A user compared the performance of two local large language models, DeepSeek V4.1 Flash and Qwen 3.8 FN, by tasking them with generating an interactive V8 engine simulation. DeepSeek V4.1 Flash, despite using a lower-quality quantization, completed the task in approximately 60 minutes, generating 250k tokens at about 94 tokens/second and maintaining a 346k context window. Qwen 3.8 FN, running with multiple subagents, achieved a higher aggregate speed of around 350 tokens/second, generating over 443k tokens from the main agent and 804k total context over 70 minutes, though it encountered issues with subagents and produced a less polished final output. AI
IMPACT Provides insight into the relative performance and capabilities of specific local LLMs for complex generation tasks.
RANK_REASON Comparison of two specific local LLM models on a benchmark task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
