MLX models for LocalLM Lab

Which open-weight models to download for a 24 GB MacBook Air and a 64 GB Mac Studio, what we've tested, and what's next. Updated 2026-10-22. How local models work, and how we test them: Local AI Models on Apple Silicon.

Updated 2026-10-22. Results are for LocalLM Lab 2.1.0 GA on a 64 GB Mac Studio unless a row says otherwise; rows with times on the Air come from LocalLM Lab 2.0. 2.1.0 GA runs mlx-swift-lm at 30d755c (upstream main plus a fix of ours for Qwen3 tool calls inside an unclosed <think> block). Gemma 4 list arguments, gpt-oss tool calls and Muse Glimmer need 2.1.0 GA; on 2.0 they don't work. LocalLM Lab 2.2.0 GA runs mlx-swift-lm at dea880a: 30d755c plus Nemotron 3.5 support and nothing else, so the 2.1.0 GA results hold on 2.2.0 GA. Nemotron 3.5 needs 2.2.0 GA, and its rows were tested on it. Sizes are the checkpoint's weight files on Hugging Face as of October 2026, not the repo's usedStorage (which can count old revisions twice — Qwen3-14B-4bit reports 16.6 GB but downloads 8.3 GB).

Every result here is a point in time. Conversions, chat templates and the runtime all move, so check your own model with the SDK's MLXModelProvider.capabilityProbe before building on it.

Target Macs

• Comfortable is about half of memory: inference needs room for context and activations on top of the weights, and macOS needs its share.
• Tight loads, but long prompts or a second app can push it into a stall. LocalLM Lab warns before downloading anything over half of memory, and 70% is where the SDK's own check (MLXModelProvider.validate) flags a model as too big.
• The Air has no fan. A long generation slows down as it heats up, so judge its speed on short and medium prompts.
• A mixture-of-experts (MoE) model, marked MoE below, needs memory for all its weights but only computes a few billion parameters per token. It is much faster than a dense model of the same size, so on the Studio it is often the better trade.

Have a Mac with 128 GB or more? Help us test the bigger models. Several models the app's runtime supports are too big for our Macs, such as GLM 4.5 Air (60 GB), Mistral Small 4 119B (68 GB) and Mistral Medium 3.5 128B (78 GB); their repo ids are under Won't run in the app. Run them through the test protocol and send your results to neuron@thisbrain.ai or share them on our Discord. Tell us your Mac and its memory, and we'll add your results to this page with full attribution.

Our picks

Every pick is a tested row below. For a smaller grounding pick on the Air, Ministral-3-14B-Instruct-2512-mxfp4 (8.0 GB) passed every test twice.

How to read the tables

Tool use

• ✅ Reliable — called the right tools and grounded its answer in what came back.
• ⚠️ Partial — calls tools, but the wrong ones, inconsistently, or the call is never dispatched; or invents details the tools didn't return.
• ❌ No tools — loads and answers, but never calls a tool.
• 🚫 Won't load or Too big — the app's runtime can't load it, the repo is broken, or it's too big for either Mac.
• 🧪 Not tested yet — a candidate: the runtime supports its architecture and we've checked its size, but haven't run it.

Fit — ✅ comfortable, ⚠️ tight, ❌ too big, using the limits in Target Macs.

Format — the model_type from the checkpoint's config.json, and the tool-call parser the runtime picks for it. A model whose format has no parser can't call tools in the app, however good it is.

Qwen

The reference family: the MLX bridge was built against Qwen's tool-call convention, and the Qwen3 models call tools reliably at every size we tried. Qwen3.5 is less steady: see its rows.

Gemma

Mistral

Two things to know before picking a Mistral checkpoint:

• Most of mlx-community's December 2025 Ministral 3 conversions are broken. Their model.safetensors.index.json names weight files the repo doesn't have, so nothing can load them. As of October 2026, every quantized 8B and 14B Instruct repo and every quantized Reasoning repo is broken; every 3B Instruct repo is intact. For the 14B, use the later MXFP4, MXFP8 or NVFP4 conversions. There is no intact quantized 8B Instruct.
• Only the newer mistral3 models get Mistral's tool-call parser. Older checkpoints (model_type mistral: Mistral 7B, Nemo, Magistral, Codestral) get the default JSON parser, but their templates write Mistral's [TOOL_CALLS] syntax. Expect the gpt-oss result: a correct call that is never dispatched. Test before relying on one. (That's on LocalLM Lab 2.0's 45693f6. From 2.1.0 GA the runtime picks the Mistral parser from the template. But Mistral 7B v0.3 writes the older [TOOL_CALLS] [{…}] list form, which that parser rejects, so its calls still fail. For tools, use a Ministral 3 model.)

Granite

GLM

Zhipu AI's GLM models (MIT license). The GLM tool-call format has its own parser in the runtime, and LocalLM Lab 2.1.0 GA also reads GLM-4-0414's calls, which come without the usual markers. GLM 4.6V and 4.1V use a vision architecture (glm4v) the runtime doesn't support, GLM 5.2 and 5.3 use glm_moe_dsa (also unsupported, and over 400 GB), and GLM-OCR is a text-recognition model, not a chat model.

Other families

Can't use tools

We tested these and found they don't call tools in the app: they load and answer, but never use a connector or MCP server. Fine for plain chat. Other models with ❌ No tools are listed with their family above.

Won't run in the app

Speed helpers and adapters

Pairs that ran end to end on the Air in a release build. Speed depends on the Mac and the pair, so measure yours.

To try on the Studio: Qwen3-0.6B-4bit as a helper for Qwen3-8B-4bit and Qwen3-14B-4bit, and Qwen3.5-0.8B-4bit as a helper for Qwen3.8-27B-4bit (same qwen3_5 vocabulary — check before relying on it).

Test protocol

1. Fit and load. Add the model by repo id in LocalLM Lab's Add Model sheet. Note the preflight result and whether it loads. A load failure with a missing weight file is usually the repo, not the app: compare its model.safetensors.index.json with its file list.
2. Tool use.
   • Run the four tool tests below.
   • Record the tools called and the elapsed time.
   • Check every commit, release and fact the answer cites against the source.
   • Treat any invented detail as a finding, even in an otherwise good answer.
3. Speed. Run the Benchmark switch on the same pasted text for each model on a given Mac, and record time to first token and tokens per second from a warmed-up run.
4. Record. Keep the full trace, then update the model's row here. Tried a model that isn't listed? Send us your results at neuron@thisbrain.ai.

The four tool tests

Each test is a single prompt with only the named tools turned on. GitHub times on this page are the GitHub test, start to answer.

• Echo: the local test MCP server's echo tool, asked to echo a marker. Checks that a tool call gets through at all.
• GitHub: the GitHub MCP server with list_commits, list_releases and a few other read-only tools, and this prompt, which names no tool:

  Repository: humanizer. Owner: ancientcomputing. Use the GitHub MCP server. Look at the repository's recent commit history and its release history, and tell me: has this project been more active in development recently than its release cadence suggests, or vice versa? Support your answer with specific commits and release dates.

• HF: the Hugging Face MCP server's hub_repo_details, asked for the license and task tag of two model repos. The answers are exact, so they're easy to check.
• Context7: resolve-library-id then query-docs, asked for code from a library's documentation. The second call needs the ID the first returns, so it checks chaining.