Model benchmarks

ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit local LLM performance

As of September 2026, ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit runs at up to 72.0 tok/s for local inference (best of 3 community benchmark runs across 1 GPU).

vLLM
ShareRedditX

Model size

Unknown

Peak speed

72.0 tok/s

Average speed

71.5 tok/s

Avg PP

1910.9 tok/s

Min memory

n/a

Max context

262.144 tokens

Avg output / run

15.625 tokens

Avg runtime / run

3m 23s

Avg quality

70.6

Benchmark runs

3

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 3 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.

TaskP5 (low)AvgP95 (high)
Overall68.070.674.8
Agent Workflow69.972.174.3
Code Generation38.452.566.9
Role Play & Narrative73.479.086.3
Research & Analysis74.578.983.4

Performance by hardware and tool

Every hardware/tool/quantization combination ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit has been benchmarked on, ranked by peak token generation speed. Last updated September 2026.

HardwareToolQuantBest tok/sAvg tok/sMemoryContextQualityRuns
Apple M4 MaxvLLM72.0 tok/s71.5 tok/sn/a262.144 tokens70.63

Benchmark runs

All 3 ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Frequently asked questions

Is ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit good for coding?
In our benchmarks, ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit scores 52.5/100 for coding. It runs at about 71.5 tok/s, so if you want more speed, unsloth/Qwen3.8-27B-GGUF:IQ3_S is faster (~104.3 tok/s) and still scores well for coding (83.0/100).
Is ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit good for agentic (tool-using) tasks?
In our benchmarks, ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit scores 72.1/100 for agentic workflows. It runs at about 71.5 tok/s, so if you want more speed, Ornith-1.5-35B-A3B-BigBang-oQ8e-mtp is faster (~102.7 tok/s) and still scores well for agentic workflows (91.1/100).
How fast is ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit for local inference?
Across 3 community benchmark runs, ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit reaches up to 72.0 tok/s and averages 71.5 tok/s, with the fastest results on Apple M4 Max.
How much memory does ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit need?
The leanest observed configuration used about n/a of memory.
Which tools have been used to run ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit?
Benchmarks were submitted using vLLM. Results are community-contributed and updated as new runs arrive.