Qwen3.8 27B
Qwen/Qwen3.8-27B
Qwen · 22 recorded configurations
Measured setup
Hardware
Runtime and tool
Quantisation
MTP draft depth
Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.
2× RTX 3090 · llama.cpp
Q4_K_M · MTP 4
PP · left axis · not measuredTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 12 | — | 71.87 |
| 2,048 | — | 113.13 |
| 4,096 | — | 82.33 |
| 8,192 | — | 79.72 |
| 16,384 | — | 68.12 |
| 32,768 | — | 70.15 |
| 65,536 | — | 65.02 |
| 131,072 | — | 50.95 |
| 261,888 | — | 39.72 |
Technical record and provenance
Run details
- Source run ID
- rtx3090x2-qwen3.8-27b-q4km-mtp4-2026-09
- Measured date
- 2026-09-10
- Hardware
- 2× RTX 3090
- Runtime
- llamaCpp · llama-server 0.3.0-dev (build 50, commit fc6545d)
- Weight placement
- gpu-resident
- Split ratio
- 1:1
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Batch size
- 2,048
- Microbatch size
- 2,048
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 4
- Measured repetitions
- 1
- Notes
- Live llama-server MTP sweep using the embedded draft-mtp head. One 32-token warm-up preceded each measured 256-token completion at each requested depth; one measured repetition. Prompt processing is intentionally omitted: the live prefilled-prompt protocol is not comparable to the llama-bench PP rows. The 261,888-token row is the actual measured prompt depth after the model context clamp.
- Source article
- /local-ai/qwen3-8-27b-local-ai/
- Raw log path
research/benchmarks/logs/rtx3090x2/2026-09-10-qwen3.8-27b-q4km/