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 off
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 1,550.49 | 58.45 |
| 2,048 | 1,501.02 | 57.96 |
| 4,096 | 1,487.12 | 57.31 |
| 8,192 | 1,441.90 | 55.69 |
| 16,384 | 1,359.20 | 53.57 |
| 32,768 | 1,233.63 | 49.66 |
| 65,536 | 1,032.98 | 42.83 |
| 131,072 | 774.93 | 33.45 |
| 262,144 | 512.93 | 22.68 |
Concurrent-stream measurements
| Streams | Context depth (tokens) | Aggregate PP (tok/s) | Aggregate TG (tok/s) | TG / stream (tok/s) | Aggregate all (tok/s) |
|---|---|---|---|---|---|
| 1 | 8,192 | 1,441.90 | 55.69 | — | — |
| 2 | 8,192 | 1,566.70 | 95.02 | 47.51 | 254.23 |
| 3 | 8,192 | 1,604.74 | 115.28 | 38.43 | 302.39 |
| 4 | 8,192 | 1,611.07 | 129.08 | 32.27 | 333.76 |
| 5 | 8,192 | 1,588.22 | 141.75 | 28.35 | 360.83 |
| 6 | 8,192 | 1,607.79 | 146.12 | 24.35 | 370.94 |
| 8 | 8,192 | 1,625.89 | 158.16 | 19.77 | 397.20 |
Technical record and provenance
Run details
- Source run ID
- rtx3090x2-qwen3.8-27b-q4km-2026-09
- Measured date
- 2026-09-10
- Hardware
- 2× RTX 3090
- Runtime
- llamaCpp · llama-bench 3466812 (build 476)
- Weight placement
- gpu-resident
- Split ratio
- 1:1
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Batch size
- 2,048
- Microbatch size
- 2,048
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- None (standard decoding)
- Measured repetitions
- 3
- Notes
- dealignai Qwen3.8 27B Q4_K_M, fully GPU-resident across two RTX 3090s at 250 W each. Non-MTP llama-bench context rows and llama-batched-bench concurrency rows use Q8_0 K/V, Flash Attention, batch 2048/2048, tensor split 1:1, and three measured repetitions after warm-up. The separate live llama-server MTP sweep is preserved beside these raw logs but is not represented by the standard BenchmarkRowResult schema.
- Source article
- /local-ai/qwen3-8-27b-local-ai/
- Raw log path
research/benchmarks/logs/rtx3090x2/2026-09-10-qwen3.8-27b-q4km/