Model benchmarks
KAT-Coder-V2.5-Dev-OptiQ-4bit local LLM performance
As of October 2026, KAT-Coder-V2.5-Dev-OptiQ-4bit runs at up to 100.5 tok/s for local inference (best of 2 community benchmark runs across 1 GPU).
Model size
35B
Peak speed
100.5 tok/s
Average speed
87.5 tok/s
Avg PP
1475.0 tok/s
Min memory
21.2 GB
Max context
262,144 tokens
Avg output / run
12,975 tokens
Avg runtime / run
2m 38s
Avg quality
80.8
Benchmark runs
2
GPUs tested
1
Quality by task
Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 2 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.
| Task | P5 (low) | Avg | P95 (high) |
|---|---|---|---|
| Overall | 80.7 | 80.8 | 80.9 |
| Agent Workflow | 80.4 | 82.8 | 85.2 |
| Code Generation | 74.1 | 74.5 | 75.0 |
| Role Play & Narrative | 76.0 | 80.4 | 84.7 |
| Research & Analysis | 84.3 | 85.4 | 86.6 |
Performance by hardware and tool
Every hardware/tool/quantization combination KAT-Coder-V2.5-Dev-OptiQ-4bit has been benchmarked on, ranked by peak token generation speed. Last updated October 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| Apple M4 Max | oMLX | 4bit | 100.5 tok/s | 100.5 tok/s | 21.3 GB | 262,144 tokens | 80.9 | 1 |
| Apple M4 Max | oMLX | — | 74.5 tok/s | 74.5 tok/s | 21.2 GB | 65,536 tokens | 80.7 | 1 |
Benchmark runs
All 2 KAT-Coder-V2.5-Dev-OptiQ-4bit runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is KAT-Coder-V2.5-Dev-OptiQ-4bit good for coding?
- In our benchmarks, KAT-Coder-V2.5-Dev-OptiQ-4bit scores 74.5/100 for coding. It runs at about 87.5 tok/s, so if you want more speed, Ornith-1.5-35B-A3B.APEX-I-NanoPlus is faster (~110.5 tok/s) and still scores well for coding (81.1/100).
- Is KAT-Coder-V2.5-Dev-OptiQ-4bit good for agentic (tool-using) tasks?
- In our benchmarks, KAT-Coder-V2.5-Dev-OptiQ-4bit scores 82.8/100 for agentic workflows. It runs at about 87.5 tok/s, so if you want more speed, cyber-tiel-coder-35b-a3b-mtp is faster (~115.9 tok/s) and still scores well for agentic workflows (86.4/100).
- How fast is KAT-Coder-V2.5-Dev-OptiQ-4bit for local inference?
- Across 2 community benchmark runs, KAT-Coder-V2.5-Dev-OptiQ-4bit reaches up to 100.5 tok/s and averages 87.5 tok/s, with the fastest results on Apple M4 Max.
- How much memory does KAT-Coder-V2.5-Dev-OptiQ-4bit need?
- The leanest observed configuration used about 21.2 GB of memory (quantizations tested: 4bit).
- Which tools have been used to run KAT-Coder-V2.5-Dev-OptiQ-4bit?
- Benchmarks were submitted using oMLX. Results are community-contributed and updated as new runs arrive.