Key takeaway: For the Qwen3.8-27B GGUFs we measured on an R9700, MTP depth 2 was the better starting point than depth 4 for decode and coding. On the same UD-Q4_K_XL target it decoded faster and completed one more coding task, while depth 4 scored 0.75 points higher on MMLU-Pro. The Q6 reference did not improve coding or MMLU-Pro in this suite, so higher nominal quantization did not translate to a better result here.
About the model
Qwen's official Qwen3.8-27B card describes a 27-billion-parameter vision-language model with a hybrid attention design: linear-attention layers are interleaved with full-attention layers. It is a dense model, so all 27B parameters are active during inference. Qwen publishes the base weights under Apache 2.0. The card documents 262,144 tokens natively and says YaRN can extend the context to 1 million tokens. It also cautions that extended context needs appropriate runtime settings.
The card presents Qwen3.8 as a model for coding and agentic tasks, highlights tool use and multi-step reasoning, and reports its own benchmark results. It includes a trained multi-token prediction (MTP) head. Recommended generation settings depend on mode: thinking uses temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 0, and repetition penalty 1; instruct/non-thinking uses temperature 0.7, top-p 0.80, top-k 20, min-p 0, presence penalty 1.5, and repetition penalty 1. The card's reasoning-effort controls adjust reasoning depth separately from sampling.
For this benchmark we used Unsloth's Qwen3.8-27B GGUFs. The draft model was a separate Qwen3.8 MTP head used alongside the target; the target weights remained the Unsloth UD quant. MTP proposes a short continuation that the target checks in a batch. It changes generation speed, not the target model's underlying answer distribution when verification is applied correctly. Depth controls how many candidate tokens are proposed before verification; greater depth can add throughput when acceptance is high, but can also add overhead.
R9700 suite results
We used the same standard-v2 suite across these Qwen runs: 400 MMLU-Pro questions, 50 GSM8K questions, and six coding tasks. Sampling temperature was zero, each response was capped at 2,048 tokens, and parallelism was four. Parenthetical counts denote items that reached the output cap. The Q4 depth comparison uses the same target weights and isolates drafter depth more closely than the Q6 row.
Setup. AMD Radeon AI PRO R9700 (32 GB GDDR6); desktop AM5 platform, DDR5-6000, Windows 11; llama.cpp llama-server, Vulkan backend, build b11836; flash attention on; 131K context; ubatch 2048; four parallel slots sharing one unified KV cache. Tested October 2026.
| Qwen3.8-27B target and draft | MMLU-Pro | GSM8K | Coding | Decode | Suite wall time |
|---|---|---|---|---|---|
| UD-Q4_K_XL, MTP d2 | 86.75% (31 at cap) | 94% (1 at cap) | 5/6 | 62 tok/s | 149 min |
| UD-Q4_K_XL, MTP d4 | 87.5% (30 at cap) | 92% (3 at cap) | 4/6 | 56 tok/s | 127 min |
| UD-Q6_K_XL, MTP d4 reference | 86.25% (32 at cap) | 96% (1 at cap) | 4/6 | 51 tok/s | 146 min |
At the same Q4 target, depth 2 gained 6 tok/s over depth 4 and completed one more coding item. Depth 4 scored 0.75 points higher on MMLU-Pro and had a shorter full-suite wall time, but GSM8K was two points lower and three answers hit the cap rather than one. There is no single unambiguous winner across every suite metric. A separate short decode sweep with 768-token code and prose prompts showed the same speed preference: Q4 depth 2 measured 62 tok/s versus 56 at depth 4. Q6 at depth 4 was slower than Q4 at either depth and did not improve coding or MMLU-Pro in this set. It did score higher on GSM8K than Q4 depth 4, which is a reminder that a small test suite can move differently across task families. The higher-bit label alone is not a guarantee of better measured results.
Recommended run settings
For a balanced starting point on this R9700, use the UD-Q4_K_XL target with MTP depth 2, the tested 131K context, ubatch 2048, and four-way parallelism. If reproducing our score suite, keep temperature 0 and the same 2,048-token answer cap. For everyday thinking-mode generations, Qwen's model card recommends temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 0, repetition penalty 1. For direct instruct mode, it recommends temperature 0.7, top-p 0.80, top-k 20, min-p 0, presence penalty 1.5, and repetition penalty 1.
Depth 4 may still be useful for a workload with high MTP acceptance, but our evidence points to d2 as the reliable speed choice: a brief decode sweep on GSQ IQ3_S measured 66/61 tok/s for code/prose at d2 versus 57/44 at d4. That sweep uses a different target quant and should not be conflated with the main Q4 table. Our measured Q4 depth 4 suite finished sooner despite a lower decode median, showing how the wider test mix and run wall time need not track a short single-stream decode number. Try both depths on your actual prompts and record task completion quality as well as token rate.
Verdict
Qwen3.8-27B has a useful MTP path, but more draft tokens were not automatically better on our hardware. Q4 with depth 2 is the most balanced configuration in this measured set; Q6/d4 did not earn its larger footprint through better suite performance. We would choose d2 for a general local assistant and evaluate depth 4 only where workload-specific acceptance demonstrably offsets its slower decode.
Related reading: “MiniCPM-V 4.7 on the R9700: a fast multimodal MoE, with a DSpark caveat”; “Sage-A1.5 on the R9700: why thinking mode changed the score”; “Swift 1.5 versus stock Qwen3.8-27B on the R9700”; “Qwen3.8-27B GSQ-RCO IQ3_S: 11.3 GiB and a useful quality floor”; and “Flint 40B-A2.5B on the R9700: fast decode, longer work”.
Disclaimer. These are measurements from one machine, one server build, and one test suite on the stated date. Results vary with hardware, software version, settings, and load. They are not a general ranking or a recommendation.
