Hollow markers & dashed spokes: axis not yet scored
G
93
A
87.5
U
—
N
91.2
T
97.1
L
93.8
E
—
T
54.7
Specification
- Parameters
- 35B total / 3B active per token
- Architecture
- qwen3_5_moe
- Size on disk
- 20.43 GB
- Quantization
- 4bit
- Format
- MLX
- Reasoning (CoT)
- Yes — emits reasoning tokens
- Internal ID
- M8
- Mean speed
- 89.4 tok/s across suites
- Stall census
- 1 stall in 35 observed tests (2.9%)
- Reasoning appetite
- 2,471 tokens mean · 16,377 max
- Model card
- lmstudio.ai
Suite results
| Suite | Score | Avg / 20 | Tests | tok/s |
|---|---|---|---|---|
| General capability (13-task real-workload suite) 1a | 241.8 / 260 | 18.6 | 13/13 | 119.9 |
| Agentic tool-calling & protocol adherence 1b | 140 / 160 | 17.5 | 8/8 | 108.9 |
| Long-context retrieval & synthesis 1g | 120 / 120 | 20 | 6/6 | 74.9 |
| Long-context multi-needle (MRCR) 1g2 | 44.1 / 60 | 14.7 | 3/3 | 53.8 |
| Live one-shot builds (runtime-verified) 1h | 75 / 80 | 18.8 | 4/4 | — |
Each test is LLM-judged 0–20 against a fixed rubric; suite max = tests × 20. Where fewer tests ran than the suite total, unrun tests count as zero toward the axis score. See methodology.