What changed
Two independent measurements of Qwen3.8 27B landed a day apart, on setups with nothing in common. On 26 August 2026 Quesma published task-level scores across the Unsloth GGUF quantizations, at around $3,000 of Modal GPU time 1. On 27 August 2026 TerminalBytes published five-run Ollama timings for the same model on a Mac Studio M3 Ultra 2. Both land on 17 GB for the 4-bit Q4_K_M file 12, and both report that an older llama.cpp build will not run the model at all 12.
The agreement neither run set out to make
Quesma used a llama.cpp build from 16 August 2026, because earlier builds do not work for
this model 1. TerminalBytes says you need a build from the last couple of weeks, and
that older ones fail with unknown model architecture: 'qwen35' 2. Two engineers, two
harnesses, two machines, one wall. The GGUF architecture tag is the tell, and I think it
generalizes: for an open-weight model this fresh, the runtime is the scarce ingredient.
| Same machine, same 17 GB Q4_K_M quant 2 | qwen3.6:27b | qwen3.8:27b |
|---|---|---|
| Generation speed, 5-run average | 28.6 tok/s | 14.0 tok/s |
| Tokens used per answer | 1,950-3,340 | 890-1,090 |
Those timings cut against the obvious reading. Generation speed halved between the two model versions on one machine, and the newer model answers the same prompts in roughly a third the tokens, so wall-clock per finished answer is close to a tie 2. Compare finished answers before you compare tokens per second.
The reproducibility cost is the part that lasts. Quesma records that Unsloth replaced the v2 quantization files on 19 August 2026, so the exact files behind most of its scores are no longer available 1.
Impact on your team
This is a decision for anyone pinning a local-inference image this quarter. Pin three things together: the llama.cpp or Ollama build, the exact GGUF file revision, and the reasoning effort. Quesma warns that the effort setting matters a lot, that the default is xhigh, and that it can overthink 1; leave it out and your numbers stop meaning anything. The migration trap runs backwards from the usual one: no card got smaller, so a capacity plan built on 17 GB still holds 12, while the container you baked in July refuses the model outright 12. Rebuild the runtime first, then benchmark.