Add Gemma 4, LFM2, and Qwen3-Coder models to catalog

[weklund-agent]

Summary

The catalog is missing several high-performing models that are available on mlx-community and score well on agentic coding benchmarks (per mlx_transformers_benchmark):

Model	Params	Quality (M4 Pro 64GB)	Gen tok/s	RAM (int4)
Qwen3-Coder-30B-A3B	30B MoE (3B active)	65.7%	80	17.8 GB
Gemma 4 E2B-it	2.3B dense	65.3%	121	3.5 GB
LFM2-24B-A2B	24B MoE (2.3B active)	67.3%	117	14.2 GB

On M5 Max 128GB, Qwen3-Coder-30B-A3B is the top model at 75.5% quality / 129 tok/s.

HuggingFace sources

Qwen3-Coder-30B-A3B:

• int4: mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit
• int8: mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit
• bf16: Qwen/Qwen3-Coder-30B-A3B-Instruct
• Capabilities: tool_calling (hermes), thinking (qwen3)

Gemma 4 E2B-it:

• int4: mlx-community/gemma-4-e2b-it-4bit
• int8: mlx-community/gemma-4-e2b-it-8bit
• bf16: google/gemma-4-E2B-it
• Note: blocked by mlx_lm gemma4 arch support — see related issue

LFM2-24B-A2B:

• int4: mlx-community/LFM2-24B-A2B-4bit
• int8: LiquidAI/LFM2-24B-A2B-MLX-8bit (published by LiquidAI, not mlx-community)
• bf16: LiquidAI/LFM2-24B-A2B-MLX-bf16
• Architecture: mamba2-hybrid

Notes

• None of these models are gated on HuggingFace
• Gemma 4 is currently blocked by vllm-mlx's bundled mlx_lm not supporting the gemma4 architecture (separate issue)
• All three were manually added to catalog YAML during testing and work correctly (except Gemma 4 due to the mlx_lm issue)