Making Qwen Faster on an RTX 3090
Updated with b11429 probabilistic MTP: 24–30% faster decode in matched screens, 68.4 tokens/s across eight coding attempts, and 7/8 strict deliveries.
Found 4 posts
Updated with b11429 probabilistic MTP: 24–30% faster decode in matched screens, 68.4 tokens/s across eight coding attempts, and 7/8 strict deliveries.
Two tuning stages on eight repeated development cases: raise the response cap from 8K to 32K, then compare medium with xhigh. More repairs finished, waiting increased, and confirmation remains unrun.
A controlled Qwen3.8-27B Q4_K_M and OpenCode comparison: MTP raised generation throughput by 56% and cut successful task time by 20–40%. Paired coding results explain why throughput alone is not enough.
Deploy Qwen3.8-27B Q4_K_M on an RTX 3090 24GB and connect it to OpenCode through llama.cpp. Measured 128K capacity, 20.9–36.4 tokens/s generation, and four coding tasks with successes, timeouts, and review findings.
Allow visitor analytics?
With your permission, we use Google Analytics and cookies to understand traffic sources, page engagement and download-link clicks. You can read everything without analytics. Privacy notice