I took some time this morning to replace the model with the newly released Qwen3.6-27B and tested it.
I used the Qwen3.6-27B-UD-Q4_K_XL.gguf model.
The execution options are
ExecStart=/usr/local/bin/llama-server \
-m /home/models/unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-UD-Q4_K_XL.gguf \
--mmproj /home/models/unsloth/Qwen3.6-27B-GGUF/mmproj-F16.gguf \
--alias Qwen3.6-27B \
--host 0.0.0.0 \
--port 8000 \
--temp 0.6 \
--top-p 0.95 \
--top-k 20 \
--min-p 0.00 \
--kv-unified \
--cache-type-k q4_0 \
--cache-type-v q4_0 \
--flash-attn on \
--fit on \
--ctx-size 262144 \
-n 32768 \
--api-key "Enter your desired api key" \
-ngl 99
I ran it with these settings and unlike the previous 35B model, I increased the ctx to 256k.
It's working well +_+;;
## Performance Benchmark (Measured)
| Category | Speed |
| Prompt Processing (Input Encoding) | 1,207 tok/s |
| Token Generation (Output) | 37.5 tok/s |
- 2,620 token input → Prefill 2.17 seconds
- Output generation fixed at 26.7ms/token
- reasoning_content included in output tokens due to thinking model characteristics
### Test Case Results
| Test | Input Tokens | Output Tokens | Time Taken | Notes |
|--------|-----------|-----------|-----------|------|
| Short Input | 20 tok | 600 tok | 15.9s | 37.8 tok/s |
| Medium Input (Code) | 20 tok | 800 tok | 21.1s | 37.8 tok/s |
| Long Input (~2600tok) | 2,620 tok | 1,000 tok | 29.0s | 34.5 tok/s |
I haven't tested it with real coding tests yet.
I have a feeling it will be better than Qwen3.6-35B-A3B.
So I'm even more excited about it.
