gpt-oss-20b-MLX-4bit-gs32

gpt-oss-20b — MLX 4-bit (group size 32)

Summary. This is a 4-bit (Q4) MLX quantization of gpt-oss-20B (sparse Mixture-of-Experts, MPx4). Group size is 32.
Built for Apple Silicon with Metal acceleration.

• Base model: openai/gpt-oss-20b (Apache-2.0)
• Quantization: MLX Q4, q_group_size=32 (some tensors remain FP16 for stability)
• Files: MLX weight shards + config.json; tokenizer files included for drop-in use
• Footprint: ~13.11 GB on disk
• Intended use: local inference / research on M-series Macs
• Not intended for: safety-critical decisions; outputs may be inaccurate or biased

Requirements

Runs on: Apple Silicon (M1 or newer) with macOS ≥ 13.5 via MLX (Metal).
Not supported: Intel macOS / Linux / Windows (use a GGUF build + llama.cpp instead).
RAM guidance: 24 GB+ required for Q4 gs=32 (does not work on Mac with only 16 GB).

How to use (MLX)

pip install mlx-lm transformers

# Python API (uses tokenizer bundled with this repo)
from mlx_lm import load, generate

model, tokenizer = load("halley-ai/gpt-oss-20b-MLX-4bit-gs32")
print(generate(
    model, tokenizer,
    prompt="Explain the Chudnovsky algorithm to compute π.",
    max_tokens=256, max_kv_size=512
))

# CLI
python -m mlx_lm generate --model halley-ai/gpt-oss-20b-MLX-4bit-gs32 \
  --prompt "Explain the Chudnovsky algorithm to compute pi." \
  --max-kv-size 512 --max-tokens 256

Performance (Apple Silicon, real-world)

LM Studio / CLI (MLX, Q4 gs=32) ≈2k-token responses:

• M1 Max (32 GB): ~62–72 tok/s, 0.30–0.40 s TTFB
• M4 Pro (24 GB): ~80-90 tok/s, 0.40–0.60 TTFB
• M3 Ultra (256 GB): ~130–140 tok/s, 0.20–0.30 s TTFB

Throughput varies with Mac model, context, and sampler settings.

Evaluation

Perplexity (PPL) streaming evaluation on WikiText-2; window=stride=4096, ~100k tokens, EOS inserted between docs.

<table> <thead> <tr><th>Variant</th><th>PPL (ctx=4096)</th></tr> </thead> <tbody> <tr><td>MLX 8-bit (reference)</td><td>10.75</td></tr> <tr><td>MLX 6-bit (gs=32)</td><td>10.46 (−2.7% vs 8-bit/gs64)</td></tr> <tr><td>MLX 5-bit (gs=32)</td><td>11.11 (+3.3% vs 8-bit/gs64, +6.2% vs 6-bit/gs32)</strong></td></tr> <tr><td><strong>MLX 4-bit (gs=32)</strong></td><td><strong>13.70 (+27.4% vs 8-bit/gs64, +31.0% vs 6-bit/gs32)</strong></td></tr> </tbody> </table>

Interpretation

• MLX 6-bit/gs32: Best of the group; edges out 8-bit/gs64 slightly at a smaller footprint.
• MLX 5-bit/gs32: Small, consistent drop vs 6-bit/gs32 and 8-bit/gs64 (~3–6% PPL); strong “fits-16GB” option when GPU buffer limits matter.
• MLX 8-bit/gs64: Solid reference; near‑FP16 quality at a larger footprint.
• MLX 4-bit/gs32: Trades accuracy for footprint; use when RAM is constrained or throughput is the priority.

Conversion details (provenance)

python -m mlx_lm convert \
  --hf-path openai/gpt-oss-20b \
  --mlx-path gpt-oss-20b-mlx-q4-gs32 \
  --q-bits 4 --q-group-size 32 -q

• Some non-expert tensors (embeddings, norms, router) remain FP16.

Sibling & reference models

• halley-ai/gpt-oss-20b-MLX-6bit-gs32
• halley-ai/gpt-oss-20b-MLX-5bit-gs32
• Reference (8-bit, upstream): lmstudio-community/gpt-oss-20b-MLX-8bit

Limitations & biases

Outputs may be factually wrong or unsafe. Don’t use for medical, legal, or financial decisions without human review. MoE models can be sensitive to prompt wording; prefer explicit instructions and structure.

License & credits

• License: Apache-2.0 (inherits from base model)
• Base model: OpenAI gpt-oss-20B
• Quantization: Halley AI Lab (MLX Q4, gs=32)
• Please cite both the base model and this repository when you use the weights.