Qwen3.8-27B OBLITERATED Mythos-Class-Agentic — MLX 6-bit
An unofficial, complete MLX conversion of medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic. The upstream model derives from Qwen/Qwen3.8-27B. No additional training was performed.
32 GiB Mac users: start with the 4-bit version. This 6-bit version passed text/tool smoke tests in oMLX with the text-only runtime. The default multimodal oMLX configuration hit its memory guard during the first text request on our M1 Max with 32 GiB. LM Studio's resource guard rejected loading. Successful image/video inference has not been validated.
The repository contains the complete multimodal weights and configuration. The text-only setting used in testing was a local oMLX runtime choice; no vision weights were removed from these downloadable files.
Conversion
| Item | Value |
|---|---|
| Source revision | 528121d7b0b85885a658dfe67e3643b8a4f9337e |
| Quantization | MLX affine, 6-bit, group size 64 |
| Weights | Approximately 22.78 GB |
| Architecture | qwen3_5, multimodal configuration retained |
| Floating-point modules | Retained according to the native converter's multimodal exclusion rules; not every tensor is 6-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. It took 86 seconds on this machine, excluding download and tests. This single observation is not a performance guarantee.
Compared with 4-bit storage, 6-bit represents weights with finer quantization but uses approximately 6.7 GB more disk space. A measurable improvement on downstream tasks has not been established here. The upstream agentic claims were not independently reproduced.
Weights, configuration, processor serialization and the shard index were regenerated. The upstream chat template and Apache license file are preserved. Source identity and software versions are recorded in conversion-provenance.json; file hashes are in checksums.sha256.
Tested local configuration
The successful text tests used oMLX 0.6.4, model-type override llm and a maximum context setting of 4,096 tokens. The actual prompts were short. Memory guards remained enabled and unchanged. The override avoids loading the vision component for these text requests; it does not modify the checkpoint.
This model is not validated for general multimodal use on a 32 GiB Mac. Hardware with more memory is the appropriate target for further full-model evaluation, but no higher-memory machine was tested here and no guaranteed minimum RAM requirement is claimed.
LM Studio 0.4.24+1 discovered the files but rejected the 4,096-context load based on its resource guard. Its estimated requirement of 29.73 GiB had LOW confidence; it is not a measured runtime footprint. This release does not claim a successful LM Studio inference test for 6-bit.
Validation
One run of three synthetic tasks on 2026-09-13:
| Check | oMLX text-only runtime |
|---|---|
| Invoice amount and due date | Correct |
| Ambiguous amount and missing date | Left unknown |
| Native demo-tool name and arguments | Correct |
| Answer after the synthetic tool result | Correct |
These are smoke tests, not a benchmark or an accuracy-retention study. The same task family also succeeded with the 4-bit conversion and stock Qwen, so no general quality advantage is established. Extraction answers were JSON, but field types were not constrained by a JSON Schema. Use an explicit schema and validate outputs in automated workflows.
The tested tool was read-only and synthetic. To build an actual agent, a host application must provide tools, execute selected calls, return results and schedule jobs. The model alone does not connect to Gmail.
validation-results.json includes the synthetic prompts, final answers, tool calls and limitations. Image/video behavior, long-context performance, real email workflows and broad agent reliability remain untested.
License and attribution
Distributed under Apache License 2.0, with the upstream copyright and license notices retained. 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 is an independent quantized conversion published by uzzero, not an official Qwen or medismera release.
See the upstream model card for its behavioral characteristics, limitations and claims. This conversion changes representation, not training or intent.
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.