Hussain-Nazary

Apple Silicon (Macs — the local-LLM favorite)

Apple’s CPU/GPU/Neural Engine share one pool of unified memory, so a Mac can hold an entire large model in RAM — no PCIe transfer bottleneck. That, plus high memory bandwidth and the excellent MLX framework, makes Macs the most popular hardware for serious local inference. The deciding spec is unified memory (how big a model fits) and bandwidth (how fast it decodes). All model picks lead with the newest releases (Aug 2026).

M4 family (2024–2025)

Spec	M4	M4 Pro	M4 Max
CPU	10-core	14-core (12/14 in late refresh)	16-core
GPU	10-core	20-core	40-core
Neural Engine	16-core	16-core	16-core
Memory	up to 32 GB	up to 64 GB	up to 128 GB
Bandwidth	~120 GB/s	~273 GB/s	~546 GB/s
Devices	MacBook Pro 14, iMac, Mac mini, iPad Pro	MacBook Pro, Mac mini	MacBook Pro, Mac Studio

Best for: Excellent local AI on a laptop; the 128 GB M4 Max MacBook Pro became the 2025 reference machine for running big models.

Recommended AI models (newest first):

M4 base (16–32 GB): Ministral 3 8B Q4 (Dec 2025) → 40–60 tok/s; Qwen3.5 / Gemma 4 sub-12B (2026) → 35–55 tok/s; Gemma 3 12B Q4 → 25–35 tok/s; Ministral 3 14B Q4 tight on 32 GB; Voxtral Transcribe 2 (on-device STT, Feb 2026).
M4 Pro (24–64 GB): Ministral 3 14B Q4 → 30–40 tok/s; Devstral 2 Q4 (24B coding) → 20–25 tok/s; Gemma 3 27B Q4 (~16.5 GB) on 48 GB+ → 18–25 tok/s; GLM-4.7-Flash Q4 on 48 GB+ → 15–20 tok/s.
M4 Max (128 GB): Mistral Medium 3 (Jul 2026) → 8–10 tok/s; gpt-oss-120B Q4 (~70 GB) → 5–7 tok/s; Gemma 3 27B → 30–35 tok/s; GLM-4.7-Flash → 35–40 tok/s; multiple models resident at once. MiniMax M2.5 (230B) is just over the edge — Q3 or API.
Vision: Gemma 3 27B (native multimodal, Jul 2026) on Pro/Max; MiniCPM-V (smaller Macs); Image: Qwen Image 3.0 via MPS/ComfyUI; Audio: Voxtral Transcribe 2 real-time, Whisper large-v3 fallback.

M5 (October 2025)

Spec	Value
CPU	12-core (MacBook Pro/Air configs; 10-core in Air)
GPU	16-core (10-core in Air)
Neural Engine	16-core
Bandwidth	~153 GB/s (Air) up to ~250 GB/s (Pro)
Process	TSMC N3 (2nd-gen)
Devices	MacBook Pro 14, MacBook Air, iPad Pro
Price (Aug 2026)	MacBook Air from $1,099 (up $100 from M4); MacBook Pro 14 from ~$1,599

Best for: The mainstream upgrade — Apple’s ML research measured 19–27% faster LLM throughput than M4 on the same workloads, driven by greater memory bandwidth.

Recommended AI models (newest first):

16–24 GB: Ministral 3 8B Q4 → 45–65 tok/s; Qwen3.5 / Gemma 4 sub-12B (2026) → 40–60 tok/s; Gemma 3 12B → 30–40 tok/s; Voxtral Transcribe 2; Kokoro TTS.
32 GB: Devstral 2 Q4 → 25–30 tok/s; Gemma 3 27B Q4 (tight, ~16.5 GB + context).

M5 Pro / M5 Max (March 2026)

Spec	M5 Pro	M5 Max
CPU	15-core (5 super + 10 perf) or 18-core (6 super + 12 perf)	18-core
GPU	16-core or 20-core	40-core (each GPU core has a Neural Accelerator)
Neural Engine	16-core	16-core
Memory	up to 48 GB (reported)	up to 128 GB
Bandwidth	~256 GB/s class	~700 GB/s (~28% over M4 Max)
Devices	MacBook Pro 14/16	MacBook Pro 14/16
Price (Aug 2026)	14″ from $2,199 (up $200 from M4); 16″ from $2,699	14″ from $3,599; 16″ from $3,899 (Amazon record lows to ~$1,984 for 24 GB M5 Pro)

Best for: The best laptop for local LLMs in 2026. Apple claims up to 4× faster LLM prompt processing than M4 Pro/Max and up to 8× AI image generation vs M1 Pro/Max.

Recommended AI models (newest first):

M5 Pro (36–48 GB): Gemma 3 27B Q4 → 20–30 tok/s; GLM-4.7-Flash Q4 → 15–25 tok/s; Devstral 2 Q4 → 30–40 tok/s; Ministral 3 14B → 40–50 tok/s.
M5 Max (128 GB): Mistral Medium 3 (Jul 2026) → 12–16 tok/s; gpt-oss-120B Q4 → 8–10 tok/s; Gemma 3 27B → 40–50 tok/s; GLM-4.7-Flash → 45–55 tok/s; Ministral 3 14B → 60–80 tok/s; Qwen3.5 / Gemma 4 sub-12B → 90–120 tok/s.
Vision: Gemma 3 27B (multimodal); Image: Qwen Image 3.0 / Muse Spark 1.2 via MPS/MLX; Audio: Voxtral Transcribe 2, Kokoro voice pipelines at real-time ×3–5.

M5 Ultra (Mac Studio, 2026)

Spec	Value (reported)
CPU	36-core
GPU	84-core
Neural Engine	32-core
Memory	up to 512 GB (768 GB per latest 2026 reports)
Bandwidth	~1 TB/s class (reported)
Storage	up to 16 TB
Price	from ~$4,000+ (reported)

Best for: The extreme end of local inference — the only mainstream desktop that holds the newest frontier open models entirely in memory.

Recommended AI models (newest first):

Flagship (512 GB): MiniMax M2.5 Q4 (230B → ~130 GB) → 15–20 tok/s; Mistral Large 3 Q4 (675B → ~380 GB) → 5–8 tok/s; GLM-5.2 Q4 (~744B → ~420 GB) → 4–6 tok/s; DeepSeek V4 Flash / Kimi K3 (open-weight MoE giants) at Q2–Q3 → slow but real. No other desktop runs these.
Practical (70B class): Mistral Medium 3 → ~25–30 tok/s; gpt-oss-120B Q4 → 20–30 tok/s; multiple 27–120B models resident simultaneously (e.g., one coder + one general + one vision model).
Fine-tuning: QLoRA on 70B-class fits comfortably.

Quick decision table

Mac	Memory ceiling	Newest picks that fit (Q4)	70B-class speed
M4/M5 (base)	24–32 GB	Ministral 3 8B/14B, Qwen3.5, Gemma 4	—
M4 Pro / M5 Pro	48–64 GB	Gemma 3 27B, GLM-4.7-Flash, Devstral 2	~8–10 tok/s (Q3)
M4 Max	128 GB	Mistral Medium 3, gpt-oss-120B	8–12 tok/s
M5 Max	128 GB	Mistral Medium 3, gpt-oss-120B	12–16 tok/s
M5 Ultra (Studio)	512 GB+	MiniMax M2.5, Mistral Large 3, GLM-5.2	~20–25 tok/s

Bottom line: Buy the most unified memory you can afford — it’s the single spec that decides which of the newest models you can run at all. Speeds are bandwidth-bound; MLX/llama.cpp/Ollama all work great.