Your Mac. Your MLX models. One measured workflow.

Melix turns fragmented local AI work into a single, repeatable studio on Apple Silicon.

Serve locally. Tune with LoRA. Benchmark with rigor. Evaluate and export. No cloud. No subscriptions. No data leaves your Mac.

Get Started View Benchmarks

Native macOS app + CLI
Apple Silicon optimized
100% local
No subscriptions

Local-first MLX serving

Run MLX models entirely on device with project-owned endpoints and artifacts.

LoRA fine-tuning

Parameter-efficient tuning with repeatable local runs and adapter records.

Benchmark comparison

Track performance across MLX models, configs, datasets, and hardware.

Evaluation artifacts

Generate reports and exports that make local model changes auditable.

Melix Studio

MLX Model Registry

Search models

Model	Type	Size	Updated
mlx-community/Mistral-7B-Instruct-v0.3	MLX	4.2 GB	Jun 8
mlx-community/Llama-3.2-3B-Instruct	MLX	2.4 GB	Jun 7
mlx-community/Qwen2.5-Coder-7B-Instruct	MLX	4.7 GB	Jun 6
support-tone-lora-r16	LoRA	184 MB	Jun 5
mlx-community/Phi-3.5-mini-instruct	MLX	2.2 GB	Jun 4

Active Local Server

Running

Model: Mistral-7B-Instruct-v0.3
Runtime: MLX / Metal
Endpoint: localhost:11434/v1
Context: 4096
Requests: 1,248

Explain rotary positional embeddings in simple terms.

Rotary positional embeddings help attention compare token positions by rotating vector pairs, preserving relative position signals as sequences get longer.

Benchmarks / Evaluation

Model	Score	95% CI
Mistral-7B-Instruct-v0.3	8.48	+/- 0.12
Qwen2.5-Coder-7B-Instruct	8.17	+/- 0.15
Llama-3.2-3B-Instruct	7.83	+/- 0.13
Phi-3.5-mini-instruct	6.91	+/- 0.14

View Report

$ melix models add --source mlx Mistral-7B-Instruct-v0.3 Model registry
$ melix serve --runtime mlx --model Mistral-7B-Instruct-v0.3 Local server
$ melix lora train --config tone.yml LoRA training
$ melix benchmark run --suite mt-bench Benchmark
$ melix eval export --format report.md,json Evaluation export

why local

Local execution changes the development loop.

Local execution by default

Run model serving, adapter tuning, benchmarks, and evaluations on your Mac. Keep prompts, adapters, benchmark outputs, and eval artifacts in a local workflow.

Fast iteration

Serve a model, test an adapter, run an evaluation, and compare results without waiting on hosted inference queues or remote experiment infrastructure.

Reproducible workflow state

Track the model, target hardware, adapter, benchmark settings, eval set, and output metrics behind each run.

workflow ledger

One local workflow from model to measurement.

Serve the model, attach the adapter, benchmark the run, evaluate the output, and inspect the evidence locally.

01 / serve

Expose MLX models as working local endpoints.

Melix keeps model configuration, native Apple Silicon target, memory profile, and localhost routing in one place so app loops can call the same local MLX runtime repeatedly.

endpointlocalhost:8080

modelMistral-7B-Instruct-v0.3

targetMLX / Metal

$ melix serve --runtime mlx --model Mistral-7B-Instruct-v0.3

02 / benchmark

Compare throughput, latency, and memory pressure.

Run local benchmark profiles across MLX models, adapters, and settings before deciding which loop is actually better on the Mac in front of you.

tokens/sec112.6

p95 latency112 ms

memory18.4 GB

$ melix benchmark run --suite local-loop --compare baseline

03 / lora

Track adapters as local project artifacts.

LoRA runs stay tied to dataset path, rank, checkpoint, and serve-time comparison instead of becoming loose files scattered across terminals and notebooks.

adaptersupport-tone-lora-r16

rank16

checkpoint4200

$ melix lora train --dataset ./data/support --rank 16 --checkpoint local

04 / evaluate

Score outputs before the workflow moves.

Evaluation lives beside the runtime: offline eval sets, regression deltas, and output notes can be compared before a prompt, model, or adapter change ships.

eval setregression-core

pass rate92.4%

delta+7.4%

$ melix eval export --format report.md,json

technical surface

Built like a datasheet, not a promise deck.

Runtime

MLX model contexts
Metal-targeted local execution
localhost endpoints for app loops

LoRA

adapter rank and checkpoint ledger
dataset locality by project path
serve-time adapter comparison

Measurement

tokens/sec and p95 latency
memory and run deltas
offline eval-set regression checks

See the proof

Independent benchmarks and real results from local Apple Silicon hardware.

View Benchmarks

MT-Bench: 8.48 Mistral-7B-Instruct-v0.3 on Mac Studio
Arena Hard: 58.3% win rate for top local run
Tokens / sec: 112.6 at 4k context
Runs: 247 across 12 datasets and 6 models

Built for local MLX operators

Choose the local loop you run most often.

PATH 01 / LOCAL APPS OK

Build local apps

Wire local MLX serving into app loops.

models add -> serve -> chat

local endpoint, request log, reproducible run state

PATH 02 / MODEL WORK OK

Tune and compare models

Train LoRA and compare adapters.

lora train -> benchmark run -> compare

score deltas, latency, adapter artifacts

PATH 03 / PRIVATE EVALS OK

Evaluate private research

Test prompts and models without remote services.

eval run -> export report

markdown/json exports, local dataset path, no data leaves Mac

View benchmarks Read docs

Docs

Everything you need in one local studio

100% Local

Roadmap

A native MLX studio for the full Apple Silicon loop.

Melix is moving toward richer runtime operations, stronger evaluation workflows, better packaging, and more native operator polish.

status

Alpha software for local-first builders.

Melix is early, explicit, and local-first. The product is built around a concrete developer workflow: serve the model on your Mac, tune or attach an adapter, benchmark the run, evaluate the output, and inspect the evidence locally.

Inspect the repository