Use open models for volume. Keep Claude and ChatGPT for the work where mistakes, taste, and deep reasoning cost more than tokens.
On this page: Short answer · Tiers · Computers · Models · Fleet · Phones · Your work · Products · Plan · Sources
Your 48 GB M4 Pro is capable, but it is also your main production computer and its disk is already tight. One carefully chosen local model makes sense. A library of giant models does not.
Devstral Small 2 on the Mac for local coding and repo work. Gemma 3 12B or Qwen 3.5 9B as an optional lighter general model. Kaggle for scheduled GPU batches. Ollama Cloud or a shut-down-after-use Runpod for larger open models. Claude and ChatGPT for architecture, hard debugging, major creative direction, high-stakes research, and final review.
This is the honest quality ladder. Move down the ladder for volume and privacy. Move up when the cost of a weak answer is higher than the model bill.
Claude and ChatGPT remain your senior partners.
Near-frontier breadth at much lower unit cost.
Your M4 Pro handles a focused 12B to 24B model well.
Fast, cheap, narrow, and useful across the older fleet.
The model matters less than putting it on the right computer. Free notebooks are excellent batch workers, but they are poor always-on assistants.
Best for private, interactive work. No per-message cost. Always available. Keep only one heavyweight model loaded.
Your best free burst worker. Officially offers a free P100, roughly 30 GPU hours weekly, and sessions up to 12 hours.
A useful backup notebook. Free GPU access is not guaranteed and changes over time.
Use Ollama Cloud for simplicity or Runpod when you need your own GPU. Shut rented machines down as soon as the job ends.
You do not need every release. These are the useful families for your mix of development, design, research, content, documents, and media.
The strongest practical coding-agent choice for this Mac. Built for exploring repositories, editing files, and using tools.
Google's clean general-purpose option. It can read images and documents, and the 12B version is light enough to remain comfortable.
A quick multilingual generalist with vision and tool support. Better for speed than for final authority.
OpenAI's open-weight reasoning model. Useful when thinking depth matters more than writing style.
Stronger general quality, but they compete more noticeably with your browser, builds, and cloud-agent sessions.
The open-model ceiling. Rent them by token or by hour instead of forcing them onto local hardware.
Small enough for free notebook GPUs and strong enough for repetitive, checkable work.
It runs, but the live test used 23.5 GB of memory and pushed this already-busy Mac toward full swap at 64K context.
The Intel machines are not useless. They can run small quantized models for background work, serve private endpoints, index your material, run checks, and keep queues moving while the M4 Pro stays responsive. Expect slower answers, not magic GPU speed.
M4 Pro, 48 GB unified memory. Already runs a 35B Qwen model through llama.cpp and has a separate 9B vision model on disk. Use the 35B for focused local work; keep one large model loaded at a time. The machine has about 98 GiB free, so no model library expansion yet.
32 GB RAM, Intel i5-6600, Radeon graphics, 71 GB free. A realistic candidate for 3B-7B Q4 models through llama.cpp, with background-speed expectations. Also a strong index, test, and overnight queue worker.
16 GB RAM, Xeon E5, 418 GB free, and older Radeon cards. Try 1B-3B Q4 CPU models for tags, extraction, and summaries. Better still for archives, document preparation, containers, and queued batch work than interactive chat.
Intel i5, 16 GB RAM, 392 GB free, with about 7 GB RAM currently available. Aim at 1B-3B Q4 models and short tasks; reserve most memory for the system and other services.
Intel i5, 16 GB RAM, macOS 26, 748 GB free, Metal 3. Try llama.cpp CPU inference with 3B-7B Q4; 9B may work for a focused test but will be slower. Great as a private API endpoint, model mirror, and batch worker.
Intel i5, 4 GB RAM, 126 GB free. Not a chat host, but useful for a lightweight dispatcher, health checks, scheduled wakeups, and perhaps a 0.5B-1B model for very short classification.
Online with SMB storage available, but Remote Login is pending. Treat it as storage until its current processor, memory, and GPU can be verified.
Live check now succeeds: Intel i5 3.3 GHz, 4 cores, 32 GB RAM, macOS 12.7.6, Radeon 2 GB. Try CPU-only llama.cpp with 3B-7B Q4; verify available disk before downloading. Older macOS may limit the easiest app choices.
I could not identify a connected phone model in this check, so these are safe capability bands rather than a claim about your exact device. Your phone can still be the pocket doorway to the M4 model even when it is too small to run one itself.
Use it as a camera, voice recorder, scanner, and chat screen for a model running on your Mac. A tiny 0.8B-1B offline model is an experiment, not a Claude substitute.
Try 1B-3B Q4 for rewriting, short summaries, note cleanup, and extraction. Expect short context and slower responses than a cloud app.
3B-4B is the sensible first try. A 7B quantized model may fit on some phones, but heat, battery drain, speed, and app support decide whether it is pleasant.
Best fit for your day: use the phone as the interface and send work over Tailscale to the Mac's private model endpoint. Keep a small on-device model only for airplane mode or quick private transformations. If your iPhone supports Apple's Foundation Models, that is another no-download option for bounded writing, extraction, summaries, and structured tasks; it is separate from Ollama and cannot load your own model files.
Your history is dominated by app building, design, debugging, deployment, research, content, documents, media, and voice. This is where the savings come from.
Local model writes routine components, tests, docs, and first-pass refactors. Cloud frontier designs architecture and resolves hard cross-system failures.
LOCAL FIRST, FRONTIER REVIEWLocal model organizes content and checks consistency. Claude or ChatGPT owns concept, taste, visual critique, and final copy.
FRONTIER LEDFleet machines can run checks and builds. A frontier model should review risky changes, live incidents, auth, billing, and data migrations.
AUTOMATE, THEN ESCALATELocal models summarize supplied material. Cloud models with web access verify current facts, laws, pricing, product details, and sources.
LOCAL DIGEST, CLOUD VERIFYLocal handles variants, cleanup, formatting, and extraction. Frontier cloud handles positioning, final narrative, and sensitive client-facing work.
MOSTLY LOCALKaggle remains the best free batch lane for ComfyUI-style jobs, Kokoro, transcription, embeddings, and repeatable media processing.
KAGGLE FIRSTLocal models earn their keep when they take bounded, testable jobs inside products you already maintain. They should not become a new product that needs constant care.
The private vault and execution layer should use local inference only where a model adds real value. Routine indexing stays deterministic.
Milo owns the plan and phone-first daily guidance. A small model can prepare options without becoming the final decision-maker.
Its devotional content, Soul Audit, subscriptions, and user data need a clear line between assistance and spiritual authority.
High-volume agency work is the clearest cost-saving opportunity because most transformations are repeatable and reviewable.
Genealogy rewards careful extraction but punishes confident invention. Local AI should assist the evidence workflow, never become the source.
The app already has tests, backtests, deterministic projections, and two clear weekly decisions. Keep the math as code.
The canvas game has a small testable codebase. A coding model can handle contained implementation work if visual quality is reviewed separately.
The glasses SDK changes and hardware constraints make current documentation more important than model size.
Kokoro and Chatterbox already prove the right pattern: notebooks for burst compute, local APIs only when responsiveness matters.
Start lean. Prove savings for a month before adding more models or another paid service.
Model catalogs and prices change quickly. These links are the live sources behind the recommendation.
Fleet facts were measured live through the canonical Tailscale aliases on September 27, 2026. jamess-imac-pro remains pending shell access; phone make/model and memory were not identified, so phone guidance is conditional. Disk capacity and model speed are device-specific; validate before downloading or assigning always-on work. Local model quality claims are recommendations, not a guarantee of Claude-level judgment.