Downloads · 30 days
77
11% of all-time downloads
kvas-25/Ektome-Qwen3.8-27B-oQ6e-MLX-MTP
Ektome-Qwen3.8-27B-oQ6e-MLX-MTP is a text generation model from kvas-25. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as apache-2.0.
Near-lossless MLX oQ6e quantization of Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored, with the native 15-tensor MTP head and full vision tower preserved.
Downloads · 30 days
77
11% of all-time downloads
All-time downloads
707
Public
Parameters
27.8B
23.7 GB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors23.7 GB · 100%
How the weights are stored.
U3227.3B · 98%
From the Hugging Face model README
Near-lossless MLX oQ6e quantization of Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored, with the native 15-tensor MTP head and full vision tower preserved.
The Ektomē refusal-removal work belongs to Zynerji. This repository contributes the MLX/oMLX quantization, independent tensor/geometry audit, and runtime validation. The untouched upstream base is Qwen/Qwen3.8-27B.
| Property | Value |
|---|---|
| Format | MLX enhanced oQ6e |
| Size | 23,716,287,912 bytes (~22.09 GiB) |
| Group size / mode | 64 / affine |
| Quantized modules | 505 |
| Q6 modules | 470 |
| Protected Q8 modules | 35 |
| Importance matrix | 128 samples × 512 tokens; 504 entries |
| Vision tensors | 333, unquantized |
| Native MTP weights | 15, retained |
The embedding table and LM head are the two expected importance-matrix exceptions and are protected at Q8. The artifact contains no NaN or infinity values.
Before quantization, the Ektome BF16 source was compared tensor-by-tensor with the pinned official checkpoint:
Runtime checks passed for loading, native MTP drafting, tool calls and tool results, vision/OCR, and retrieval across a roughly 96K-token prompt. A true cold 95,995-token prefill completed without a memory-guard rejection on a 48 GB Apple Silicon host. The first visible answer was clipped by an overly small completion cap; a corrected recheck returned all distributed retrieval targets.
This is a quantized derivative, so “near-lossless” does not mean mathematically lossless. No post-quant benchmark suite is claimed beyond the validation above.
Download the repository into an oMLX model directory:
hf download kvas-25/Ektome-Qwen3.8-27B-oQ6e-MLX-MTP \
--local-dir ~/.omlx/models/Qwen3.8-27B-Ektome-oQ6e-mtp
omlx serve --model-dir ~/.omlx/models
Then select Qwen3.8-27B-Ektome-oQ6e-mtp through the oMLX admin UI or its
OpenAI-compatible API. Native speculative decoding requires enabling MTP in the
model settings; the validated configuration used five draft tokens. A practical
48 GB profile used a 98,304-token context window and Q4 TurboQuant KV cache.
Vision assets are included and were validated, but a text-only loader override can be used when a client does not need image input.
| Component | Revision/version |
|---|---|
| Ektome BF16 source | 7be830b2e33f9e91a1bd31628a06980a1b075803 |
| Official Qwen source | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 |
| Refusal directions audit | 4481c7d78458be4e5e33d16bfe6193d05400981b |
| Rank-1 forensics | 6772ac141aeda4e58d97118da7c6e40b1bfc2259 |
| oMLX build | 0.6.2 / f2d36f3d25a7e7a2401a92eecafc28b8f8968ec7 |
| MLX / mlx-lm / mlx-vlm | 0.32.0 / 0.31.3 / 0.6.3 |
Apache-2.0, inherited from Qwen3.8-27B and the Ektome source. See the upstream model cards for their full methodology, limitations, and usage notes. No warranty is provided; users are responsible for lawful and responsible use.