Local Benchmark: DeepSeek V4.1 Builds a Working V8 Engine at 94 tok/s While Faster Qwen 3.8 Fails
jesdga95 · reddit · 2026-11-10
A Redditor ran the same one-shot prompt (a fully interactive Three.js V8 engine simulation) on two locally hosted MoE models, using RTX 5090 + 5060 Ti and 128GB DDR5:
- DeepSeek V4.1 Flash (IQ1M quant): single agent, 60 min, 250k tokens at 94 tok/s, 346k context uncompacted. Engine ran on the first try with all controls working, though the cutaway was thin and small parts missing.
- Qwen 3.8 FN (NVFP4): main agent + 5 subagents, 350 tok/s aggregate, 70 min, 443k+ tokens, two compactions. Built more parts with a better cutaway, but two subagents claimed to write files without doing so, and the engine never actually starts.
Author's takeaway: DeepSeek is notably better at figuring things out autonomously; Qwen works well for refactors and feature additions but shouldn't be left unattended on critical tasks. Full prompt and playable builds are shared publicly.
More from coding & agent
- verity-gate 0.2.0 ships its agent action-vetting gate as MCP servers for pip and npm — W1RELESS · 2026-11-10
- HuggingFace SafeAgents: a safety layer between your harness and your model — blaizedsouza · 2026-11-10
- rakyll teases Agent Substrate as 'agentic compute fabric' ahead of KubeCon keynote — rakyll · 2026-11-10
- Developer uses a 1.9B decision model on a laptop to flag dirty restaurant tables in real time — JosephJacks_ · 2026-11-10
- Microsoft FORGE Lab uses agentic AI tools to find memory-safety bugs in Rust core libs — jedisct1 · 2026-11-10
- CTM-AI: open-source multi-agent framework inspired by consciousness theory, NeurIPS 2026 paper — pliang279 · 2026-11-10