Roadmap
Available. In progress. Planned.
What works today, what is being built, and what is intended. Status is copied from the repository; this page makes no delivery commitments and lists no dates.
- Available
- 15 items
- In progress
- 4 items
- Planned
- 6 items
- StableReady for everyday use.
- PreviewWorks and passes the test suite. Feedback welcome.
- In progressActively being built.
- PlannedIntended, not started or not yet usable.
Preview items work and are listed under Available with their own label. Source: ROADMAP.md and README.md (opens in a new tab)
Model compatibility
Available
- StableLlama, Qwen, Gemma, Phi, Mistral, Granite familiesSee README for exact versions.
- StableGGUF, SafeTensors, GPTQ and AWQ loading
In progress
—No items
Planned
- PlannedRemaining i-quant types (IQ1–IQ3), TQ1_0/TQ2_0, MXFP4Refused at load today, with the tensor named.
- PlannedVision-language models
CPU inference
Available
- StableAVX2 and NEON SIMD kernels
- PreviewWindows on ARM
In progress
- In progressIntel AMX / AVX-512 matrix acceleration
Planned
—No items
GPU execution
Available
- PreviewApple Metal backend
- PreviewVulkan backend (AMD, Intel, NVIDIA, Qualcomm)
- PreviewNVIDIA CUDA backendBuilt and unit-tested in CI; not yet measured on NVIDIA hardware.
- StableAutomatic CPU / GPU selection
In progress
- In progressGPU multi-user throughput
Planned
—No items
Scheduling and batching
Available
- StableContinuous batching and chunked prefill
- StableSpeculative decoding
In progress
—No items
Planned
- PlannedDisaggregated prefill and decode serving
Memory optimization
Available
- StablePaged KV cache with prefix sharing
- StableQ8_0 KV cache pages
In progress
—No items
Planned
- PlannedFP8 / FP4 microscaling formats
Performance and benchmarking
Available
- StableBuilt-in benchmark command
In progress
- In progressPublic, reproducible benchmark dashboard on this site
Planned
—No items
Developer tooling
Available
- StablePrometheus metrics endpoint
In progress
—No items
Planned
- PlannedStructured JSON output (grammar masking)
- PlannedTool / function calling in the chat API
Documentation
Available
- StableReference guides in the repository
In progress
- In progressDocumentation hosted on dynalm.in
Planned
—No items
Influence
Shape what comes next.
Priorities follow real use. Propose features and report problems on GitHub.