Quail, Ollama, oMLX, Rapid-MLX and Nativ

Five ways to run AI models on a Mac and give your tools an API. Here's what each does, with a source for every claim, and how they compare when measured on the same Mac with the same weights.

What each one does

Feature	Quail1.0	Ollama0.34.4	oMLX0.6.4	Rapid-MLX0.15.2	Nativ0.3.11
How you install it	Signed, notarized Mac app (DMG or Homebrew cask); no Python source	Signed, notarized Mac app or install script; no Python source	Signed, notarized Mac app with Python inside, or pip source	Homebrew or pip (Python); a separate signed desktop app source	Signed, notarized Mac app (DMG or Homebrew cask) with Python inside source
Model formats	GGUF and MLX, from one server source	GGUF and MLX (imported into its own store) source	MLX only source	MLX only source	MLX only source
OpenAI Chat Completions and Responses	Both source	Both (Responses without stored state) source	Both source	Both source	Both source
Anthropic Messages (Claude Code)	Yes, with count_tokens source	Yes, without count_tokens source	Yes, with count_tokens source	Yes, with count_tokens source	Yes, with count_tokens source
API key	On from the start source	None for the local server source	Asked for by the app's setup; off for the plain server source	Off unless set (the desktop app makes one) source	Off unless set source
Requests from web pages	Refused unless you allow an origin source	Only from this Mac's own pages and apps source	Any site (CORS *), changed only in settings.json source	Any site (CORS *) unless --cors-origins source	Any site (CORS *) source
Several requests at once	Yes: llama.cpp slots on GGUF, batched decoding on MLX source	One at a time by default (OLLAMA_NUM_PARALLEL); always one for some model families source	Yes, continuous batching (8 by default) source	Yes, continuous batching source	Yes, continuous batching (no limit by default) source
Prompt cache	Per conversation; MLX's also kept on disk source	In memory (MLX runner) source	In memory plus an SSD tier that survives restarts source	In memory, saved at shutdown source	Off by default; in memory plus an SSD tier when on source
Speculative decoding	Prompt lookup (fast when a reply repeats its prompt) source	MTP and DFlash on MLX source	DFlash and MTP, per model, off by default source	MTP, DFlash and more; on by default for some models source	DFlash, EAGLE-3 or MTP draft model; off by default source
Embeddings	Yes, and reranking, on GGUF and MLX source	Yes source	Yes, and reranking source	With an optional extra source	Yes source
Image generation	No (planned) source	No (experimental support removed in 0.32.6) source	No source	Yes, generation and edits (an optional extra) source	Yes, generation and edits source
Speech (transcription, speech)	No (planned) source	Transcription only (not in its docs) source	Both (the app includes the audio runtime) source	Both (an optional extra) source	Both, and translation source
MCP	No (planned: as a server, and its tools through the API) source	No source	MCP servers' tools through the API and its chat source	MCP servers' tools through its API and rapid-mlx chat source	MCP servers' tools in its chat, not through the API source
Starting coding tools	quail launch, and set-up snippets with a Test button source	ollama launch (about 25 tools) source	omlx launch, and one-click setup in its admin page source	rapid-mlx launch and agent setup source	Sets up and launches 10 terminal agents, and sets up 6 editors source
Usage data sent	None (the update check says which version is asking) source	None found (update checks send OS and version) source	Its benchmark uploads results with a hashed device ID source	On by default since 0.15.0 (can be turned off) source	None found (update checks go to GitHub) source
Platforms	macOS 14+, Apple silicon source	macOS, Windows, Linux source	macOS 15+ source	macOS source	macOS 26+, Apple silicon source
Licence	MIT source	MIT source	Apache-2.0 source	Apache-2.0 source	MIT source

Checked 2026-10-03 against each project's own docs, source or release notes at the version shown. Something wrong or out of date? Tell us.

Measured

Quail runs both formats, so it's compared in two lanes, each on identical weights: the same .gguf file against llama.cpp's own server and Ollama, and the same MLX folder against oMLX and Rapid-MLX. Nativ isn't measured yet (#188). Every engine gets the same prompts, the same settings and the same number of requests, one server at a time, on a Mac that's cooled down first.

How the comparison works, and what Quail changed after the first one: How it works.

Measured 2026-09-30 on Apple M1 Max, 64 GB, macOS 27.0.1. The full report has every level, the intervals and the caveats.

Qwen3 8B

Output tokens per second by requests at once, GGUF lane

GGUF lane	1 request, tok/s	8 at once, tok/s	First token (1 request)	GSM8K	MMLU-Pro	BFCL
Quail	38.9	55.7	1,108 ms	91%	62%	88%
llama-server	34.5	57.8	1,355 ms	91%	62%	88%
Ollama	37.4	60.0	1,140 ms	91%	63%	89%

Output tokens per second by requests at once, MLX lane

MLX lane	1 request, tok/s	8 at once, tok/s	First token (1 request)	GSM8K	MMLU-Pro	BFCL
Quail	44.9	49.8	1,439 ms	91%	61%	91%
oMLX	42.5	42.3	1,953 ms	89%	64%	91%
Rapid-MLX	45.9	54.7	1,434 ms	91%	62%	90%

Qwen3.6 35B-A3B

GGUF lane	1 request, tok/s	8 at once, tok/s	First token (1 request)	GSM8K	MMLU-Pro	BFCL
Quail	53.8	84.3	690 ms	96%	82%	92%
llama-server	54.0	85.6	667 ms	95%	80%	91%
Ollama	54.7	54.5	634 ms	95%	82%	90%

MLX lane	1 request, tok/s	8 at once, tok/s	First token (1 request)	GSM8K	MMLU-Pro	BFCL
Quail	59.8	88.0	884 ms	96%	85%	92%
oMLX	56.7	72.2	1,248 ms	94%	82%	91%
Rapid-MLX	63.7	101.3	870 ms	95%	79%	32%

Gemma 4 26B-A4B

GGUF lane	1 request, tok/s	8 at once, tok/s	First token (1 request)	GSM8K	MMLU-Pro	BFCL
Quail	52.3	82.9	692 ms	96%	86%	90%
llama-server	50.2	78.8	777 ms	95%	90%	90%
Ollama	50.3	81.3	726 ms	95%	91%	90%

MLX lane	1 request, tok/s	8 at once, tok/s	First token (1 request)	GSM8K	MMLU-Pro	BFCL
Quail	47.3	68.1	1,022 ms	91%	86%	90%
oMLX	45.2	61.0	1,395 ms	89%	85%	89%
Rapid-MLX	51.8	81.6	939 ms	90%	85%	20%

How it's measured

Speed: GuideLLM, at 1, 2, 4 and 8 requests at once, with prompts of exactly the same length for every engine. Each engine's own benchmark tool is run too, so a server isn't blamed for its engine.
Quality: EleutherAI's lm-evaluation-harness on GSM8K and MMLU-Pro, the same questions for every engine.
Tool calling: the Berkeley Function-Calling Leaderboard's own checker, which scores what an agent actually receives.
Fair settings: the same context and number of slots, a full-precision KV cache, temperature 0 and thinking off everywhere. A speed difference only counts if it's over 5% and bigger than the spread between runs.

The harness, its settings and every raw result are in the repository: bench/, with the reasoning in ADR-0063. You can run it yourself.