Uploaded August 2026 | Updated September 2026, 2 weeks ago
🚀 The LangChain 10 Days FREE Bootcamp is live: 10 lessons, free AI models only, from your first API call to a production grade RAG agent. Start with Day 0 for the roadmap and setup.
📺 Full playlist: youtube.com/watch?v=KJ3_NExk7-Q&list=PLW4pPr9JCovI&index=1
----------
Qwen 3.8 27B vs Nemotron 3.5 Lightning vs Muse Glimmer, all three running locally on a single RTX 5090, tested on 16 hard math, coding and reasoning problems with verified answers.
Qwen 3.8 came out on 14 August, so I ran it against the two models that landed just before it. Same runner, same quantization, same context window, same temperature for all three. You will see the real scoreboard, how long each model took, how much VRAM it needed, how many tokens it burned, and which models got stuck thinking until they ran out of budget. This video covers the full benchmark from setup to verdict, so you can decide which 27B model to keep on your machine.
⏱ Chapters:
0:00 Intro and what we are testing
0:48 Official benchmark numbers from the model makers
2:05 Why paper benchmarks are not enough
4:04 Qwen 3.8 volcano simulation test
4:51 Specs compared: parameters, MoE vs dense, context window
5:44 Embedding width and VRAM each model needs
6:38 Vision input and separate thinking field
7:26 My test setup: Ollama, 64K context, Q4, 32K cap
7:56 The 16 hard problems: math, coding, reasoning
8:17 Scoreboard: how many each model solved
9:15 Where each model failed and who ran out of tokens
10:12 Efficiency: suite time and tokens generated
11:28 Where the Qwen 3.8 speed actually comes from
12:20 Where Qwen 3.8 falls behind on math
13:03 Token speed vs effective problem solving speed
14:45 Verdict on Qwen 3.8 for daily use
🔗 Resources:
Full write up with all 16 problems and the raw results: kgptalkie.com/tutorials/generative-ai/qwen-3-8-27b-vs-nemotron-3-5-vs-muse-glimmer
Ollama: ollama.com
My setup: RTX 5090 32GB, Ollama 0.32.12 on Windows 11, Q4_K_M for all three models, 65,536 token context, 32,768 token generation cap, temperature 0.2.
Results in short: Muse Glimmer 15/16 but took 16 minutes. Nemotron 3.5 14/16 in 8.3 minutes and needed 25GB VRAM. Qwen 3.8 14/16 in 6.3 minutes on 17GB, and it was the only model that never ran out of its token budget.
📺 Watch next:
Nemotron 3.5 Lightning vs Muse Glimmer on RTX 5090: Full Benchmark
youtube.com/watch?v=n74N6p5so7o
🎓 Go deeper with my Udemy course:
Master Langchain v1 and Ollama, Chatbot, RAG and AI Agents: kgptalkie.com/langchain
If this benchmark helped you pick a model, hit like. Tell me in the comments which model you want tested next, or drop a problem statement you want me to run, and I will benchmark it for you. Subscribe and turn on the bell so you catch the next local model test.
#Qwen3 #LocalLLM #Ollama #RTX5090
🚀 The LangChain 10 Days FREE Bootcamp is live: 10 lessons, free AI models only, from your first API call to a production grade RAG agent. Start with Day 0 for the roadmap and setup.
📺 Full playlist: youtube.com/watch?v=KJ3_NExk7-Q&list=PLW4pPr9JCovI&index=1
----------
Qwen 3.8 27B vs Nemotron 3.5 Lightning vs Muse Glimmer, all three running locally on a single RTX 5090, tested on 16 hard math, coding and reasoning problems with verified answers.
Qwen 3.8 came out on 14 August, so I ran it against the two models that landed just before it. Same runner, same quantization, same context window, same temperature for all three. You will see the real scoreboard, how long each model took, how much VRAM it needed, how many tokens it burned, and which models got stuck thinking until they ran out of budget. This video covers the full benchmark from setup to verdict, so you can decide which 27B model to keep on your machine.
⏱ Chapters:
0:00 Intro and what we are testing
0:48 Official benchmark numbers from the model makers
2:05 Why paper benchmarks are not enough
4:04 Qwen 3.8 volcano simulation test
4:51 Specs compared: parameters, MoE vs dense, context window
5:44 Embedding width and VRAM each model needs
6:38 Vision input and separate thinking field
7:26 My test setup: Ollama, 64K context, Q4, 32K cap
7:56 The 16 hard problems: math, coding, reasoning
8:17 Scoreboard: how many each model solved
9:15 Where each model failed and who ran out of tokens
10:12 Efficiency: suite time and tokens generated
11:28 Where the Qwen 3.8 speed actually comes from
12:20 Where Qwen 3.8 falls behind on math
13:03 Token speed vs effective problem solving speed
14:45 Verdict on Qwen 3.8 for daily use
🔗 Resources:
Full write up with all 16 problems and the raw results: kgptalkie.com/tutorials/generative-ai/qwen-3-8-27b-vs-nemotron-3-5-vs-muse-glimmer
Ollama: ollama.com
My setup: RTX 5090 32GB, Ollama 0.32.12 on Windows 11, Q4_K_M for all three models, 65,536 token context, 32,768 token generation cap, temperature 0.2.
Results in short: Muse Glimmer 15/16 but took 16 minutes. Nemotron 3.5 14/16 in 8.3 minutes and needed 25GB VRAM. Qwen 3.8 14/16 in 6.3 minutes on 17GB, and it was the only model that never ran out of its token budget.
📺 Watch next:
Nemotron 3.5 Lightning vs Muse Glimmer on RTX 5090: Full Benchmark
youtube.com/watch?v=n74N6p5so7o
🎓 Go deeper with my Udemy course:
Master Langchain v1 and Ollama, Chatbot, RAG and AI Agents: kgptalkie.com/langchain
If this benchmark helped you pick a model, hit like. Tell me in the comments which model you want tested next, or drop a problem statement you want me to run, and I will benchmark it for you. Subscribe and turn on the bell so you catch the next local model test.
#Qwen3 #LocalLLM #Ollama #RTX5090
