Qwen3.6 35B
unsloth/qwen3.6-35b-a3b
unsloth/qwen3.6-35b-a3b-gguf on HuggingFace ↗
Throughput
Decode tok/s after a fixed context fill (PBM) — same agent-style workload as bench v2.
Recipe
- Profile
- qwen36-q4-llama
- Engine
- llama.cpp
- Context
- 256k · q8_0 KV
- Served as
- qwen3.6-35b-a3b-q4
Run on your Spark
single node
shell
spark inference up qwen36-q4-llama
Why we run it
Golden fleet target — auto-scaffolded from recipe qwen36-q4-llama.
Bench notes
PBM 4k @ 46.8 tok/s — perfbench-metrics — profile=qwen36-q4-llama
Measurement history
Context ladder
Older bench-v2 / golden cells at each benched context window (single measurement).
| Context | KV | Throughput |
|---|---|---|
| @ 32k peak golden | — | 33.6t/s |
Benchmark runs
Recorded inference benchmark sessions for this model's profile.
| Date | Profile | Method | Avg | Session t/s | Range | Fill | Tool |
|---|---|---|---|---|---|---|---|
| 2026-06-28latest | qwen36-q4-llama | bench-agent-v2v2.0 | 33.6t/s | — | — | ~50,000 | fail |
| golden 256k/q8_0 @ 33.6 tok/s — fill~50000 — bench-agent-v2 — tool_ok=False | |||||||
| 2026-06-21 | qwen36-q4-llama | bench-agent | 48.6t/s | 48.7 · 48.6 · 48.6 | 48.6–48.7 | @ 32k | — |
| agent bench avg 48.6 tok/s over 3 sessions × 3 turns (2304 tok in 47.4s) | |||||||