Robot-Haus/Qwen3.8-Flash-Next-MLX-oQ3-MTP

M1 Ultra Speed Benchmark and Anecdotal Intelligence Review Figures Below

Qwen3.8 Flash Next — MLX oQ3 with native MTP

A sensitivity-guided, mixed-precision MLX conversion of Qwen/Qwen3.8-Flash-Next, rebuilt directly from the official BF16 checkpoint with the model's matching native MTP block preserved.

Original model · Qwen overview · MLX-VLM · Qwen Community License 1.0

About this conversion

oQ3 uses a 3-bit affine base and spends additional precision on sensitive modules. Layer sensitivity was measured with a validated quantized calibration proxy, while every released weight was quantized from the official BF16 checkpoint. The result is a compact model with 746 higher-precision module overrides rather than a uniform 3-bit layout.

Item	Value
Repository	Vontra/Qwen3.8-Flash-Next-MLX-oQ3-MTP
Base model	Qwen/Qwen3.8-Flash-Next
Source weights	Official BF16 checkpoint
Format	MLX safetensors
Quantisation	oQ3 mixed precision, 3-bit affine base
Base group size	32
Protected modules	314 at 4-bit, 82 at 5-bit, 127 at 6-bit, and 223 at 8-bit
Native MTP	Included, one Qwen4Exp draft block
Indexed tensors	3,747 total, including 76 MTP entries
Weight shards	19
Weight size	92.505 GB / 86.152 GiB
Configured context	262,144 tokens
Architecture	qwen4_exp vision-language sparse MoE

The upstream tokenizer, current chat template, vision processor, generation configuration, licence, and native MTP configuration are retained.

Use a runtime with explicit qwen4_exp and native-MTP support. A runtime that does not construct the Qwen4Exp MTP module may reject the 76 MTP tensor entries during strict loading.

Download and use

python -m pip install --upgrade huggingface_hub


hf download Vontra/Qwen3.8-Flash-Next-MLX-oQ3-MTP \
  --local-dir ./Qwen3.8-Flash-Next-MLX-oQ3-MTP

In a compatible oMLX build, add the downloaded directory to the model directories, refresh the registry, select the model, and enable native MTP. MTP can be disabled for baseline comparisons and troubleshooting.

Apple M3 Studio performance

Benchmark environment: oMLX 0.6.3rc3 (build 2475), MLX-VLM 0.6.3, and MLX 0.32.0 on an Apple M3 Studio. Measurements use greedy decoding, seed 6330, identical prompts, a separate warm-up, and three measured 512-token runs per mode.

Runtime mode	Runs	Output per run	Median generation speed	Drafted	Accepted	Acceptance
Native MTP disabled	3	512 tokens	26.5352 tokens/s	Not applicable	Not applicable	Not applicable
Native MTP enabled	3	512 tokens	29.0822 tokens/s	847	583	68.83%

Native MTP improved median decode throughput by 9.60% in this test. The MTP-off and MTP-on runs produced exactly matching output hashes. All six sustained runs reached 512 generated tokens; exact instruction, factual, arithmetic, and coherent long-generation gates passed in both modes.

Results vary with prompt length, context growth, cache state, runtime version, memory pressure, and thermal conditions. The first request after loading includes model and kernel warm-up and is excluded from the steady-state result.

Apple M1 Ultra (128 GB) performance

Benchmark environment: oMLX 0.6.4, engine auto, context Code (Python), on an Apple M1 Ultra 128 GB. As of this writing, this is the preferred oMLX pick for Qwen3.8 Flash Next on this system.

Single request

Test	TTFT (ms)	TPOT (ms)	Prefill TPS	Decode TPS	E2E (s)	Throughput	Peak Mem
pp1024/tg128	8157.6	49.00	125.5	20.6	14.394	80.0	64.60 GB
pp4096/tg128	22503.5	50.09	182.0	20.1	28.878	146.3	67.25 GB
pp8192/tg128	35326.7	58.74	231.9	18.0	36.342	225.9	67.39 GB
pp16384/tg128	62766.1	160.83	261.0	12.4	63.079	259.8	67.68 GB
pp32768/tg128	118533.7	69.03	276.4	15.5	119.704	273.9	68.26 GB
pp65536/tg128	231924.8	48.68	282.6	20.7	238.127	275.8	69.42 GB

Continuous batching — pp1024/tg128

Batch	Decode TPS	Speedup	Prefill TPS	Prefill TPS/req	TTFT (ms)	E2E (s)
1×	20.6	1.00×	125.5	125.5	8157.6	14.394
2×	30.7	1.49×	187.9	94.0	7979.6	19.241
4×	52.8	2.56×	155.8	39.0	15141.8	35.992
8×	102.0	4.95×	144.2	18.0	30174.5	65.264

Peak memory stays below 70 GB across all tested context lengths, leaving comfortable headroom on a 128 GB system.

Intelligence benchmarks (sample 30, no-think mode)

Benchmark	Accuracy	Correct	Total	Time (s)
MMLU	80.0%	24	30	125.4
TruthfulQA	100.0%	30	30	37.7
HumanEval	90.0%	27	30	272.0
MBPP	90.0%	27	30	144.5

Architecture

Qwen3.8 Flash Next combines Gated DeltaNet, Qwen Sparse Attention, sparse mixture-of-experts layers, widened gated residual streams, hashed bigram and trigram embeddings, and a native next-token-prediction block for speculative decoding.

Architecture detail	Upstream value
Language-model parameters	125B total / 6B active
N-gram embedding	51B parameters, 20,000,000 entries
Native MTP	4B parameters, one draft layer
Hidden size	2,560
Layers	48
Routed / active experts	512 / 10, plus 1 shared expert
Native context	262,144 tokens, extensible upstream to 1,000,000

For upstream evaluations, intended use, limitations, safety guidance, and the full architecture discussion, see the original model card.

Conversion and validation

• The converter read the official BF16 checkpoint directly.
• A 3-bit affine base at group size 32 was combined with 746 sensitivity-guided 4/5/6/8-bit overrides.
• Structural validation passed for all 3,747 indexed tensors, all 19 weight shards, and all 76 native-MTP tensor entries.
• Three deterministic 512-token runs passed in each MTP mode with exact output parity.
• Native-MTP telemetry, exact-answer gates, and coherent long-generation checks passed.

This is a community conversion, not an official Qwen release.

Limitations

• Quantisation can reduce quality relative to BF16; 3-bit models should be evaluated on the intended workload.
• The native 262,144-token context does not guarantee that every Apple-silicon system can run that length within available unified memory.
• Native MTP needs a compatible runtime and may not improve every prompt or context length.
• This is an MLX release for Apple silicon, not a GGUF, CUDA, TensorRT-LLM, or vLLM checkpoint.
• Upstream model limitations and safety considerations still apply.

Licence and attribution

The upstream model is released under the Qwen Community License 1.0. The required licence text is included in this repository and should be reviewed before use or redistribution.

Model design, training, evaluations, and upstream documentation belong to Qwen and the original contributors. The MLX conversion, native-MTP preservation, Apple-silicon validation, and packaging are provided by Vontra.