Qwen3.8-27B OBLITERATED Mythos-Class-Agentic — MLX 4-bit
An unofficial MLX conversion of medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic, made for local inference on Apple Silicon. No additional training was performed. The upstream model derives from Qwen/Qwen3.8-27B.
This repository changes the storage format and quantization. It does not establish improved agent performance, reproduce the upstream benchmarks, or imply endorsement by the upstream authors.
Conversion
| Item | Value |
|---|---|
| Source revision | 528121d7b0b85885a658dfe67e3643b8a4f9337e |
| Quantization | MLX affine, 4-bit, group size 64 |
| Weights | Approximately 16.05 GB, 16 safetensors shards |
| Architecture | qwen3_5, multimodal configuration retained |
| Floating-point modules | Retained according to the converter's native multimodal exclusion rules; not every tensor is 4-bit |
| Software | MLX 0.32.0, mlx-vlm 0.6.3, mlx-lm 0.31.3, Transformers 5.12.1 |
| Conversion hardware | Apple M1 Max, 32 GiB unified memory |
The conversion used lazy loading, CPU execution, a 256 MiB MLX cache limit, and approximately 1 GiB output shards to keep conversion memory modest. The measured conversion time was 79 seconds on this machine; download and validation time are excluded. This is a single observation, not a performance guarantee.
Model weights, configuration, processor serialization and the shard index were regenerated. The upstream chat template and Apache license file are preserved. See conversion-provenance.json and checksums.sha256 for provenance and file identity.
Use in oMLX or LM Studio
Download this repository as an MLX model. It was successfully loaded in oMLX 0.6.4 and LM Studio 0.4.24+1. In LM Studio the load test used an 8,192-token context setting; the test prompts were short. Adjust the context setting to available memory, and avoid loading another large model at the same time on a 32 GiB machine.
Both applications can discover the same local model directory when configured to use it. To use tools, the client must supply tool definitions, execute the chosen tools, and return tool results. This model alone does not connect to Gmail or schedule daily jobs.
The upstream chat template is included. Application support for reasoning controls varies: an enable_thinking=false request suppressed the reasoning field in the tested oMLX calls, but LM Studio still returned reasoning content. Do not assume that this API extension behaves identically across hosts.
Validation
Tested on 2026-09-13 with synthetic data only:
| Check | oMLX | LM Studio |
|---|---|---|
| Load this converted model | Passed | Passed |
| Extract invoice identifier, amount and due date | Correct values | Correct values |
| Leave ambiguous amount and missing date unknown | Passed | Passed |
| Request the correct read-only demo tool and arguments | Passed | Passed |
| Answer using the returned demo tool result | Passed | Passed |
The stock Qwen3.8-27B MLX 4-bit model was also tested on the same three tasks in oMLX with identical request bodies and sampling settings. Both models solved the tested tasks. These few examples do not establish a quality advantage for Mythos. This is not a formal benchmark or a measurement of quality retained after quantization.
Extraction answers were valid JSON, but the prompts did not impose a typed JSON Schema. Representations of the uncertainty field varied between null, boolean and explanatory text; LM Studio returned the invoice amount as a string. Use an explicit schema and validate output types for automated workflows.
The full synthetic prompts, final answers, tool calls and timings are in validation-results.json. Generated reasoning text is omitted. Timing was not controlled sufficiently for speed rankings.
Vision/video behavior, long-context performance, real email workflows and broad agent reliability have not been evaluated. The multimodal configuration and associated weights are retained, but successful text tests do not validate vision capabilities.
License and attribution
Distributed under Apache License 2.0, retaining the upstream license and copyright notices. Credit for the base model belongs to Qwen/Alibaba Cloud and credit for the upstream derivative belongs to its authors, including the referenced OBLITERATUS lineage. This repository is an independent quantized conversion published by uzzero.
The source model's behavioral characteristics and limitations remain relevant. See the upstream model card for its description and claims; those claims were not independently reproduced here.
Reasoning controls in LM Studio
For Reasoning Effort (Low / Medium / Extra High), Enable Thinking and Preserve Thinking, use the published LM Studio Hub entry via lms get uzzero/qwen3.8-27b-mythos-mlx. See the short setup guide for the manual alternative and oMLX settings. Both quantizations share the same definition. Model weights and the upstream chat template are unchanged. The controls require selecting the installed model-definition entry in LM Studio; downloading the checkpoint alone does not activate them.