Reading speed (prefill)
Measured in tok/s
How fast the model takes in everything you send: your prompt, files, an app’s instructions and the conversation so far. It sets how long big jobs wait before anything happens.
The Mac app for running local AI
Maximize local AI performance on your Mac
Run the right model at full speed. LMM shows which models and context sizes fit your Mac, and finds the bottleneck when something slows down.
For Macs with Apple silicon and macOS 14 or later
Overview · Live Unified Memory & Active Model Pool
M1–M4 Apple Silicon · Light & Dark Mode
Built for Apple siliconKnows how much memory your Mac can give to models
Works with your serversOllama, LM Studio, llama.cpp and MLX
No telemetryYour prompts and files never come to us
Nothing changes without youSafe fixes can be undone
Planner
Plan for the work you want to do
Tell LMM what you want to do. It compares the setups your Mac can run and recommends the best match for your work, balancing the task (coding, writing and more), memory use, quantization loss, unlocked models and the intelligence index.
Example requestExample request: coding with 2 agents on an M3 Max with 128 GB
Every setup gets a hardware fit rating
- Runs greatQuick replies with room to spare
- Runs wellComfortable for the task you chose
- WorkableUsable, with trade‑offs you can see
- FrustratingIt fits, but the waits will show
- Won’t fitNot enough room, with what would help
Planner · Hardware Fit & Multi-Agent Recommendations
Weights + KV cache calculated before loading
Models · Connected Local Pools & Live Catalog
oMLX · Ollama · LM Studio · llama.cpp
How to read the numbers
Speed is more than tokens per second
Most people judge local AI by one number, tokens per second. That only tells you how fast the answer is written. Before the first word appears, the model has to read everything you sent, and for a coding agent that can be 30,000 tokens or more.
You press EnterFirst word ≈33 sDone ≈54 s
Reading (prefill)30,000 tokens at ≈900 tok/s
Writing (decode)800 tokens at ≈38 tok/s
First word (time to first token)
Measured in seconds
The pause before the answer starts. It’s mostly reading time, so it grows with long prompts, big files and long conversations.
Writing speed (decode)
Measured in tok/s
How fast the answer appears once it starts. It’s the number most people quote, but once it outpaces your reading, the wait for the first word matters more.
TokenA short word or part of one. 1,000 tokens is roughly 750 English words.
Memory and context
Know what fits before you press Load
Every model needs memory for itself, and more for its context window, the limit most apps hide. Pick a context size and LMM shows what it costs and whether it fits, before anything loads.
Recommended for this MacQwen 3.8 27B with 64k context for coding agents, based on your memory and how you work
This MacM3 Max · 128 GB · GPU limit 96 GB
83GB free13 GB in use
Qwen 3.8 27B17 GB model, plus context
24 GBneeded
RecommendedLightning 30B19 GB model, plus context
22 GBneeded
FitsLlama 3.3 70B40 GB model, plus context
50 GBneeded
FitsPick a context size, then press Load. LMM shows the memory before anything loads.
ModelMemory to load it
ContextGrows with the context size
Apps and macOS13 GB
FreeRoom left for models
Near the limitThe last 10% of room
Example: M3 Max with 128 GB. Try the dropdowns and Load buttons.
Checkup
Find what to fix, and see why first
One scan looks at memory, running models, connected apps and server settings. Each finding explains itself before anything changes.
3 things to check
Frees 27 GB of memoryFixes 1 context problem
Performance & Checkup · Live Speedtest & Bottleneck Diagnostics
Decode tok/s · Prefill tok/s · First reply latency
Warning
Claude Code is close to its context limit
118kof 128k
Long sessions will start losing early files. A larger context for Qwen 3.8 27B fits in the room you have.
Suggestion
Lightning 30B has been idle for 2 hours
27GB of memory
Unload it to free memory now, and load it again when you need it.
Suggestion
A download you haven’t used in 30 days
40GB on disk
Llama 3.3 70B is taking up disk space. Keep it for later or remove it.
Setup help
Let a model set it up, with your approval
Connect a cloud model you choose and LMM can walk you through the setup: install a tool, download a model, change a setting, run a test prompt. Pick a low-cost model from a provider with zero data retention, and nothing changes until you approve it.
- 1Choose your modelUse your own provider account. LMM shows each provider’s data policy, where it came from and when it was checked.
- 2Review the planEvery step is written in plain words, with what it changes, before it runs.
- 3Approve and verifyLMM runs the steps you approve, then finishes with a test prompt.
Optional. LMM works without a cloud model, and nothing leaves your Mac for setup unless you connect one.
Assistant · Local Multi-Turn Chat, UI Inspect & Guided Tour
Works 100% locally or with optional zero-retention cloud profile
Local Agents · 1-Click Coding Triad & Agent Routing
Claude Code · Codex · Goose · Hermes · OpenClaw
Privacy
Your data stays yours
LMM runs on your Mac. Your prompts, conversations and files never come to us. A prompt goes only to the model server you choose, whether that’s this Mac, a paired headless instance or an endpoint you set up.
Connections · Local AI Servers, Gateways & Optional BYO Profiles
Zero prompt telemetry · All model traffic stays on your network
Every connection LMM makes
- Knowledge feedModel and benchmark updates, downloaded to your Mac
- LicenseActivation and renewal
- App updatesChecks for new versions of LMM
- Paired Macs & headless instancesDirectly over mutual TLS on your own network
Off until you turn it on
- Conversation detailsSizes and timings for each conversation, not message text
- Saving message textA separate setting, off by default
- Cloud setup helpOnly if you connect a cloud model, and only to the provider you choose
Headless Instances & Multi-Mac Nodes
Supports headless Mac instances
Deploy a headless Mac mini or Mac Studio instance running the lmm-host daemon over mutual TLS (mTLS). Monitor unified memory headroom, inspect active context windows, and load or unload models remotely from your primary Mac.
Mac(s) · Headless Instance Pairing & Multi-Mac Telemetry
Mutual TLS (mTLS) · 6-digit pairing · Bonjour & Tailscale
Works Better Together: LMM × PromptWorkshop
Sister Product: Supercharge Your Local Models with PromptWorkshop
PromptWorkshop optimizes prompts, system instructions, and tool definitions specifically for local models (Gemma 4, Qwen 3.6/3.8, Bonsai) — making a free 27B–35B local model perform like a paid frontier model while saving 30–50% of your KV cache tokens.
Model-Specific Prompt Optimization
Gemma 4 · Qwen 3.6/3.8 · Bonsai
Compiles crisp, unambiguous instructions tuned for local dense & MoE architectures (Gemma 4 31B, Qwen 3.6 35B-A3B, Bonsai 27B).
30–50% Context & KV Cache Savings
More headroom in unified memory
Compresses bloated system prompts and agent instructions so more agents fit simultaneously in your Mac’s unified memory.
Automated Local Eval & Audit
$0.00 token cost
Tests prompt variants directly against your local oMLX (:8003) or Ollama (:11434) endpoints at $0.00 token cost.
Native macOS App · Light & Dark Appearance
See Local Model Manager on macOS
Every workspace page shares a calm, unified Apple Human Interface design with an instant Light/Dark appearance toggle, real-time unified memory gauges, and 1-click progressive disclosure for advanced tools.
Overview & Checkup
Unified Memory · Loaded Models · System Snapshot
See available unified memory, loaded models, and actionable Checkup findings in two seconds.
Models & Live Worker Catalog
On This Mac · Discover · Remote Updates
Filter verified MLX, GGUF, and Ollama models by capability with instant M1–M4 hardware fit ratings.
Side-by-Side Test Workbench
1–3 Models · PromptWorkshop Harden · Speed Metrics
Compare up to three local or cloud models simultaneously with live prefill, first-token, and decode speed telemetry.
Hardware Fit Planner
M1–M4 Base / Pro / Max / Ultra · Context Sizing
Calculate exact weight + KV cache memory footprints for 1, 2, or 4 parallel coding agents before loading.
Local Agents & Coding Triad
Claude Code · Codex · Goose · Hermes · OpenClaw
Bind local coding agents to your fast local model server in one click with automatic rollback protection.
Performance & System Telemetry
Real-Time Throughput · Thermal & Swap Watch
Inspect live decode tok/s, prefill latency, and unified memory pressure across this Mac and paired headless instances.
Headless Instances & Multi-Mac Host
lmm-host Daemon · Mutual TLS · Bonjour & Tailscale
Pair headless Mac mini and Mac Studio instances over mutual TLS to monitor memory and manage models remotely.
Interactive Training & AI Dictionary
6 Video Modules · 129-Term Dictionary · Quick Guides
Built-in Apple Silicon local AI curriculum covering unified memory, quantization, KV cache sizing, and coding triads.
Free Trials & Pro-Rata Upgrade Ladder
3-Day Monthly Trial · 1-Week Annual & One-Time Trial
Try Monthly free for 3 days, or try Annual or One-Time Purchase free for a full week (7 days). Upgrade up the ladder anytime (Monthly → Annual → One-Time, or 1 → 2 → 5 Macs) and LMM automatically refunds the unused pro-rata balance of your lower purchase.
- Monthly: 3-Day Free TrialFull access for 3 days before your monthly billing starts
- Annual & One-Time: 1-Week Free TrialFull access for 7 days on Annual subscriptions and One-Time purchases
- Pro-Rata Upgrade LadderUpgrade anytime — unused lower-tier balance is refunded automatically pro rata
Launch offer20% off the first payment with code LAUNCH20
Each headless Mac instance counts as one Mac seat
Monthly · 3-Day Free Trial
$7.96/monthRegular price $9.95
Includes a 3-day free trial. Launch price with LAUNCH20 on first payment; upgrades to Annual or One-Time automatically refund unused days pro rata.
Annual · 1-Week Free Trial
$31.96/yearRegular price $39.95
Includes a 7-day (1-week) free trial. Launch price with LAUNCH20 on first payment; upgrades to One-Time automatically refund unused time pro rata.
One-time · 1-Week Free Trial
$95.96onceRegular price $119.95
Includes a 7-day (1-week) free trial. One payment, no renewal; includes all 1.x updates and automatic pro-rata credit when adding Mac seats.
Prices are in US dollars before applicable tax. LAUNCH20 discounts the first payment only. Monthly includes a 3-day free trial; Annual and One-time include a 1-week (7-day) free trial. Upgrading up the ladder automatically refunds the unused pro-rata balance of your lower purchase. Already purchased? Recover your license by email.
FAQ
Before you install
Straight answers about fit, privacy and your license.
It depends on the task, the model, the context size and what else is using memory. LMM weighs all of it and rates how well each setup fits your hardware. On a smaller Mac it tells you honestly when a setup is Frustrating or Won’t fit, and what would help.
Context is how much text a model keeps in mind at once: your conversation, files and an app’s instructions. A bigger context needs more memory, which is why a model that fits at 32k may not fit at 128k.
All Apple silicon Macs (M1 through M4, Base, Pro, Max, and Ultra) running macOS 14 or later, including dedicated headless Mac mini and Mac Studio instances running lmm-host. It works with Ollama, LM Studio, llama.cpp, and MLX servers.
No. LMM requires an Apple silicon Mac running macOS 14 or later. Every Apple silicon chip from M1 onward is supported, in the regular, Pro, Max and Ultra versions. Intel Macs aren’t supported.
LMM explains every finding before anything changes. Safe fixes can be undone, and everything else waits for your approval. Your prompts, conversations and files never come to us.
Monthly includes a 3-day free trial and renews monthly until you cancel. Annual includes a 1-week (7-day) free trial and costs about a third of paying monthly. One-time includes a 1-week (7-day) free trial, a single payment with no renewal, and all 1.x updates. You can upgrade up the ladder anytime (Monthly → Annual → One-Time, or 1 → 2 → 5 Macs) and LMM automatically refunds the unused pro-rata balance of your lower purchase.
Monthly subscriptions include a 3-day free trial, and Annual subscriptions and One-time purchases include a 1-week (7-day) free trial. All features are unlocked during your trial.
Each headless instance running LMM or the lmm-host daemon counts as one Mac seat in your plan, the same as a desktop or laptop. Choose a 2-Mac or 5-Mac plan under Pricing to cover multiple instances.
Yes, if you want it. Connect a cloud model you choose, ideally a low-cost one with zero data retention, and LMM writes out each step in plain words before anything runs. It’s optional, and LMM works without it.
When you upgrade up the ladder (Monthly → Annual → One-Time, or 1 → 2 → 5 Macs), our Stripe webhook automatically refunds the unused pro-rata balance of your lower purchase, cancels any superseded lower subscription, and migrates your active Mac seats. Every first-time purchase also includes a 30-day money-back guarantee.
See what your Mac can run
Download LMM for macOS and start with a 3-day Monthly trial or 1-week Annual / One-Time trial.