Qwen3.8-Flash-Next-oQ8e-mtp

This model was quantized using oQ in oMLX v0.6.4. oQe (imatrix). MTP kept.

Quantization details

• Model type: qwen4_exp
• Bits: 8
• Group size: 64
• Format: MLX safetensors
• oMLX: Jundot v0.6.4 (source)
• Parameters: 125B total, 6B activated, 51B n-gram embedding, 4B MTP
• MTP: preserved (76 mtp.* tensors)
• Source: Qwen/Qwen3.8-Flash-Next, converted to MLX, then oQe with preserve_mtp

Benchmark

Apple M3 Ultra, 256 GB. oMLX 0.6.4. Thinking off, temperature 0, Lightning MTP depth 3. oQ4e is Jundot/Qwen3.8-Flash-Next-oQ4e-mtp on the same machine.

Write speed is how fast the reply is generated (generation_tokens_per_second). Read speed is how fast the prompt is ingested (prompt_tokens_per_second).

Coding task: write a Python Fibonacci function, no comments. All three packs stopped on their own.

	oQ4e	oQ6e	oQ8e
Time to completion	1.74 s	2.07 s	2.57 s
Time to first token	0.51 s	0.64 s	0.67 s
Tokens written	79	80	80
Write speed (tok/s)	64.4	56.1	42.1
Read speed (tok/s)	51.2	40.3	38.9

Longer prompts (256-token generation cap):

Write speed (tok/s)

Context	oQ4e	oQ6e	oQ8e
Short code	64.4	56.1	42.1
4K	56.5	44.9	41.9
8K	56.3	44.8	43.7
16K	52.9	48.3	47.5

Read speed (tok/s)

Context	oQ4e	oQ6e	oQ8e
Short code	51.2	40.3	38.9
4K	732	597	581
8K	772	728	716
16K	729	785	763

Raw timings: bench.json (this pack) and bench-compare.json (all three).