Qwen3.8 27B Quant Fits 24GB VRAM at 100k Context, Sparking Local Model Profit-Threat Debate
ChopSticksPlease · reddit · 2026-09-05
The poster credits the Qwen and Unsloth teams' Qwen3.8 27B UD Q4KXL quant: it runs with 100k context (Q8 KV cache) on a single 3090's 24GB VRAM and performs phenomenally well for agentic coding.
Their thesis: the real threat to Anthropic/OpenAI profits isn't another frontier model, but small fast local models that handle 80-90% of mundane coding work for hours at zero API cost. They also flag a market gap — Kimi-K3 is out of reach for most businesses and prosumers — and ask whether a dual/quad GPU server (96-192GB VRAM + >256GB DDR4) or a 2-4 machine DGX cluster could run MiniMax-M3 at >30tps for agentic coding. On their dual 3090 + 128GB rig, Qwen3.8-Flash-Next Q4 is fast but doesn't feel much smarter than the 27B, and their workhorse Minimax-M2.7 is too slow for coding at larger quants.
More from coding & agent
- 105 Real Bugs Tested: GPT-6.1 Sol Over 10x Cheaper Than GPT-5.6 Sol — PawelHuryn · 2026-10-23
- Dev builds SkiFree-style sled game in one prompt, praises DeepSeek Flash — victormustar · 2026-10-23
- Every ditched personal agents for a shared Claude Managed Agents coworker in one week — every · 2026-10-23
- WorkOS shows how to let AI agents safely sign up for your app with auth.md and ID-JAG — AI Engineer · 2026-10-23
- Devs on Reddit swap tactics for stopping prompt regressions in AI agents — SleepFlimsy6963 · 2026-10-23
- Claude Code + ComfyUI MCP workflow for client work: paid models or open weights? — newaccount47 · 2026-10-23