ur-dad-matt's picture
Card rebuilt from outlier.host: site figures with build and date, current tiers, working links
c698aec verified
|
Raw History Blame Contribute Delete
3.94 kB
metadata
language:
  - en
  - zh
  - fr
  - es
  - pt
  - de
  - it
  - ru
  - ja
  - ko
  - ar
  - vi
  - th
  - nl
  - pl
license: apache-2.0
library_name: mlx
base_model: Qwen/Qwen3.6-27B
tags:
  - 4-bit
  - 4bit
  - apple-silicon
  - best-quality
  - chat
  - conversational
  - edge-ai
  - flagship
  - function-calling
  - humaneval
  - instruct
  - local-llm
  - m1
  - m2
  - m3
  - m4
  - mac
  - mac-mini
  - mac-studio
  - macbook-air
  - macbook-pro
  - macos
  - metal
  - mlx
  - mlx-lm
  - no-cloud
  - offline
  - on-device
  - outlier
  - outlier-app
  - private
  - private-ai
  - quantized
  - qwen
  - qwen3.6
  - qwen3_5
  - reasoning
  - safetensors
  - text-generation
  - thinking
  - tool-use
pipeline_tag: text-generation
model-index:
  - name: Outlier-Ai/Outlier-Core-27B-MLX-4bit
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (200-question sample, build 1.11.804, 2026-08-24)
          type: cais/mmlu
        metrics:
          - type: acc
            name: accuracy
            value: 0.895
            verified: false
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HumanEval (all 164, build 1.11.804, 2026-08-24)
          type: openai_humaneval
        metrics:
          - type: pass@1
            name: pass@1
            value: 0.9512
            verified: false

Part of the Outlier shipping lineup. Outlier is a free macOS app that runs this model locally, with one click. Apple Silicon only. Outlier for Mac: free Nano and Lite; Pro is $49 once, on up to 3 Macs · brew install --cask outlier-host/tap/outlier

Outlier Core 27B (MLX 4-bit)

What this is: Qwen3.6-27B, text-only, converted to MLX 4-bit. Nothing else is changed: no fine-tune, no merge. It is the Core tier of the Outlier app for Mac, which downloads exactly these files.

  • Download size: 15.13 GB · minimum unified memory: 24 GB · Apple silicon
  • Licence: Apache 2.0, the base model's licence. The Outlier team converted the weights; the model is the base model's authors' work.
  • In the app: part of Pro ($49 once)

Measured results

These are the figures on outlier.host, with how each was run. Small samples: at n=200 an MMLU figure can be off by up to 7 points either way.

Test Result How
MMLU 89.5% (179/200) a 200-question sample, build 1.11.804, 2026-08-24
HumanEval 95.1% (156/164) all 164 problems, official tests, build 1.11.804, 2026-08-24
SWE-bench Verified (agentic) 46.0% (23/50), blind, official Docker harness, build 1.11.757, 2026-08-09 a random 50-task sample (seed 42), no test patch or file hint
Decode speed, M1 Ultra 20.7 tok/s (measured 2026-04-29)

The app you download is a later build than the one measured; these figures have not been re-run on it.

Use it

The easiest way is the Outlier app: pick Outlier Core in the model picker and it downloads these files.

To load the weights yourself:

pip install mlx-lm
python -m mlx_lm.generate --model Outlier-Ai/Outlier-Core-27B-MLX-4bit --prompt "Hello" --max-tokens 256

The other Outlier tiers

Earlier experiments in this organisation (Outlier-10B, -40B, -70B, -150B and others) are research checkpoints. The app does not ship them.