Hardware Updated October 24, 2026 · 10 min read

Apple Silicon for AI: M1 to M6 Chips Compared (2026)

According to the LLMCheck index, decode speed tracks memory bandwidth: the M5 Ultra (1,228 GB/s) is the fastest Apple Silicon for local LLMs, the M5 Max (614 GB/s) the fastest laptop chip, and the Mac mini M5 Pro (307 GB/s, up to 64 GB) the best value. An M5 Max has six times the bandwidth of a base M3.

Correction, 5 October 2026. Earlier versions of this page gave the base M3’s bandwidth as ~200 GB/s and the M5 Max’s as ~600 (Apple: 100 and 614), claimed the M5 Max was about 3x faster than the M1 Max (decode is about 1.5x — the bandwidth ratio), and printed a 70B speed for the M5 Max above that chip’s memory-bandwidth ceiling. The table now reads its figures from the LLMCheck index, with provenance.

Updated 24 October 2026. Community-measured runs on the M5 Pro replaced its estimates: the index now shows Qwen 3.6-35B-A3B at 79 tok/s there (it had estimated about 92), and LFM2.5-24B-A2B at 114, which makes LFM2.5-24B-A2B the index’s pick for every Mac mini M5 Pro and for 64 GB Macs. Qwen 3.6-35B-A3B is the index’s #1 and still its pick for 24 and 32 GB Macs.

Updated 18 October 2026. The budget table now uses Apple’s launch prices and the index’s pick for each Mac, with the Mac mini M5 Pro as the value pick; the generation table’s decode gains now equal the bandwidth ratios (it had credited the M4 Max to M5 Max step with +25%); the memory table no longer lists a 405B model on 128 GB, which does not fit; and Apple’s own M5 Ultra vs M3 Ultra measurements are added.

The Complete Apple Silicon AI Benchmark Table

According to the LLMCheck index, this table shows single-stream decode speed across every major Apple Silicon variant, at 4-bit (Q4_K_M), for three reference models. Figures are estimates from the published bandwidth model unless marked as measured; LLMCheck does not run its own tests.

Chip	GPU cores	Bandwidth	Max RAM	Qwen 3.5 9B	Qwen3.8-27B	Llama 3.3 70B
M6 Generation (2026)
M6	12	153–170 GB/s	32 GB	estimated~25 tok/s	community9 tok/s	N/A
M5 Generation (2025–2026)
M5 Ultra	64–80	1,228 GB/s	512 GB	community167 tok/s	sourced48 tok/s	estimated~23 tok/s
M5 Max	32–40	614 GB/s	128 GB	estimated~82 tok/s	estimated~29 tok/s	sourced13 tok/s
M5 Pro	16–20	307 GB/s	64 GB	estimated~44 tok/s	sourced17 tok/s	estimated~6 tok/s
M5	10	153 GB/s	32 GB	estimated~23 tok/s	estimated~8 tok/s	N/A
M4 Generation (2024–2025)
M4 Max	32–40	410–546 GB/s	128 GB	estimated~74 tok/s	estimated~26 tok/s	estimated~12 tok/s
M4 Pro	16–20	273 GB/s	64 GB	estimated~39 tok/s	estimated~13 tok/s	estimated~5 tok/s
M4	10	120 GB/s	32 GB	estimated~18 tok/s	estimated~7 tok/s	N/A
M3 Generation (2023–2025)
M3 Ultra	60–80	819 GB/s	512 GB	estimated~105 tok/s	community38 tok/s	estimated~16 tok/s
M3 Max	30–40	300–400 GB/s	128 GB	estimated~56 tok/s	estimated~19 tok/s	estimated~8 tok/s
M3 Pro	14–18	150 GB/s	36 GB	estimated~22 tok/s	estimated~8 tok/s	N/A
M3	8–10	100 GB/s	24 GB	estimated~15 tok/s	estimated~5 tok/s	N/A
M2 Generation (2022–2023)
M2 Ultra	60–76	800 GB/s	192 GB	estimated~103 tok/s	estimated~38 tok/s	estimated~15 tok/s
M2 Max	30–38	400 GB/s	96 GB	estimated~56 tok/s	estimated~19 tok/s	estimated~8 tok/s
M2 Pro	16–19	200 GB/s	32 GB	estimated~29 tok/s	estimated~11 tok/s	N/A
M2	8–10	100 GB/s	24 GB	estimated~15 tok/s	estimated~5 tok/s	N/A
M1 Generation (2020–2022)
M1 Ultra	48–64	800 GB/s	128 GB	estimated~103 tok/s	estimated~38 tok/s	estimated~15 tok/s
M1 Max	24–32	400 GB/s	64 GB	estimated~56 tok/s	estimated~19 tok/s	estimated~8 tok/s
M1 Pro	14–16	200 GB/s	32 GB	estimated~29 tok/s	estimated~11 tok/s	N/A
M1	7–8	68 GB/s	16 GB	estimated~10 tok/s	N/A	N/A

N/A = the model’s smallest Mac build does not fit the chip’s largest memory option. The index’s minimum RAM for a model is its smallest build of at least 3 bits with a working Mac runtime, plus about 2 GB; a 4-bit 70B needs 44 GB, a 4-bit Qwen3.8-27B 19 GB.

The Key Insight: Memory Bandwidth Determines Speed

According to the LLMCheck index, memory bandwidth predicts decode speed better than GPU core count, Neural Engine TOPS or chip generation. Here’s why:

During token generation, the model’s active weights are read from memory once for every token. A 9B model at 4-bit is about 5.1 GB read per token. At about 82 tok/s — the index’s estimate for an M5 Max — that is about 420 GB/s of sustained reads, roughly 70% of the chip’s 614 GB/s. This is why bandwidth is the bottleneck, not compute.

For large dense models, doubling the bandwidth roughly doubles tok/s. Small models gain a little less, because a fixed per-token cost takes a larger share of each token’s time, and small mixture-of-experts models less again (methodology). Compute matters for the other phase, prompt processing, which is where the M5 generation’s Neural Accelerators pay off.

Best Value Recommendations by Budget

Starting prices are Apple’s US launch prices; the pick is the index’s best model for that memory (by RAM tier), with its speed on that Mac.

Mac	Starts at	RAM / bandwidth	Index pick	Speed on that Mac
Mac mini M6	$899 (16 GB)	16 GB / 153 GB/s	Gemma 4 12B	estimated~17 tok/s
Mac mini M6, 32 GB	configure to order	32 GB / 170 GB/s	Qwen 3.6-35B-A3B	community64 tok/s
Mac mini M5 Pro	$1,699 (24 GB)	24–64 GB / 307 GB/s	LFM2.5-24B-A2B	community114 tok/s
Mac Studio M5 Max	$2,499 (36 GB, 460 GB/s)	36–128 GB / 460–614 GB/s	Qwen3.8-Flash-Next at 128 GB	community46 tok/s
MacBook Pro M5 Max	$3,599 (14-inch)	36–128 GB / 460–614 GB/s	Qwen 3.6-35B-A3B	sourced118 tok/s (40-core GPU)
Mac Studio M5 Ultra	$5,499 (96 GB)	96–512 GB / 1,228 GB/s	MiMo-V2.6-Flash at 256 GB+	community62 tok/s

Prices: Apple’s announcements for the Mac Studio (25 August 2026) and MacBook Pro (3 March 2026); Mac mini prices as announced on 25 August 2026. The cheapest M5 Max in both the Mac Studio and the MacBook Pro has the 32-core GPU at 460 GB/s; 614 GB/s needs the 40-core GPU.

LLMCheck’s pick for most buyers: the Mac mini M5 Pro. It has 1.8× the bandwidth of the Mac mini M6 (307 against 170 GB/s) and runs the index’s #1, Qwen 3.6-35B-A3B, at community79 tok/s, and the index’s pick for that Mac, the faster LFM2.5-24B-A2B, at community114 tok/s. Take 48 or 64 GB if you can: 24 GB is Qwen 3.6-35B-A3B’s minimum, and more memory adds room for long contexts and the 80–120B mixture-of-experts models.

Generation-Over-Generation Improvement

Because decode follows bandwidth, the decode gain of an upgrade is close to its bandwidth ratio. Bandwidths are Apple’s figures for the full version of each chip.

Upgrade path	Bandwidth	Decode gain	Worth it for local LLMs?
M1 Max → M2 Max	400 → 400 GB/s	about 0%	No
M2 Max → M3 Max	400 → 400 GB/s	about 0% (the 30-core M3 Max has 300 GB/s)	No
M3 Max → M4 Max	400 → 546 GB/s	about +37%	Yes
M4 Max → M5 Max	546 → 614 GB/s	about +12%	For long prompts: Apple quotes up to 4× faster prompt processing
M4 Pro → M5 Pro	273 → 307 GB/s	about +12%	For long prompts: Apple quotes up to 4×
M1 Max → M5 Max	400 → 614 GB/s	about +54%	Yes
M1 → M5	68 → 153 GB/s	about +125%	Yes
M4 → M6	120 → 153–170 GB/s	about +28–42%	Yes, at 24 GB or more
M3 Ultra → M5 Ultra	819 → 1,228 GB/s	about +50% (Apple measured 1.58×)	Yes, and time to first token is 4.4× faster

Apple’s own measurement: M5 Ultra vs M3 Ultra

Apple Machine Learning Research published single-Mac figures for both Ultra chips on 15 October 2026 (“Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio”). With Qwen3.8-27B in BF16 and an 8,192-token prompt (mlx_lm.benchmark -p 8192 -g 128 -b 1), the M5 Ultra decoded at 19.69 tok/s against 12.46 on the M3 Ultra, 1.58× faster and close to the 1.5× bandwidth ratio, while time to first token fell from 18.2 s to 4.1 s, 4.4× faster. That is the M5 generation in one table: decode tracks bandwidth, prompt processing jumps. Both figures are published (vendor) rows in the index.

A rule of thumb from the index: an upgrade that adds less than about 20% bandwidth is felt mainly in prompt processing (on M5-generation chips) or through extra memory, not in chat speed. For the M5 Max decision in detail, see M5 Max vs M4 Max.

Which Models Run on Which Chips?

Memory decides what fits; the chip decides how fast. The index’s picks for each memory tier as of 24 October 2026 (the live list is Best LLM by Mac RAM, and how much RAM you need explains the tiers):

Your RAM	Index pick	Also fits	Chips that offer it
8 GB	Nanbeige4.2-3B	Gemma 4 E4B, Phi-4 Mini	M1–M3 base
16 GB	Gemma 4 12B	GPT-oss 20B (tight)	M1–M6 base
24–32 GB	Qwen 3.6-35B-A3B	Qwen3.8-27B, Gemma 4 26B-A4B	M2–M6 base, Pro
48–64 GB	LFM2.5-24B-A2B	Qwen 3.6-35B-A3B, Qwen3-Coder-Next, GPT-oss 120B (64 GB), Llama 3.3 70B	Pro, Max
128 GB	Qwen3.8-Flash-Next	Qwen 3.5 122B-A10B, DeepSeek V4 Flash at 2-bit	M3, M4, M5 Max; Ultra
192–256 GB	MiMo-V2.6-Flash	GLM 5.2 (256 GB)	M2, M3, M5 Ultra
512 GB	MiMo-V2.6-Flash	1T-parameter models at about 3-bit (Kimi K2.6, Kimi K2.7 Code)	M3, M5 Ultra

LLMCheck Research Team

LLMCheck maintains the open LLMCheck index of local-LLM performance on Apple Silicon — transparent estimates from a published model, sourced third-party benchmarks, and community submissions, with every figure carrying a provenance label. Full data is open (CC BY 4.0) at llmcheck.net/data.

Frequently Asked Questions

Which Apple Silicon chip is best for running local AI?

For speed and memory, the M5 Ultra (1,228 GB/s, up to 512 GB, Mac Studio only); for a laptop, the M5 Max (614 GB/s, up to 128 GB). According to the LLMCheck index, Qwen3.8-27B runs at about 29 tok/s estimated on an M5 Max and 48 tok/s published on an M5 Ultra. For value, the Mac mini M5 Pro (307 GB/s, from $1,699).

How much faster is the M5 than the M1 for AI?

According to the LLMCheck index, the M5 Max decodes about 1.5x faster than the M1 Max on the same model, 614 GB/s against 400, and the base M5 a little over 2x faster than the base M1 (153 against 68 GB/s). Prompt processing gains are much larger, from the Neural Accelerators in every M5 GPU core.

Is an M5 Pro or M4 Pro good enough for local AI?

Yes, especially with 48 or 64 GB. According to the LLMCheck index, an M5 Pro runs Qwen 3.6-35B-A3B at 79 tok/s (community-measured) and an M4 Pro at about 74 (estimated), and the dense Qwen3.8-27B at 17 (published) and about 13 (estimated). A 64 GB Pro also fits Qwen3-Coder-Next and GPT-oss 120B.

Why does memory bandwidth matter more than GPU cores?

Because every generated token reads the model's active weights from memory once, so decode speed is capped by memory bandwidth divided by the bytes read. For large models, doubling bandwidth roughly doubles tokens per second; small models gain a little less. GPU compute matters more for prompt processing.

Can I run 70B parameter models on a MacBook Air?

No. A 70B model at 4-bit is about 40 GB of weights, and the LLMCheck index's minimum for it is 44 GB, more than any MacBook Air offers. You need a Mac with 48 GB or more: an M4 or M5 Pro or Max with 48 to 64 GB, or a Mac Studio. On an M5 Max it runs at about 13 tok/s (tbreak's published figure).

What is the best budget Mac for AI in 2026?

The Mac mini M6 at $899 (16 GB) runs Gemma 4 12B at about 17 tok/s estimated. For the best value, the Mac mini M5 Pro from $1,699 (24 GB, 307 GB/s) runs Qwen 3.6-35B-A3B, the LLMCheck index's #1, at 79 tok/s community-measured, and LFM2.5-24B-A2B, the index's pick for that Mac, at 114 tok/s community-measured.

Is the M2 Ultra still worth it for AI in 2026?

Used and well priced, yes: its 800 GB/s and up to 192 GB still hold MiMo-V2.6-Flash, which needs about 170 GB, at about 54 tok/s estimated on a 192 GB M2 Ultra. New, the M3 Ultra and M5 Ultra go to 512 GB, and the M5 Ultra has about 1.5x the M2 Ultra's bandwidth.

Find the Right Model for Your Chip

Filter LLMCheck's leaderboard by your Mac's RAM to see exactly what runs on your hardware.

View Leaderboard →