Labels

Dense transformer decoder

Hybrid attention stack (linear/sliding/full or attention+SSM interleave)

Sparse mixture-of-experts decoder

State-space / recurrent model (Mamba, RWKV)

Architecture and code structure changes

Benchmark harness and performance measurement (bench_*.sh, /update-benchmarks)

Command-line interface / CLI flags

mlxcel-core: MLX FFI, primitives, KV cache, layers

User and developer documentation

Generation, sampling, decoding (incl. speculative, DRY)

Model architectures, weights, loading, metadata

Model surgery operations (merge, interpolate, replace, etc.)

This issue or pull request already exists

Good for newcomers

Extra attention is needed

Breaking change requiring migration

Feature marked for future removal

4-bit checkpoint 50-100GB; runnable but dominates the machine

4-bit checkpoint 10-50GB; comfortable on a 128GB dev box

4-bit checkpoint <= 10GB; fast iteration, safe for smoke tests

4-bit checkpoint > 100GB; exceeds a 128GB dev box, needs bigger hardware

Audio or speech model

Multi-modal omni model (text + vision + audio)

Text-only language model

Vision-language model

Linux (CUDA / packaging) specific

macOS (Apple Silicon) specific

Windows (native) specific

Future considerations