Local test: Qwen 3.8 Flash Next Q4 flops at Tetris coding vs 27B Q8
wgaca2 · reddit · 2026-10-17
A Reddit user benchmarked Qwen 3.8 Flash Next (UD-Q4KXL) on 4x 3090s against Qwen 3.8 27B Q8 on Tetris-building prompts, pushing back on community hype. The 27B model produced a working standalone .exe with two self-fixed bugs, while Flash Next under Strata ran 4 hours without a playable exe and under llama.cpp still shipped buggy output within a 1-hour cap. The author admits the Q4-vs-Q8 comparison is unfair but rejects claims that Flash Next at Q3 matches the 27B model. A shared results table shows GPT Luna/Astra passing all checks in 4-6 minutes.
More from coding & agent
- aiand_ sponsors Tokyo Agent Hackathon, offering Kimi and DeepSeek via compatible API — DavidBennett__ · 2026-10-17
- OpenCall playbook: agent swarm does on-call triage with human approval and auto fix PRs — please-dont-deploy · 2026-10-17
- Anthropic launches Claude Haiku 5.5: 75% cheaper, 39.2% on Terminal-Bench vs GPT-6 Luna's 16.4% — dl_weekly · 2026-10-17
- Spotify MCP Server lets AI assistants like Cursor and Claude control playback — modelcontextprotocol · 2026-10-17
- pqs-mcp-server scores any prompt across 8 dimensions before the model runs — modelcontextprotocol · 2026-10-17
- AI Industry Worker Says Nobody He Knows Uses xAI's Grok Build — SimonBalmain · 2026-10-17