SmallThinker-4BA0.6B-Instruct-mlx-4bit

SmallThinker-4BA0.6B-Instruct (MLX 4-bit)

An MLX 4-bit (group size 64, affine) conversion of PowerInfer/SmallThinker-4BA0.6B-Instruct: a 4B-total / roughly 0.6B-active dense-attention Mixture-of-Experts model (32 experts, top-4, ReLU-gated), non-thinking Instruct.

Requires an mlx_lm architecture module

Stock mlx_lm does not yet ship a smallthinker architecture, so this model will not load with an unmodified install. Until the upstream mlx-lm PR lands, copy the included smallthinker.py into your mlx_lm/models/ directory:

cp smallthinker.py "$(python -c 'import mlx_lm,os;print(os.path.join(os.path.dirname(mlx_lm.__file__),"models"))')/"

Then:

python -m mlx_lm generate \
  --model Aa09876/SmallThinker-4BA0.6B-Instruct-mlx-4bit \
  --prompt "Explain a mixture-of-experts model in two sentences."

Tool calling

The model uses the native Hermes tool-call format, <tool_call>{json}</tool_call>, and works with servers that select a Hermes tool parser (for example rapid-mlx --tool-call-parser hermes, no reasoning parser).

Provenance and lineage

• Base weights: PowerInfer/SmallThinker-4BA0.6B-Instruct (Apache-2.0), pinned revision b51db6d
• Architecture support: a from-scratch mlx_lm implementation of the custom SmallThinkerForCausalLM (ReLU-gated packed-expert MoE, GQA with 12 query / 2 KV heads, and the model's pre-attention router-input routing). It was numerically parity-checked against the reference PyTorch model: BF16 greedy generation matched token-for-token, router top-k selection matched 100% (tie-adjusted), and per-tensor error stayed at the BF16 rounding floor.
• Quantization: MLX 4-bit, group size 64, affine.

License

Apache-2.0, inherited from the base model. Attribution: PowerInfer / IPADS (SmallThinker). This repository redistributes a quantized derivative under the same license.