The best local LLM for a MacBook Air M2 (24 GB) is Qwen 3.6-35B-A3B at ~43 tok/s (estimated), the top-ranked model that reaches 15 tok/s here. Its 24 GB fits 63 of the 130 models in the LLMCheck index, up to 39.8B-class. The most capable, Qwen3.8-27B, runs at ~5 tok/s (estimated): slow for chat.
Considering this Mac?
MacBook Air M5 on Amazon → · compare every Mac for local AI →
Ranked by LLMCheck suitability (capability balanced against speed on the M2); the top pick is the best-ranked model that reaches 15 tok/s here, has a standard Mac runtime and fits the GPU memory limit (raised if need be), and slower models are tagged slow. Click a model for its full figures and setup. Every speed carries its provenance mark: most are index estimates from the memory-bandwidth model, a filled square is a published figure and a hollow circle a community run — submit a real run →
| # | Model | Size | License | Speed | 64K context | Capability |
|---|---|---|---|---|---|---|
| 1 | Qwen 3.6-35B-A3B raise GPU limit | 35B MoE | Apache 2.0 | estimated43 tok/s | 38/50 | |
| 2 | Qwen3.8-27B slow raise GPU limit | 27.8B | Apache 2.0 | estimated5 tok/s | 46/50 | |
| 3 | Nanbeige4.3-33B slow raise GPU limit (macOS 15) | 33.1B | Apache 2.0 | estimated4 tok/s | 46/50 | |
| 4 | Qwen 3.6-27B slow raise GPU limit | 27B | Apache 2.0 | estimated5 tok/s | 44/50 | |
| 5 | Recursive Flint raise GPU limit | 39.8B MoE | Apache 2.0 | estimated49 tok/s | 37/50 | |
| 6 | KAT-Coder-V2.5 over GPU limit | 35B MoE | Apache 2.0 | estimated31 tok/s | 39/50 | |
| 7 | Muse Glimmer 30B slow raise GPU limit | 30B | Apache 2.0 | estimated4 tok/s | agent-ready | 42/50 |
| 8 | LFM2.5-24B-A2B raise GPU limit (macOS 15) | 24B MoE | LFM Open | estimated52 tok/s | agent-ready | 35/50 |
| 9 | Gemma 4 31B slow raise GPU limit | 31B | Apache 2.0 | estimated5 tok/s | 40/50 | |
| 10 | MiniCPM-V 4.7 35B-A3B over GPU limit | 35.2B MoE | Apache 2.0 | estimated38 tok/s | 35/50 | |
| 11 | Nemotron 3.5 Lightning raise GPU limit | 30B MoE | OpenMDW | estimated29 tok/s | agent-ready | 36/50 |
| 12 | Gemma 4 26B-A4B raise GPU limit | 26B MoE | Apache 2.0 | estimated34 tok/s | agent-ready | 35/50 |
Showing the top 12 of 63 models that fit in 24 GB. See the full leaderboard or all benchmarks.
Context: “fits” means the weights at a short prompt. Coding agents (Claude Code, OpenCode, Codex CLI) hold 64K tokens or more, and every token adds KV cache: agent-ready marks a model whose weights plus a 64K-token 16-bit KV cache still fit in the 20 GB this Mac’s GPU can use with the limit raised (33 of the 58 models here with a KV figure). That is an estimate from each model’s config.json, not a measurement; a q8 KV cache halves the cache part, and — means the index has no KV figure for the model yet. The GPU memory limit below applies on top.
GPU memory limit: macOS lets the GPU use about 16 GB of this Mac’s 24 GB by default on macOS 15 and earlier, and about 17.8 GB on macOS 26 and later. Of the 63 models that fit, 25 need more than that at a short prompt: 17 need a raised limit and are tagged raise GPU limit (Qwen 3.6-35B-A3B and Gemma 4 31B only past the safe setting below, as measured runs on this much memory show); 8 need more than even the raised limit and are tagged over GPU limit. sudo sysctl iogpu.wired_limit_mb=20480 raises it to 20 GB, leaving 4 GB for macOS, until the Mac restarts (sudo sysctl iogpu.wired_limit_mb=0 restores the default). How the limit is read →
GPU memory limit: Qwen 3.6-35B-A3B needs about 24 GB, more than macOS lets the GPU use on this Mac by default (about 16 GB)., past the index’s safe setting (sudo sysctl iogpu.wired_limit_mb=20480, 20 GB), so macOS keeps less than 4 GB: quit other apps first.
The fastest way to get started is Ollama. Install it, then pull the model by its tag (checked on ollama.com, 2026-11-14):
ollama run qwen3.6:35b-a3b · model card →
Prefer a GUI? LM Studio gives you a one-click download and chat window. For step-by-step help see our Ollama install guide, or open the Qwen 3.6-35B-A3B on M2 page for its figures and setup.
Twice as fast: the cheapest Mac in the index that runs Qwen 3.6-35B-A3B at 2× or more of this Mac’s estimated~43 tok/s is the Mac Studio M2 Max (32 GB) at estimated~92 tok/s, about $1,300 refurbished or used: search Amazon Renewed listings →. New: the Mac Studio M5 Max (36 GB) at estimated~100 tok/s, $2,499 new (Apple list price, checked 2026-11-14): Mac Studio M5 Max (36GB) on Amazon →.
For less: no cheaper Mac in the index runs Qwen 3.6-35B-A3B at 15+ tok/s.
Speeds are each Mac’s own figure from the index, with its provenance mark. Prices come from data/macs.json: Apple’s list price for a new Mac (dated), an approximate street price for a used or refurbished one. Amazon stock checked 2026-11-18. Amazon links are affiliate links: as an Amazon Associate, LLMCheck earns from qualifying purchases.
Apple no longer sells the MacBook Air M2 (24 GB) new. A current MacBook Air M5 (24 GB) runs the same models with 153 GB/s of memory bandwidth instead of 100 GB/s, so they generate faster. Amazon stock of the 24 GB configuration is not verified: check the listing’s memory.
As an Amazon Associate, LLMCheck earns from qualifying purchases. Affiliate links cost you nothing extra and never influence our rankings.
LLMCheck. (2026). Best Local LLMs for the MacBook Air M2 (24 GB). llmcheck.net. https://llmcheck.net/best-llm/macbook-air-m2-24gb/ (updated ). CC BY 4.0.
Qwen 3.6-35B-A3B (35B MoE, Apache 2.0) is the best all-round pick at ~43 tok/s (estimated) on the MacBook Air M2. If you want maximum speed, Ling-3.0-tiny hits ~106 tok/s (estimated); Qwen3.8-27B is more capable but runs at only ~5 tok/s (estimated) here.
About 63 of the 130 models in the LLMCheck index fit in 24 GB of unified memory, from compact models up to Recursive Flint (39.8B MoE).
No. A dense 70B such as Llama 3.3 70B needs about 44 GB at 4-bit; with 24 GB, stick to models like Qwen 3.6-35B-A3B. For a 70B, look at a 64 GB+ Mac.
24 GB is enough for dense models up to about 33B at 4-bit (the largest in the index that fits is Nanbeige4.3-33B), and for mixture-of-experts models such as KAT-Coder-V2.5 (~31 tok/s (estimated)). As a rule of thumb a 4-bit model needs about 0.57 GB per billion parameters plus ~2 GB, and macOS needs its own share of memory too. Because Apple Silicon uses unified memory, the same pool serves as system RAM and GPU memory.