Kalsa

A database of local inference engines and forks

What each one is, who made it, its licence, the benchmarks its author published, and how we read them.

What are you running?

Pick a chip and meet the engines that run on it, A to Z.

17 engines run on Apple M1, in alphabetical order. Listed engines only. Solid numbers were measured on this chip family. Dashed ones were measured on another chip, so they say less about yours.

dmlxby zouyee

Zig runtime for Apple Silicon built around model-specific speculative decoding and SSD-streamed MoE models.

Apple M1–M6EngineMIT
Faster decodeFaster start
+20%decode +20% vs omlx-mtpMeasured on M4 Pro
Claimed by the author
DwarfStar 4by antirez

Inference engine in C built around a few large models (DeepSeek V4 Flash and PRO, GLM 5.x, Qwen3.8), with Metal, CUDA and ROCm backends.

Apple M1–M6NVIDIAAMDEngineMIT
Bigger modelsMultiple machinesNew hardware
No comparison published for this hardware yet.
frinkby antonellof

Pure-Rust GGUF loader with quantized Metal kernels and lossless speculative decoding.

Apple M1–M5EngineApache-2.0
Faster decode
+47%decode +47% vs llama.cpp (build unknown)Measured on M2 Pro
Claimed by the author
llama.cppby ggml-org

Reference upstream across Metal, CUDA, ROCm/HIP, Vulkan, SYCL and CPU.

Apple M1–M5NVIDIAAMDIntelCPUReferenceMIT
Reference engine: the baseline others compare against.
mistral.rsby EricLBuehler

Rust engine with ISQ/UQFF quantization.

Apple M1–M5NVIDIAEngineMIT
Faster prefill
No comparison published for this hardware yet.
mlx-lmby ml-explore (Apple)

Reference upstream for Apple Silicon (MLX).

Apple M1–M6ReferenceMIT
Reference engine: the baseline others compare against.
mlxcelby lablup

MLX engine for text and VLM.

Apple M1, M5EngineApache-2.0
Faster decode
+16%decode +16% vs mlx-lm 0.31.3Measured on M5 Max
Claimed by the author
MTPLXby youssofal

Mac app and command line that runs Qwen models on Apple Silicon using the model's own multi-token prediction heads.

Apple M1–M6ReferenceApache-2.0
Reference engine: the baseline others compare against.
MTPLX M1-family buildby shunya1810

Neural Engine prefill, M1 flash-attention and 8-bit paged KV.

Fork of youssofal/MTPLX
Apple M1ForkApache-2.0
Faster decode
2.2×decode 2.2× vs upstream MTPLX 2.12.2Measured on M1 Max
Claimed by the author
Ollamaby ollama

Reference upstream: UX wrapper with a model library.

Apple M1–M5NVIDIAAMDCPUReferenceMIT
Reference engine: the baseline others compare against.
oMLXby jundot

LLM inference server for Apple Silicon with continuous batching and SSD caching, managed from the macOS menu bar.

Apple M1–M5EngineApache-2.0
Faster prefill
2.8×prefill 2.8× vs This Mac aloneMeasured on M3 Ultra
Claimed by the author
paperniuk/splashby paperniuk

Unofficial port of Splash to M1 and M2 Macs, with Metal kernels written for those GPUs.

Fork of incoai/splash
Apple M1, M2ForkApache-2.0
New hardware
No comparison published for this hardware yet.
QVAC Fabricby tetherto

Core ML and XDNA NPU offload.

Fork of ggml-org/llama.cpp
Apple M1–M6AMDForkMIT
Faster decodeFaster prefill
+38%prefill +38% vs Core ML offMeasured on M3 Ultra
Claimed by the author
Rapid-MLXby raullenchai

Pure MLX kernels with continuous batching and prefix cache.

Built on ml-explore/mlx-lm
Apple M1–M6EngineApache-2.0
Faster decode
+48%decode +48% vs oMLXMeasured on M5 Max
Claimed by the author
Strataby Niko1221

Desktop app and engine for running Qwen3.8 models locally on NVIDIA, AMD and Apple Silicon, with an OpenAI-compatible local API.

Apple M1–M6NVIDIAAMDReferenceMIT
Reference engine: the baseline others compare against.
SwiftLMby SharpAI

Native Swift server for MLX models on Apple Silicon, with SSD streaming of MoE experts and a compressed KV cache.

Apple M1–M6EngineMIT
Less memory
No comparison published for this hardware yet.
vMLXby jjang-ai

MLX model server for Apple Silicon with adaptive mixed-precision quantization and partial loading of MoE experts from SSD.

Apple M1–M5EngineApache-2.0
Less memory
No comparison published for this hardware yet.

Published comparisons

Against llama.cpp or mlx-lm, as each author ran them. Listed A to Z, not ranked. Each names the chip it was measured on.

Recently added

The last entries in the catalog file. Dates will come from the repository history.

The people behind them

40 developers and organisations so far.

How we choose

What gets listed, what a review looks at, and how a published number is read. Read how we choose.