Downloads · 30 days
73
31% of all-time downloads
mensaprodigy/Qwen3.8-Flash-Next-REAP-384-oQ5e-BF16-MTP-PLE
Qwen3.8-Flash-Next-REAP-384-oQ5e-BF16-MTP-PLE is a machine learning model from mensaprodigy. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
This Qwen3.8-Flash-Next derivative has a 384-expert target MoE bank and retains the native one-layer, 512-expert MTP predictor in the published configuration.[1] It is packaged for qwen4exp-compatible runtimes.[1]
Downloads · 30 days
73
31% of all-time downloads
All-time downloads
239
Public
Parameters
150B
175 GB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors175 GB · 100%
How the weights are stored.
U3295.5B · 64%
From the Hugging Face model README
This Qwen3.8-Flash-Next derivative has a 384-expert target MoE bank and retains the native one-layer, 512-expert MTP predictor in the published configuration.[1] It is packaged for qwen4_exp-compatible runtimes.[1]
| Field | Published value |
|---|---|
| Architecture | Qwen4ExpForConditionalGeneration / qwen4_exp |
| Target transformer layers | 48 |
| Target experts per layer | 384 |
| Target routing width | top-10 |
| Native MTP | 1 layer, 512 experts, top-10 routing |
| Context configuration | 262,144 positions |
| Safetensor shards | 32 |
| Indexed tensors | 3,448 total; 32 mtp.* tensors |
| Primary quantization configuration | 5-bit, group size 64, with 303 named overrides |
The repository name is retained as the release identifier.[1] The authoritative artifact metadata is config.json: it contains a mixed-precision quantization configuration, so this release must not be described as a uniformly BF16 checkpoint.[1]
The charts below use only published oMLX runs for this exact 384-expert oQ5e artifact.[9][10][12][13] Accuracy runs used an M4 Max (40-core GPU), 128 GB unified memory, macOS 26.5.1, oMLX v0.7.0.dev4, 5-bit model loading, batch size 4, seed 42 stratified sampling, TurboQuant KV at 4-bit, and Lightning MTP.[9][10][12][13] Each reported accuracy result is a 100-question sample, not a full-dataset result.[9][10][12][13]
Comparison boundary: this card deliberately excludes Sh0wie REAP-288 results.[1][2] That model has a different active-expert count and build, so placing it on the same chart would not be apples-to-apples.[1][2] A matching Mensa REAP-288 card can carry a clearly labeled 288-to-288 comparison after its own release and matching measurements are published.[unverified]
This matrix preserves useful community context while keeping unlike builds, harnesses, and hosts out of a single winner-ranking chart.[1][7][27] The 384-expert Sh0wie build is the closest community artifact by active-expert count; Jundot's oQ4e MTP package is a 512-expert, 4-bit community baseline.[7][27]
| Published build | Active experts | Package | HumanEval evidence | LiveCodeBench evidence | Published throughput snapshot | Comparison scope |
|---|---|---|---|---|---|---|
| Mensa REAP-384 oQ5e | 384 | 5-bit mixed configuration | 97/100 (Thinking No); 96/100 (Thinking Yes) | 61/100 (No); 62/100 (Yes) | M4 Max 40c, 4k: 679.1 PP / 53.9 TG tok/s | Exact release; its own oMLX runs |
| Sh0wie REAP-384 MLX | 384 | 4-bit MLX | 92.1% on 164 HumanEval problems | No value used here | Not included here | Same active-expert count; different package and publisher harness |
| Jundot oQ4e MTP | 512 | 4-bit mixed configuration | 152/164 = 92.7%, full HumanEval | 55/100 | M5 Max 40c, 4k: 1,286 PP / 62.5 TG tok/s | Different expert count, quantization, runtime, batch, and performance host |
The Mensa row is based on its exact-release oMLX records and artifact metadata.[1][9][10] The Sh0wie row is publisher-reported for its 384-expert 4-bit MLX build, using a 164-problem HumanEval harness.[7] The Jundot row is tied to its oMLX HumanEval, LiveCodeBench, and M5 Max performance records.[31][32][30]
| Build / oMLX run | Hardware | Package | Batch | Sample | Score | Interpretation |
|---|---|---|---|---|---|---|
| Mensa REAP-384 oQ5e, Thinking Yes | M4 Max (40c), 128 GB | 5-bit | 4 | 100/1,054 | 62.0% | Exact release result |
| Jundot oQ4e MTP, Thinking Yes | M4 Max (40c), 128 GB | 4-bit | 1 | 100/1,054 | 55.0% | Same suite and sample size; not a controlled head-to-head |
Both LiveCodeBench rows use the same benchmark family, 100-question sample size, M4 Max (40c) / 128 GB host class, seed 42 stratification, TurboQuant KV, and Lightning MTP.[13][32] They still differ in active-expert count, quantization, oMLX version, macOS version, and batch size, so this table intentionally provides no score delta, winner label, or common bar scale.[13][32]
| 4k published run | Chip / RAM | Package | PP tok/s | TG tok/s | Peak memory | Scope |
|---|---|---|---|---|---|---|
| Mensa REAP-384 oQ5e | M4 Max (40c) / 128 GB | 5-bit | 679.1 | 53.9 | 72.8 GB | Exact-release session |
| Jundot oQ4e MTP | M5 Max (40c) / 128 GB | 4-bit | 1,286 | 62.5 | 72.3 GB | Different 512-expert model and newer chip generation |
The two 4k throughput records are useful deployment snapshots, not a speed ranking; the M5-vs-M4 host difference alone prevents an apples-to-apples conclusion.[18][30]
| Run label reported by oMLX | Correct / sampled | Accuracy | Visual |
|---|---|---|---|
| Thinking No | 97 / 100 | 97.0% | ███████████████████░ |
| Thinking Yes | 96 / 100 | 96.0% | ███████████████████░ |
The HumanEval sample is 100 of 164 problems.[9][10] The run labels, timing, and score records are published separately: 4.5 minutes for the 97.0% result and 7.5 minutes for the 96.0% result.[9][10] A one-question difference in a 100-question sample is descriptive—not a statistically established ranking between settings.[unverified]
| Run label reported by oMLX | Correct / sampled | Accuracy | Visual |
|---|---|---|---|
| Thinking No | 61 / 100 | 61.0% | ████████████░░░░░░░░ |
| Thinking Yes | 62 / 100 | 62.0% | ████████████░░░░░░░░ |
These oMLX records are explicitly labeled LiveCodeBench and use a 100-question sample of the 1,054-question benchmark.[12][13] They are shown as LiveCodeBench, not as “LiveBench,” to avoid conflating distinct benchmarks.[12][13] The recorded run times were 38.4 minutes (Thinking No) and 1.4 hours (Thinking Yes).[12][13]
HumanEval and LiveCodeBench are intentionally charted separately: they are different datasets and should not be combined into a single quality score.[9][10][12][13]
The first table is one comparable oMLX session: same model, host, 5-bit load, and Lightning MTP.[18] PP is prompt-processing throughput; TG is token-generation throughput.[18] Throughput is workload- and configuration-specific, not a general speed guarantee.[18]
| Context | PP tok/s | TG tok/s | Peak memory | PP visual | TG visual |
|---|---|---|---|---|---|
| 1k | 610.0 | 54.3 | 70.0 GB | █████████████████ | ███████████████████ |
| 4k | 679.1 | 53.9 | 72.8 GB | ███████████████████ | ███████████████████ |
| 8k | 725.7 | 56.5 | 72.9 GB | ████████████████████ | ████████████████████ |
| Batch | TG tok/s | Relative to 1× | Visual |
|---|---|---|---|
| 1× | 54.3 | 1.00× | █████ |
| 2× | 130.9 | 2.41× | ████████████ |
| 4× | 187.9 | 3.46× | ██████████████████ |
A separately published 16k-context run on the same M4 Max (40c) / 128 GB, 5-bit configuration reports 709.2 PP tok/s, 56.5 TG tok/s, and 73.2 GB peak memory.[20] It is listed separately because it was not part of the three-context session above.[18][20]
This release needs a runtime that supports the qwen4_exp architecture and the asymmetric target/MTP expert counts recorded in config.json.[1] A model appearing in a server catalog 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.[18][20]
The base model is Qwen/Qwen3.8-Flash-Next, whose official card publishes its own base-model evaluation tables and license.[8] This derivative is distributed under the Qwen Community License 1.0; the verbatim upstream license is included as LICENSE in this repository.[1][8] Review the license before any commercial deployment.[8]
[1] https://huggingface.co/mensaprodigy/Qwen3.8-Flash-Next-REAP-384-oQ5e-BF16-MTP-PLE [2] https://huggingface.co/sh0wie/Qwen3.8-Flash-Next-REAP-288-bf16 [7] https://huggingface.co/sh0wie/Qwen3.8-Flash-Next-REAP-384-MLX-4bit [8] https://huggingface.co/Qwen/Qwen3.8-Flash-Next [9] https://omlx.ai/benchmarks/intelligence/q2adqprz [10] https://omlx.ai/benchmarks/intelligence/rvjjhv65 [12] https://omlx.ai/benchmarks/intelligence/b6vnuvc6 [13] https://omlx.ai/benchmarks/intelligence/2tc37530 [18] https://omlx.ai/benchmarks/performance/dedea6ut [20] https://omlx.ai/benchmarks/performance/pfeuc75o [27] https://huggingface.co/Jundot/Qwen3.8-Flash-Next-oQ4e-mtp [30] https://omlx.ai/benchmarks/performance/00l1yizf [31] https://omlx.ai/benchmarks/intelligence/kmdnbouq [32] https://omlx.ai/benchmarks/intelligence/r6k92mrp