muse-Glimmer 30B—on coding and computer-use tasks. Qwen 3.8 proved the most accurate, matching GPT-5.6 Luna in computer use, while Ornith impressed with speed and efficiency, achieving 50 TPS on Strix Halo. Nemotron and Muse-Glimmer lagged in coding and reliability, respectively.", "body_md": "Srovnání Qwen 3.8 27B, Nemotron Lightning 30B, Ornith 35B a Muse-Glimmer 30B v reálných testech. Qwen 3.8 ovládá coding, Ornith překvapuje rychlostí, Nemotron a Muse-Glimmer ztrácejí.\n\nUživatel provedl detailní srovnání čtyř populárních open-source modelů na coding úloze. Test běžel 25 minut a na Qwen 3.8 vyprodukoval 50 000 tokenů — jde o těžký thinking model, ale výsledek je fenomenální. Ornith se ukázal jako velmi schopný model, zejména vzhledem k dané rychlosti.\n\nVýsledky detailně: [llm-bench.io porovnání](https://llm-bench.io/compare/runs?runs=cmt6ecf8g000001p45vwzux53%2Ccmt6ergk5000701p41hqdyy78%2Ccmt6f2oob000e01p49o9592cb%2Ccmt6fqddm000l01p4l1vm7skd)\n\n**Qwen 3.8 27B** — jasný vítěz v codingu a architektuře. V testech computer use MCP byl bezchybný, na úrovni GPT-5.6 Luna. Jak uživatel popsal: Qwen3.8 was flawless, on par with gpt 5.6 luna. Ale je pomalý — na některých strojích sotva 15 TPS. Pro náročné úlohy, kde záleží na přesnosti, je to jasná volba. Podle dalšího uživatele: Glad to see i was justified in focusing mainly on 3.8 since its release.\n\n**Ornith 1.5 35B-A3B** — překvapení soutěže. Ornith is seriously amazing especially given the fact it only has 3b active. Na Strix Halo dosahuje 50 TPS při 50k+ tokenech a zvládne plný kontext. Podle jednoho uživatele: After more testing Ornith is absolute sorcery, this is GPT oss20b levels of optimization. 50TPS on strix halo at 50k+ tokens and can load full context, compared to Qwen 3.8 that struggles to serve 15TPs, and I honestly can't notice the difference in front end work maybe 2% at best.\n\n**Nemotron 3.5 Lightning 30B-A3B** — určen pro ne-technické agentic úlohy. Nemotron is for non technical agentic tasks. Solidní rychlost, ale v codingu zaostává.\n\n**Muse-Glimmer 30B** — zaměřen na technické psaní. Muse Glimmer take care of 3 stages of technical writing. S dflash běží rychleji (150-200 TPS na 3090), ale v testech computer use dělal chyby: Muse glimmer was almost good, but would miss some step, misclick some button or forget to activate a window and that would doom it.\n\n**Zdroj:** [Reddit](https://www.reddit.com/r/LocalLLM/comments/1vwmeow/qwen3827b_nemotron35lightning30ba3b/)", "url": "https://wpnews.pro/news/qwen-3-8-vs-ornith-vs-nemotron-vs-muse-glimmer", "canonical_source": "https://dev.to/petr_baloun/qwen-38-vs-ornith-vs-nemotron-vs-muse-glimmer-1na4", "published_at": "2026-08-25 10:56:55+00:00", "updated_at": "2026-08-25 11:14:15.598019+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "machine-learning", "ai-products", "ai-tools"], "entities": ["Qwen 3.8", "Ornith 1.5", "Nemotron 3.5 Lightning", "Muse-Glimmer", "GPT-5.6 Luna", "Strix Halo", "Reddit", "llm-bench.io"], "alternates": {"html": "https://wpnews.pro/news/qwen-3-8-vs-ornith-vs-nemotron-vs-muse-glimmer", "markdown": "https://wpnews.pro/news/qwen-3-8-vs-ornith-vs-nemotron-vs-muse-glimmer.md", "text": "https://wpnews.pro/news/qwen-3-8-vs-ornith-vs-nemotron-vs-muse-glimmer.txt", "jsonld": "https://wpnews.pro/news/qwen-3-8-vs-ornith-vs-nemotron-vs-muse-glimmer.jsonld"}}