Qwen3.8-27B runs on one 24GB GPU—and Q4 matches the full model
Independent tests settle which GGUF to grab: near-lossless at 17GB, still usable near 10GB, worthless at 1-bit—plus why long context still fits.
Copy markdown
Q4_K_M is the sweet spot
Independent benchmarking puts Unsloth's Q4_K_M (17.1GB) neck-and-neck with the full BF16 weights—~91% on GPQA Diamond and ~78% on Terminal-Bench 2.1, matching the unquantized model. Quality holds down through 2-bit (~90% GPQA), then 1-bit collapses to random-guessing on reasoning tasks.
Pick your quant by VRAM
On a 24GB card (RTX 4090/3090), Q4_K_M at 17.1GB is the pick. On 16GB, IQ4_XS (15.7GB) fits, or Q3_K_M (13.8GB) for context headroom. On 12GB, UD-IQ2_XXS (9.0GB) is the only comfortable fit—usable, with visible loss you take knowingly.
Hybrid attention buys you context
Only ~16 of the model's layers use full attention; the rest are linear, so the KV cache runs roughly a quarter smaller than a typical 27B. That's why a 17GB quant still leaves room for long documents and agent histories on a single GPU.
Build this weekend: fully offline
Pull the Q4_K_M GGUF, load it in Ollama or llama.cpp, and you have a frontier-ish coding and research model running entirely on your own box—no API bill, no rate limits, no data leaving the machine. Enough for a private repo assistant or an offline deep-research agent.
Elsewhere: AI's contested Navier–Stokes proof
OpenAI published Lean ‘certificates’ for finite-time blowup in the 3D Euler equations—a step toward the Navier–Stokes Millennium Prize problem—reportedly generated by a 10,000-agent swarm. Mathematicians Buckmaster and Alpöge dispute the credit and say the headline result is not fully verified. It is topping Hacker News as builders argue over what ‘AI-proved’ should mean.