README.md

---
library_name: mlx
pipeline_tag: text-generation
inference: false         # MLX is macOS-only; HF Inference API won't run it
license: apache-2.0
base_model: openai/gpt-oss-20b
base_model_relation: quantized
language:
- en
- ro
tags:
- apple-silicon
- metal
- arm64
- 5-bit
- group-size-32
- moe
- mpx4
- openai
- halley-ai
---
# gpt-oss-20b — MLX 5-bit (group size 32)

**Summary.** This is a 5-bit (**Q5**) **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 Q5, `q_group_size=32` (some tensors remain FP16 for stability)
- **Files:** MLX weight shards + `config.json`; tokenizer files included for drop-in use
- **Footprint:** ~**15.76 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 minimum** for Q5 gs=32. Extra RAM improves headroom.

## How to use (MLX)

```bash
pip install mlx-lm transformers
```

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

model, tokenizer = load("halley-ai/gpt-oss-20b-MLX-5bit-gs32")
print(generate(
    model, tokenizer,
    prompt="Explain the Chudnovsky algorithm to compute π.",
    max_tokens=256, max_kv_size=512
))
```

```bash
# CLI
python -m mlx_lm generate --model halley-ai/gpt-oss-20b-MLX-5bit-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, Q5 gs=32) ≈2k-token responses:
- M1 Max (32 GB): ~45–50 tok/s, 0.40–0.60 s TTFB
- M4 Pro (24 GB): ~65–70 tok/s, 0.25–0.45 s TTFB
- M3 Ultra (256 GB): pending

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><strong>MLX 5-bit (gs=32)</strong></td><td><strong>11.11 (+3.3% vs 8-bit/gs64, +6.2% vs 6-bit/gs32)</strong></td></tr>
    <tr><td>MLX 4-bit (gs=32)</td><td>13.70 (+27.4% vs 8-bit/gs64, +31.0% vs 6-bit/gs32)</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)

```bash
python -m mlx_lm convert \
  --hf-path openai/gpt-oss-20b \
  --mlx-path gpt-oss-20b-mlx-q5-gs32 \
  --q-bits 5 --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-4bit-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 Q5, gs=32)
- Please cite both the base model and this repository when you use the weights.