A four-minute video on the local AI models on a 32GB graphics card: what Qwen3.8 27B UD Q6_K is, how it compares with the other six files in the picker, and when to turn Reasoning on.
About 3:58 • text overlays • sources credited on screen • as of 10:05 PM PT, Saturday, October 17, 2026
Key takeaways (click to open)
Your selected model, Qwen3.8 27B UD Q6_K: a 22.0 GB file (BF16 original: 55.6 GB) plus a 0.93 GB vision file (Unsloth on Hugging Face). A third-party test found 98.1% of top tokens match BF16 on English text (ARCHsloth KL-divergence measurement). An independent test by Quesma found even 4-bit Q4_K_M matched BF16 on Terminal-Bench 2.1 and GPQA (Quesma).
Fit in 32 GB: NVIDIA’s datasheet for a 32 GB workstation card lists 896 GB/s of memory bandwidth (NVIDIA). Our estimate: UD Q6_K leaves room for roughly 100K tokens of context with a 16-bit cache, or about 190K with a q8_0 cache; UD Q6_K_XL (25.3 GB) about 60K or 110K. These are arithmetic, not measurements.
Speed: a third-party test of Qwen3.8-27B (UD Q4_K_XL) on a 32GB workstation card measured 24.5 tokens/s without MTP and 72.4 at short context with MTP (lcz.me forum). We estimate roughly 20–30 tokens/s for Q6_K without MTP; that is not measured.
The other files: Gemma 4 31B (Google; 17.7 GB QAT Q4_0), Muse Glimmer (Meta; 16.8 or 19.7 GB Q4), Bonsai 27B (PrismML; about 3.8 GB 1-bit, a rebuild of the older Qwen3.6-27B) all fit. Nemotron 3.5 Super VL (NVIDIA; 46 to 132 GB) does not fit 32 GB of video memory.
Independent ranking: Artificial Analysis Intelligence Index v4.3.2 scores Qwen3.8-27B at 34 (xhigh reasoning), 28 (medium) and 20 (non-reasoning), versus Nemotron 3.5 Super VL 26, Muse Glimmer 17 and Gemma 4 31B 15 (Artificial Analysis). Bonsai is not listed.
Recommendation: keep UD Q6_K for drafting, coding and vision; use Reasoning Off or low for quick drafts and On for coding and multi-step analysis; use a smaller Q4 build or a q8_0 cache for very long documents.
What could not be verified: which quant the plain “Qwen3.8 27B GGUF” entry is (likely LM Studio’s 16.8 GB Q4_K_M); which Bonsai build is in the picker; any independent quality measurement of UD Q6_K_XL versus UD Q6_K; measured Q6_K speed on a 32GB card; Nemotron speed with RAM offload; the “Bionic” app’s own reasoning settings. Benchmarks published by Alibaba, Meta, PrismML and Unsloth are company-reported. Meta’s Glimmer charts compare with Qwen3.6, not Qwen3.8.
Notes (click to open)
Context-fit numbers are our arithmetic from file sizes and a published KV-cache figure (about 2.3 GB per 32K tokens at 16-bit), with 1.5 GB assumed for the runtime and desktop. Not measured.
Community forum posts and Reddit tests are anecdotal.
Full credits appear in the video.
Original Rio Dell Times video. Company names and marks belong to their owners and are used for news reporting and education. Facts from cited sources.
