Qwen3.8-Flash-Next-REAP-384-oQ4e-BF16-MTP-PLE

Qwen3.8-Flash-Next REAP-384 oQ4e + Native MTP + PLE

This Qwen3.8-Flash-Next derivative has a 384-expert target MoE bank and a native MTP predictor with 512 experts. It is packaged for qwen4_exp-compatible runtimes.[1]

Release facts — inspected from this repository

The repository identifier is retained as the release name.[1] The authoritative metadata is config.json: it describes a mixed 4-bit package with named 5-, 6-, and 8-bit overrides, so this release is not a uniformly BF16 checkpoint.[1]

Reproducible oMLX benchmark envelope

The benchmark tables below use published oMLX runs for this exact 384-expert oQ4e artifact: HumanEval,[3][4] LiveCodeBench,[5][6] and performance.[7][8][9] The accuracy runs used an M4 Max (40-core GPU), 128 GB unified memory, macOS 26.5.1, oMLX v0.7.0.dev4, 4-bit model loading, batch size 4, seed 42 stratified sampling, TurboQuant KV at 4-bit, and Lightning MTP.[3][4] The LiveCodeBench records report the same host, software, sampling, and acceleration envelope.[5][6] Each accuracy result is a 100-question sample, not a full-dataset result.[3][5]

Accuracy charts — same artifact, same host, same sample size

HumanEval

The HumanEval sample is 100 of 164 problems. The recorded run times were 3.5 minutes for Thinking No and 8.5 minutes for Thinking Yes.[3][4] A four-question difference in a 100-question sample is descriptive, not a controlled ranking between settings.[unverified]

LiveCodeBench (not relabeled as LiveBench)

These records are explicitly labeled LiveCodeBench and use a 100-question sample of the 1,054-question benchmark. They are not relabeled as “LiveBench,” which is a different benchmark.[5][6] The recorded run times were 35.9 minutes for Thinking No and 2.5 hours for Thinking Yes.[5][6]

HumanEval and LiveCodeBench are intentionally shown separately because they measure different datasets.[3][4][5][6]

Throughput charts — exact artifact on M4 Max (40c) / 128 GB

PP is prompt-processing throughput; TG is token-generation throughput. The records below are one oMLX session using a Code (Python) context, 4-bit loading, TurboQuant KV at 4-bit, and Lightning MTP.[7][8][9] They are workload-specific deployment snapshots, not general speed guarantees.[7][8][9]

Single-stream throughput by context

Generation throughput by concurrent stream batch

The batch table comes from the same published performance session as the context table.[7][8][9]

Runtime and compatibility

This release needs a runtime that supports the qwen4_exp architecture and its asymmetric target/MTP expert counts recorded in config.json.[1] A catalog listing alone is not proof of a successful load.[unverified] Memory use and speed depend on runtime version, context length, KV-cache policy, storage path, concurrency, MTP configuration, and hardware.[7][8][9]

Lineage and license

The base model is Qwen/Qwen3.8-Flash-Next, whose official card documents the base architecture and licensing.[2] This derivative is distributed under the Qwen Community License 1.0; the verbatim upstream license is included as LICENSE in this repository.[1][2] Review the license before commercial deployment.[2]

Limitations

• All accuracy values above are 100-question stratified samples; evaluate the full workload before relying on them.[3][4][5][6]
• Do not treat HumanEval and LiveCodeBench as a combined quality score.[3][4][5][6]
• Vision or multimodal quality has not been independently characterized for this release.[unverified]
• The visual bars are presentation aids; the numeric columns are authoritative.[unverified]

Sources

[1] https://huggingface.co/mensaprodigy/Qwen3.8-Flash-Next-REAP-384-oQ4e-BF16-MTP-PLE [2] https://huggingface.co/Qwen/Qwen3.8-Flash-Next [3] https://omlx.ai/benchmarks/intelligence/nnw0eqdx [4] https://omlx.ai/benchmarks/intelligence/026c9etj [5] https://omlx.ai/benchmarks/intelligence/h5hbdt8q [6] https://omlx.ai/benchmarks/intelligence/moonsh71 [7] https://omlx.ai/benchmarks/performance/ybqngdig [8] https://omlx.ai/benchmarks/performance/ab61ivk2 [9] https://omlx.ai/benchmarks/performance/5sqxj1ug