Kalsa

Engines

Browse all 40 engines, A to Z. Filter them by chip, licence or what they do better. Listed and community engines explains what the Listed mark needs.

Review
Kind
achyudh/llama.cppby achyudh

Speeds up decoding in llama.cpp on AMD GPUs.

Fork of ggml-org/llama.cpp
AMDForkMIT
Faster decode
+8% faster decoding than llama.cpp on AMD GPU
architectds/Strataby architectds

Speeds up prompt processing in Strata on NVIDIA GPUs.

Fork of Niko1221/Strata
NVIDIAForkMIT
Faster prefill
+94% faster prompt reading on NVIDIA GPU
Aristo94/EngramHalo.cppby Aristo94

Speeds up Qwen 3.8 Flash-Next decoding on AMD Strix Halo with MTP, in a llama.cpp fork.

Fork of ggml-org/llama.cpp
AMDForkMIT
Faster decode
+61% faster decoding than without it on AMD GPU
ax-engineby defai-digital

Runs Qwen 3.8 27B with AXQ quantization and MTP on MLX for Apple M5.

Apple M5EngineApache-2.0
Faster decode
2.7× faster decoding than mlx-lm on M5 Max
bstnxbt/dflash-mlxby bstnxbt

Runs DFlash speculative decoding on MLX for Apple M5.

Apple M5EngineApache-2.0
Faster decode
2.4× faster decoding than mlx-lm on M5 Max
dmlxby zouyee

Decodes faster and starts answering sooner on Apple Silicon, using an MTP drafter.

Apple M1–M6EngineMIT
Faster decodeFaster start
+20% faster decoding on M4 Pro
DwarfStar 4by antirez

Runs large models such as DeepSeek V4 Flash and GLM 5.3 Flash on Apple Silicon, NVIDIA and AMD.

Apple M1–M6NVIDIAAMDEngineMIT
Bigger modelsMultiple machinesNew hardware
No benchmarks yet
frinkby antonellof

Loads GGUF models in pure Rust, with quantized Metal kernels and lossless speculative decoding.

Apple M1–M5EngineApache-2.0
Faster decode
+47% faster decoding than llama.cpp on M2 Pro
GenerelSchwerz/llama.cppby GenerelSchwerz

Changes how llama.cpp places its cache on NVIDIA GPUs, and decodes Qwen3.6 35B faster than stock.

Fork of ggml-org/llama.cpp
NVIDIAForkMIT
Faster decode
2.3× faster decoding than llama.cpp on NVIDIA GPU
IngeniousIdiocy/ds4by IngeniousIdiocy

Speeds up GLM-5.3-Flash on Apple M3 in a ds4 fork.

Fork of antirez/ds4
Apple M3ForkMIT
Faster decodeFaster prefill
+60% faster decoding on M3 Ultra
kapitsaby nace-ai

Runs Qwen3.6-35B-A3B 4-bit on Apple M5 with faster decoding and prompt reading.

Apple M5EngineApache-2.0
Faster decodeFaster prefill
+49% faster decoding on M5 Max
Lidenburg/llama.cppby Lidenburg

Adds an expert cache with GPU and RAM tiers to llama.cpp on NVIDIA GPUs.

Fork of ggml-org/llama.cpp
NVIDIAForkMIT
Faster decode
+60% faster decoding than without it on NVIDIA GPU
llama.cppby ggml-org

Serves as the reference upstream and runs on Apple Silicon, NVIDIA, AMD, Intel and CPU.

Apple M1–M5NVIDIAAMDIntelCPUReferenceMIT
The reference engine
llama.cpp adaptive KV streamingby RaymondHuang210129

Streams the KV cache between host RAM and VRAM in a llama.cpp fork for NVIDIA GPUs.

Fork of ggml-org/llama.cpp
NVIDIAForkMIT
Faster decode
10.4× faster decoding on NVIDIA GPU
llama.cpp adaptive spec decoding + Strix Halo Vulkanby LaurentZuijdwijk

Speeds up prompt reading in llama.cpp on AMD Strix Halo, using adaptive speculative decoding.

Fork of ggml-org/llama.cpp
AMDForkMIT
Faster prefill
+89% faster prompt reading than llama.cpp on AMD GPU
llama.cpp MoE expert cacheby thecodacus

Keeps MoE experts in a VRAM-resident cache in a llama.cpp fork for NVIDIA GPUs.

Fork of ggml-org/llama.cpp
NVIDIAForkMIT
Faster decodeFaster prefill
+21% faster decoding than without it on NVIDIA GPU
Localsby Generated-Computers

Serves many requests at the same time on Apple M2 with higher total throughput.

Apple M2EngineApache-2.0
Parallel throughput
No benchmarks yet
mistral.rsby EricLBuehler

Runs models in Rust with ISQ and UQFF quantization on Apple Silicon and NVIDIA GPUs.

Apple M1–M5NVIDIAEngineMIT
Faster prefill
+35% faster prompt reading than llama.cpp on NVIDIA GPU
mlx-lmby ml-explore (Apple)

Serves as the reference upstream for Apple Silicon, built on MLX.

Apple M1–M6ReferenceMIT
The reference engine
mlxcelby lablup

Decodes Qwen3-30B-A3B faster than mlx-lm on an M5 Max.

Apple M1, M5EngineApache-2.0
Faster decode
+16% faster decoding than mlx-lm on M5 Max
MTPLXby youssofal

Runs on Apple Silicon and is the upstream of the MTPLX M1-family build.

Apple M1–M6ReferenceApache-2.0
The reference engine
MTPLX M1-family buildby shunya1810

Speeds up MTPLX decoding on M1-family Macs.

Fork of youssofal/MTPLX
Apple M1ForkApache-2.0
Faster decode
2.2× faster decoding than MTPLX on M1 Max
neurall/llama.cppby neurall

Streams experts from SSD or RAM into GPU passes, which speeds up prompt processing in llama.cpp on NVIDIA GPUs.

Fork of ggml-org/llama.cpp
NVIDIAForkMIT
Faster decodeFaster prefill
+62% faster decoding on NVIDIA GPU
neutronstarby giannisanni

Adds batch prefill for GLM models to ds4 on NVIDIA GPUs.

Fork of antirez/ds4
NVIDIAForkMIT
Faster prefill
Not recommended: ~0.4 tok/s generation (about 2.5 s per token; the README's own interactive figure is ~2.1 tok/s)21.7× faster prompt reading than without it on NVIDIA GPU
Ollamaby ollama

Serves as the reference upstream: a UX wrapper with a model library.

Apple M1–M5NVIDIAAMDCPUReferenceMIT
The reference engine
oMLXby jundot

Speeds up prompt reading on Apple Silicon by pairing a second Mac.

Apple M1–M5EngineApache-2.0
Faster prefill
2.8× faster prompt reading on M3 Ultra
paperniuk/splashby paperniuk

Runs Splash on M1 and M2 Macs, where the upstream needs an M3 or newer.

Fork of incoai/splash
Apple M1, M2ForkApache-2.0
New hardware
No benchmarks yet
QVAC Fabricby tetherto

Offloads work to Core ML on Apple Silicon and to the XDNA NPU on AMD, in a llama.cpp fork.

Fork of ggml-org/llama.cpp
Apple M1–M6AMDForkMIT
Faster decodeFaster prefill
+38% faster prompt reading than without it on M3 Ultra
Rapid-MLXby raullenchai

Builds on mlx-lm to decode Qwen3.5-9B faster than oMLX on Apple Silicon.

Built on ml-explore/mlx-lm
Apple M1–M6EngineApache-2.0
Faster decode
+48% faster decoding than oMLX on M5 Max
ROCmForgeby oldnordic

Decodes Qwen2.5-0.5B faster than llama.cpp on AMD GPUs.

AMDEngineGPL-3.0
Faster decode
+42% faster decoding on AMD GPU
SelfRef/llama.cpp-rdnaby SelfRef

Shrinks Qwen3.8-27B with the ROCmFPX format and decodes it faster, in a llama.cpp fork for AMD GPUs.

Fork of ggml-org/llama.cpp
AMDForkMIT
Faster decodeLess memory
+7% faster decoding on AMD GPU
spiritbuun/buun-llama-cppby spiritbuun

Adds a trellis-coded 3-bit KV cache codec to llama.cpp for NVIDIA and AMD GPUs.

Fork of ggml-org/llama.cpp
NVIDIAAMDForkMIT
Quality
No benchmarks yet
Splashby incoai (Inco AI org)

Decodes Qwen3.8-27B and Qwen3.6-35B-A3B faster than llama.cpp on Apple M3 and newer.

Apple M3–M5EngineApache-2.0
Faster decode
5.4× faster decoding than llama.cpp on M3 Max
stew675/llama-cpp-rdna-boostsby stew675

Speeds up decoding in llama.cpp on AMD GPUs with RDNA patches and a speculative-decoding clamp.

Fork of ggml-org/llama.cpp
AMDForkMIT
Faster decode
2.3× faster decoding than without it on AMD GPU
Strataby Niko1221

Serves as the reference upstream for Strata forks on NVIDIA, AMD and Apple Silicon.

Apple M1–M6NVIDIAAMDReferenceMIT
The reference engine
SwiftLMby SharpAI

Streams experts from the SSD to cut GPU memory use on Apple Silicon Macs.

Apple M1–M6EngineMIT
Less memory
No benchmarks yet
tandem.cppby ewhacc

Stores MoE experts in RAM tiers for models running on NVIDIA GPUs.

NVIDIAEngineMIT
Faster decode
+55% faster decoding on NVIDIA GPU
turbo-tan/llama.cpp-tq3by turbo-tan

Adds TQ3 weight formats to llama.cpp and an FP4 prompt path on Blackwell GPUs.

Fork of ggml-org/llama.cpp
NVIDIAForkMIT
Faster prefill
2.6× faster prompt reading than without it on NVIDIA GPU
vLLMby vllm-project

Serves as the reference upstream for NVIDIA, AMD and Intel GPUs and CPU.

NVIDIAAMDIntelCPUReferenceApache-2.0
The reference engine
vMLXby jjang-ai

Cuts memory use on Apple Silicon by loading only some of each MoE model's experts.

Apple M1–M5EngineApache-2.0
Less memory
No benchmarks yet