Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE

Qwen3.8-Flash-Next REAP-288 oQ4e + Native MTP + PLE

A Qwen3.8-Flash-Next derivative with 288 target MoE experts per layer and the native one-layer MTP predictor retained in the published configuration.[1] It is packaged for qwen4_exp-compatible runtimes.[1]

Release facts — inspected from this repository

The repository name is retained as the release identifier.[1] The authoritative artifact metadata is config.json: this package has a mixed-precision 4-bit configuration and should not be described as a uniformly BF16 checkpoint.[1]

Published oMLX tests — matching source artifact

The oMLX records use the local source identifier Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-PLE-mtp; its published benchmark configuration is the matching 288-expert, 4-bit MTP source whose config.json and safetensor index match this Hub release.[1][12][13] All accuracy runs below used M4 Max (40c), 128 GB unified memory, macOS 26.5.1, oMLX v0.7.0.dev4, 4-bit loading, batch size 4, seed 42 stratification, TurboQuant KV at 4-bit, and Lightning MTP.[12][13]

Accuracy

The HumanEval and LiveCodeBench results are separate 100-question samples; they are not combined into a single quality score.[12][13] The benchmark record calls the code task LiveCodeBench, not LiveBench.[12]

Throughput — exact source session on M4 Max (40c) / 128 GB

PP is prompt-processing throughput and TG is token-generation throughput.[14]

These measurements are workload- and configuration-specific; treat them as a reproducible snapshot, not a hardware-independent speed guarantee.[14]

Community context — scoped, not a leaderboard

The table below is included to make the community landscape easier to inspect, not to claim a direct ranking.[1][2][3] Sh0wie REAP-288 is the closest public comparison by active-expert count; Jundot oQ4e MTP is an unpruned 512-expert, 4-bit baseline.[2][3]

The Mensa row is from the matching-source oMLX HumanEval, LiveCodeBench, and M4 Max performance records.[12][13][14] Sh0wie's HumanEval figure is from its published 288-expert card; its cited oMLX performance run uses M3 Ultra / 96 GB, 4-bit loading, 1,024-token context, and MTP disabled.[2][6] Jundot's oMLX records use its 512-expert 4-bit package; the HumanEval run is full-dataset on M4 Max / 128 GB, while the listed 4k throughput snapshot is on M5 Max / 128 GB.[10][11]

What is comparable today

• Architecture match: Mensa and Sh0wie both retain 288 target experts with top-10 routing.[1][2]
• Accuracy: Mensa's published HumanEval result is a 100-question sample, while Sh0wie and Jundot publish 164-problem results under different packages and harnesses; show all three as context, not as a controlled ranking.[13][2][10]
• Throughput: the published Sh0wie and Jundot runs use different chips, contexts, runtime settings, and MTP state; do not use them as a speed prediction for this artifact.[6][11]

Runtime and compatibility

This release needs a runtime that supports qwen4_exp, 288 target experts, and the asymmetric 512-expert native MTP layer recorded in config.json.[1] Memory use and speed depend on the runtime version, context length, KV-cache policy, storage path, concurrency, and MTP configuration.[unverified]

Lineage and license

The base model is Qwen/Qwen3.8-Flash-Next, whose official card publishes its evaluation tables and license.[4] This derivative is distributed under the Qwen Community License 1.0; the verbatim upstream license is included as LICENSE in this repository.[1][4]

Limitations

• Evaluate the exact release on your own workload before relying on any benchmark snapshot.[12][13][14]
• Do not treat the Sh0wie or Jundot results as performance or quality guarantees for this build.[2][3][10]
• Vision or multimodal quality has not been independently characterized for this release.[unverified]
• The model name is not authoritative evidence of runtime precision; use config.json when selecting a compatible runtime.[1]

Sources

[1] https://huggingface.co/mensaprodigy/Qwen3.8-Flash-Next-REAP-288-oQ4e-BF16-MTP-PLE [2] https://huggingface.co/sh0wie/Qwen3.8-Flash-Next-REAP-288-MLX-4bit [3] https://huggingface.co/Jundot/Qwen3.8-Flash-Next-oQ4e-mtp [4] https://huggingface.co/Qwen/Qwen3.8-Flash-Next [6] https://omlx.ai/benchmarks/performance/2xs5i144 [10] https://omlx.ai/benchmarks/intelligence/r6k92mrp [11] https://omlx.ai/benchmarks/performance/00l1yizf [12] https://omlx.ai/benchmarks/intelligence/b7jdhtsf [13] https://omlx.ai/benchmarks/intelligence/jwmf2z73 [14] https://omlx.ai/benchmarks/performance/z4hn1h6n