Roadmap · rapid-mlx 0.15.7

Roadmap

Rapid-MLX ships small releases often rather than working to a fixed long-range plan. The best view of what is coming is the open work on GitHub; the best view of what is settled is everything outside the experimental list below.

Where to follow what's next

GitHub issues — planned work, model requests and bugs. When pull or serve refuses a Hugging Face model, adding --request files a support request for it.
Pull requests — what is being built right now.
Changelog — what each release changed.
Discord — tell us what you want (rapid-mlx feedback opens it).

Experimental in 0.15.7

These work today but are marked experimental: their behaviour, flags or hardware requirements may still change.

Area	What it is
Large-Mac model lanes	`glm5.3-flash-4bit`, `mimo-v2.6-flash-4bit`, `qwen3.8-flash-next-4bit` and `deepseek-v41-flash-reap-2bit` — frontier MoE models for 192–256 GB Mac Studios. `rapid-mlx info <alias>` shows each one's limits.
Accelerated profiles	`qwen3.8-27b-tensorfold`, `glm5.3-flash-tensorfold`, `nemotron-3.5-lightning-tensorfold` (48 GB) and `qwen3.8-flash-next-tensorfold` (192 GB) — opt-in speculative profiles that need the separately installed TensorFold runtime. They serve text chat without tools, media or grammar.
Other experimental aliases	`k2-horizon-7b-4bit`, `neohorse-9b-4bit`, `qwen3.8-27b-abliterated-4bit`.
Decision models	`rapid-mlx system-one` — a separate typed-decision server (Laya, CLM, and Clef / Clef-Flash through the `[clef]` extra) for `/v1/systemone` and `/v1/rank`.
Computer use	`rapid-mlx cua` — a native macOS accessibility agent with a configurable planner.
Speculative decoding	DDTree, DFlash pairs below 8-bit, and SuffixDecoding on aliases not yet benchmarked for it (see performance flags).
Memory techniques	`--use-paged-cache`, `--kv-cache-turboquant v4`, `--disk-stream` for MoE experts, and `--kv-disk-checkpoint-interval` (snapshots are written but not yet reloaded).

Contribute

Issues labelled for contributors and the contribution guide are on GitHub. Benchmarks from your own Mac help too: rapid-mlx benchmark run <alias> (any alias from rapid-mlx benchmark catalog), then rapid-mlx benchmark share <run-id> adds them to the leaderboard.