base_model:
- jedisct1/Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP
base_model_relation: quantized
library_name: mlx
pipeline_tag: image-text-to-text
license: other
license_name: qwen-community-1.0
license_link: LICENSE
tags: - omlx
- mlx
- qwen4
- mtp
- vision
- tool-calling
- plefp8
Qwen3.8 Flash Next Uncensored oQ4e for oMLX with FP8 PLE n-gram table
Those are the quantized weights from jedisct1/Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP with FP8 PLE n-gram table transplanted into it.
The effect of the less-lossy PLE table data is unknown, this is an experiment. Feedback is welcome.
This model passes the intelligence benchmarks with effectively the same results as the original quantized model and is subjectively coherent.
Mode Sampled Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8 Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP
--------------------------------------------------------------------------------------------------------------------------------------------
MMLU Sample 300/14042 87.7% 87.0%
MMLU_PRO Sample 300/12032 68.0% 68.7%
HUMANEVAL Full 164 92.1% 93.3%
--- Detail ---
Model: Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP-plefp8
Benchmark Accuracy Correct Total Time(s) Think
--------------------------------------------------------------
MMLU 87.7% 263 300 417.7 No
MMLU_PRO 68.0% 204 300 425.1 No
Hit token limit: 1/300 (0 of them scored correct) · Accuracy on finished answers: 68.2% (n = 299)
HUMANEVAL 92.1% 151 164 297.7 No
Model: Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP
Benchmark Accuracy Correct Total Time(s) Think
--------------------------------------------------------------
MMLU 87.0% 261 300 442.2 No
Hit token limit: 1/300 (0 of them scored correct) · Accuracy on finished answers: 87.3% (n = 299)
MMLU_PRO 68.7% 206 300 419 No
Hit token limit: 2/300 (0 of them scored correct) · Accuracy on finished answers: 69.1% (n = 298)
HUMANEVAL 93.3% 153 164 303.3 No
Performance and memory use appear to be identical to the original.
License
The pinned source repository contains the Qwen Community License 1.0, which is
preserved in LICENSE and includes additional commercial terms.
The source model card labels the repository Apache-2.0, but that label does not
match the license file in the pinned source revision.
Review the included license before using or distributing this artifact.