license: apache-2.0
library_name: mlx
pipeline_tag: image-text-to-text
base_model: Jiunsong/SuperQwen3.8-27b-abliterated
base_model_relation: quantized
tags:
- mlx
- mlx-lm
- qwen3.8
- qwen3.5
- apple-silicon
- 4-bit
- quantized
- multimodal
- reasoning
- tool-calling
- uncensored
- abliterated
- obliteratus
- supertune
language: - en
- ko
SuperQwen3.8-27b-abliterated-MLX-4bit
A native Apple-silicon build of SuperQwen3.8: 4-bit affine weights, corrected overthinking, tool use, and 30 tok/s single-stream decode on M5 Max.
This is the directly loadable MLX release ofJiunsong/SuperQwen3.8-27b-abliterated,
built on the pinnedmlx-community/Qwen3.8-27B-4bit@3e6447f.
It is not an adapter: the changed tensors were dequantized, the abliteration delta was
applied, and each affected matrix was quantized back into native MLX weight/scales/biases.
Release highlights
| Format | MLX affine 4-bit, group size 64, three safetensors shards (~14.96 GiB) |
| Single-stream decode | 29.665 / 29.960 / 30.389 tok/s, median 29.960 tok/s |
| Memory | 15.94 GB peak in the fixed 512-token C1 benchmark |
| Abliteration transfer | 100 source tensors → 300 native MLX weight/scales/biases entries |
| Reasoning regression gate | 36/36 PASS across default, low, medium, and xhigh |
| Behavior | Capability 7/8 paired-parent floor, tool PASS, benign-sensitive refusal 0/8 |
| Vision storage | 333 vision tensors retained; this release's measured runtime gate is text/tool, not MLX vision inference |
Why this build
- Runs the 27B model locally on a Mac without a server-class GPU.
- Fast where it matters: the headline number is C1 single-stream decode, not a concurrency aggregate.
- Reasoning that stops: explicit
xhighis preserved, but carries the same repeat/restart stop guard as the BF16 release. - No loose delta at runtime: the abliteration edit is baked into the MLX shards.
- Auditable conversion: the pinned base revision, every changed key, and shard hashes are recorded in
mlx_requantization_manifest.json.
Measured on Mac
Measured locally on an Apple M5 Max with MLX-LM 0.31.3. The benchmark uses one
stream, a 256-token prompt, 512 fixed generation tokens, greedy decoding, one warmup,
and three timed trials.
| Trial | C1 decode | Prompt processing | Peak memory |
|---|---|---|---|
| 1 | 30.389 tok/s | 682.571 tok/s | 15.942 GB |
| 2 | 29.960 tok/s | 661.518 tok/s | 15.942 GB |
| 3 | 29.665 tok/s | 537.247 tok/s | 15.943 GB |
| Median / average | 29.960 tok/s median | 627.112 tok/s average | 15.943 GB |
The quality run loaded the model once, executed all 53 checks in 176.6 seconds, and
peaked at 16.06 GB. Exact reports and hashes are under evidence/.
Reasoning that stops
The upstream template's implicit reasoning default is changed to bounded medium.
Explicit xhigh remains available and tells the model to stop once the answer is
established instead of repeating or restarting its deliberation.
| Effort | Deterministic overthinking gate | Median reasoning tokens | Maximum |
|---|---|---|---|
| default | 9/9 | 41 | 177 |
| low | 9/9 | 53 | 106 |
| medium | 9/9 | 41 | 177 |
| xhigh | 9/9 | 40 | 51 |
| Total | 36/36 PASS |
Run it
python3.11 -m pip install -U mlx-lm
python3.11 -m mlx_lm generate \
--model Jiunsong/SuperQwen3.8-27b-abliterated-MLX-4bit \
--prompt "Explain why speculative decoding is lossless." \
--max-tokens 512
Use explicit bounded or deep reasoning through the chat-template configuration:
python3.11 -m mlx_lm generate \
--model Jiunsong/SuperQwen3.8-27b-abliterated-MLX-4bit \
--chat-template-config '{"enable_thinking":true,"reasoning_effort":"xhigh","preserve_thinking":false}' \
--prompt "Solve this carefully: ..." \
--max-tokens 4096
Serve an OpenAI-compatible local endpoint:
python3.11 -m mlx_lm server \
--model Jiunsong/SuperQwen3.8-27b-abliterated-MLX-4bit \
--host 127.0.0.1 \
--port 8080
DFlash 2 compatibility
incoai/Qwen3.8-27B-DFlash2
was tested rather than merely listed. It is a DFlash2DraftModel with block-diffusion
and selector layers, while MLX-LM 0.31.3 accepts a standard causal draft model through
its speculative interface. The checkpoint therefore does not load as an MLX draft
model, so it is not bundled and no unmeasured speedup is claimed.
Quantization and integrity
- Mode: affine 4-bit, group size 64.
- Pinned MLX base index SHA-256:
13b840162b4cb35c66fef7df072f7dbb4717908204364f5e5d9f9655a2758fa8. - Pinned delta manifest SHA-256:
bf6d63af84e42b2e620992729f73e2ccc98fbc1532a15aa329a57b4474600f70. - Conversion manifest SHA-256:
57a416917d47f816319cd009b8df89934fb07cd06ed4e02f3be7646f60490ade. - Three output shards are independently hashed in both the manifest and
SHA256SUMS.json.
Evidence identities
| Evidence | SHA-256 |
|---|---|
| Fixed 512-token MLX C1 benchmark | 63576f44d61d77f0765fc903c640017299d16b066b95fcc61d7d3cbaa1bdeff0 |
| MLX quality and overthinking gate | 028628e3787dd4d361fce641cd8dfc6e18fb991c939bd4f8108a18eaae6121f1 |
| MLX requantization manifest | 57a416917d47f816319cd009b8df89934fb07cd06ed4e02f3be7646f60490ade |
| DFlash 2 compatibility decision | Recorded in evidence/dflash2_compatibility.json and sealed by SHA256SUMS.json |
Limitations
- Four-bit affine quantization can regress tasks outside the measured suite.
- Capability is 7/8, matching the accepted paired-parent floor; it is not presented as 8/8.
- The stored multimodal tensors are retained, but this release was performance- and quality-tested through MLX-LM's text/tool path.
- DFlash 2 requires a dedicated block-diffusion runtime implementation and is not enabled in this MLX build.
- Local speed varies with Apple chip, memory pressure, prompt length, and MLX version.
License
Apache-2.0, following the upstream Qwen3.8 release.