How I Got Frontier Performance on 16GB VRAM (Qwen 3.8 27B)

Qwen 3.8 27B really can hit frontier scores on a 16GB GPU, but only at xhigh reasoning effort, and most 16GB VRAM guides quietly turn that down to medium. Same weights, same file, one line in the config, and the score drops from 34 to 28. Artificial Analysis ran Qwen 3.8 27B at every reasoning effort: xhigh 34, medium 28, low 26, thinking off 20. At 34 it beats February's frontier models (Claude Opus 4.6 Max about 32, GPT-5.3 Codex about 32.5, Gemini 3.1 Pro Preview 30). At 28 it loses to all three. Qwen's own chat template defaults to xhigh, gives low and xhigh written instructions, and gives medium nothing at all. In this video I go through why the popular 16GB setups run medium (or a 512-token reasoning budget), what xhigh actually costs you in time on an RTX 5060 Ti or RX 9070, how much context a 16GB card really holds next to the model, and how Swift 1.5 from UkisAI gets the thinking down without turning the setting down. I also cover the tests where xhigh bought nothing, and the 24-point gap to today's top model. WHAT YOU'LL LEARN - Why the "frontier" score only exists at xhigh, and what medium scores instead - The 86.6 Terminal-Bench number that isn't on Qwen's card (it's 73.0, and Opus 4.6 Max wins at 78.2) - Why the "52 on Artificial Analysis" figure is from an older index version (the current score is 34) - What xhigh costs on a 16GB card: about 15,000 thinking tokens per hard question, around five and a half minutes - How much KV cache fits on 16GB (about 110K tokens at 8-bit, 128K at 4-bit) - Swift 1.5: 15,014 to 8,717 thinking tokens on GPQA Diamond at xhigh, with the score slightly up - When low effort is just as good, and when xhigh earns the wait - The exact 16GB setup I'd run, and when renting is the better deal MY 16GB SETUP - Model: Swift 1.5 Qwen 3.8 27B, GSQ-RCO IQ3_S with MTP (12.12 GB) - Reasoning effort: leave it on the default (xhigh) - Run your monitor off the integrated GPU to free 1-2 GB of VRAM - Skip the vision projector unless you need vision - Largest context that fits: ~110K at q8_0 KV cache, ~128K at q4_0 - Multi-token prediction on - Use effort as a dial: low for quick jobs, xhigh for hard ones CHAPTERS 00:00 The annoying afternoon 02:25 What's actually right about Qwen 3.8 27B 04:45 Two numbers that are wrong 06:31 The one setting: reasoning effort 08:45 xhigh vs medium vs low, scored 10:52 Why the 16GB guides turn it down 13:08 The context the benchmarks used 15:02 Swift 1.5: thinking less without turning it down 17:28 When xhigh buys you nothing 19:03 The 16GB setup I'd run 21:14 February's frontier, not today's 23:28 The one-line version SOURCES Qwen3.8-27B model card and chat template: https://huggingface.co/Qwen/Qwen3.8-27B Artificial Analysis, Qwen3.8 27B (xhigh / medium / low / non-reasoning): https://artificialanalysis.ai/models/qwen3-8-27b Artificial Analysis, Claude Opus 4.6 (max): https://artificialanalysis.ai/models/claude-opus-4-6-adaptive Artificial Analysis, GPT-5.3 Codex: https://artificialanalysis.ai/models/gpt-5-3-codex Artificial Analysis, Gemini 3.1 Pro Preview: https://artificialanalysis.ai/models/gemini-3-1-pro-preview Artificial Analysis, Claude Opus 5.5: https://artificialanalysis.ai/models/claude-opus-5-5 Swift 1.5 (UkisAI): https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b Swift 1.5 GSQ-RCO GGUF: https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF llama.cpp server options (reasoning budget, fit target): https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md ARC Prize on Qwen3.8-27B effort levels: https://x.com/arcprize/status/2105751611570069908 Benjamin Marie, 23,130 prompts per effort: https://x.com/bnjmn_marie/status/2092728700437688637 4,800 tasks on four RTX 3090s (superalesha): https://x.com/superalesha/status/2090318716693074390 VulcanBench results: https://x.com/morganlinton/status/2091173950651080765 xhigh vs medium thinking tokens: https://www.reddit.com/r/LocalLLaMA/comments/1vohpc8 RTX 5060 Ti 16GB, 45.6 / 31.3 tok/s: https://www.reddit.com/r/LocalLLaMA/comments/1vupiyh RTX 4070 Ti Super 16GB setup: https://www.reddit.com/r/LocalLLaMA/comments/1w1lq7u RX 9070 16GB, 128K context at q4_0: https://www.reddit.com/r/LocalLLM/comments/1x14c5u Swift vs Qwen, Aider suite: https://www.reddit.com/r/LocalLLaMA/comments/1wp5vqr Swift 1.5 on RX 9070 XT, 43.7 / 37.4 tok/s: https://x.com/TeksEdge/status/2104361696869703933 OpenRouter Qwen3.8-27B pricing: https://openrouter.ai/qwen/qwen3.8-27b #Qwen #LocalLLM #LocalAI