license: apache-2.0
language:
- en
- ru
library_name: mlx
pipeline_tag: image-text-to-text
base_model: - heretic-org/Qwen3.8-27B-heretic-ara
tags: - mlx
- apple-silicon
- qwen3_5
- qwen3.8
- vision
- mtp
- uncensored
Sovereign v2 — standard MLX, vision and native MTP
Separate 2026-10-12 release of Sovereign for Apple Silicon. The existing ARA 27B text checkpoint is preserved byte-for-byte at affine 4-bit, group size 64. BF16 vision and native MTP are restored from the same pinned ARA source, heretic-org/Qwen3.8-27B-heretic-ara at 2dc9b364104881cbb85e390f00195ba6b9d745e9.
The checkpoint uses the standard qwen3_5 MLX architecture and safetensors layout. There is no custom model implementation or trust_remote_code requirement. Text weights were not merged, retrained or requantized. The uncensored/ARA lineage is preserved; this is not a promise of zero refusals for every prompt.
Components
| Component | Packaging |
|---|---|
| Text | 3 unchanged MLX Q4 shards, 1,847 tensors |
| Vision | vision.safetensors, 333 BF16 tensors; standard Conv3D transpose |
| Native MTP | mtp/, 15 BF16 tensors, official MLX-VLM splitter |
| Code Guard | tools/code-guard/, optional external Python validation/repair |
| Compact Repair | tools/compact-repair/, optional exact-edit workflow |
MTP is a standalone sidecar and is explicitly selected at inference. The root index contains text and vision only. This prevents plain MLX loaders from treating drafter tensors as target weights.
Standard MLX usage
Tested environment: mlx-vlm==0.7.6, mlx-lm==0.32.0, mlx==0.32.3, Python 3.12, M4 Max / 48 GB. Use a separate virtual environment; requirements.lock.txt contains the complete tested dependency set.
Download this separate version, then run from its directory:
hf download proofofvitalik/Sovereign-Uncensored-ARA-27B-MLX-4bit-v2 --revision v2.0.0 --local-dir ./sovereign-v2
cd sovereign-v2
mlx_vlm.generate --model . --prompt 'Explain this image briefly.' --image image.png --max-tokens 256
mlx_vlm.generate --model . --draft-model ./mtp --draft-kind mtp --prompt 'Write a short Python function that adds two numbers.' --temperature 0 --max-tokens 256
To combine vision with MTP, add --image image.png to the second command. The default CLI mode disables thinking; add --enable-thinking when desired. Native MTP acceptance and speed depend on the prompt and target quantization; consult the measured results rather than assuming acceleration.
Splash 1.3.1
Splash 1.3.1 consumes standard MLX checkpoints. It uses its own DFlash2 drafter, independently of the native MTP sidecar. With Splash 1.3.1 installed, load this repository directly:
splash serve --model proofofvitalik/Sovereign-Uncensored-ARA-27B-MLX-4bit-v2 --revision v2.0.0 --port 8019 --max-context 65536 --max-memory 28G
Local API: http://127.0.0.1:8019/v1. Use the model ID returned by /v1/models in API requests and workflow --model arguments. The publisher's local validation launcher uses alias proofofvitalik/sovereign-v2. The command above reserves a 65,536-token context and a 28 GiB memory budget. Declared architectural context is 262,144; see the validation record for the actual lengths exercised.
Splash downloads the matching DFlash2 separately; DFlash2 and the runtime are not duplicated in this model repository. The tested draft revision and archive checksum are recorded in PROVENANCE.json. To enable Splash's optional disk cache, pass both --persistent-cache and an explicit separate --cache-dir.
LM Studio
Install the model in a separate directory under ~/.lmstudio/models/<publisher>/Sovereign-Uncensored-ARA-27B-MLX-4bit-v2/. The installed model appears as Sovereign Uncensored ARA 27B v2, with vision enabled. Text and image inference were tested on LM Studio MLX runtime 1.13.1 with a 65,536-token context and one concurrent request.
The mtp/ folder is a native MTP sidecar for MLX-VLM, not a standalone chat model. LM Studio may list it separately as Mtp; select the main Sovereign v2 entry. Native MTP was validated through MLX-VLM 0.7.6, and is not enabled in the tested LM Studio MLX installation. The generic LM Studio Draft MTP option should not be taken as proof of support for this MLX sidecar; see the upstream MLX MTP tracking issue.
Preserved Python workflows
Code Guard checks Python syntax and isolated module import, with at most one repair attempt. Compact Repair validates source hashes and exact, unique, non-overlapping edits, optionally checks supplied assertions, and writes a new candidate file. Both default to the separate v2 API at port 8019. Their algorithms are unchanged from Sovereign v1.
Use Python.org Python 3.14 on macOS for these wrappers; their existing sandbox profile permits that interpreter. The MLX virtual environment is for inference, not the sandboxed Python check worker. For a different local server, set --endpoint and --model explicitly. Passing these checks does not establish general program correctness.
For the publisher's LM Studio installation:
python3 tools/code-guard/code_guard_cli.py --prompt-file task.txt --output answer.py --endpoint http://127.0.0.1:1234/v1/chat/completions --model proofofvitalik/sovereign-v2
python3 tools/compact-repair/repair_cli.py --source module.py --task-file task.txt --output module.fixed.py --mode compact --tests-file checks.json --endpoint http://127.0.0.1:1234/v1/chat/completions --model proofofvitalik/sovereign-v2
Versioning and rollback
This release has its own Hugging Face repository and v2.0.0 tag. The publisher's local installation uses independent APFS copy-on-write clones with distinct inodes and matching SHA-256 values. Editing or deleting those new files does not modify the previous release.
Stop the v2 server with Ctrl-C and use the previous Sovereign model/runtime to roll back. No default-model switch is needed because this release does not replace the old one. The previous public release remains Sovereign-Uncensored-Splash-Q4.
See PROVENANCE.json, SHA256SUMS, and EVALUATION.md for exact components and completed checks. Historical v1 benchmark numbers are not new v2 measurements.