IngiaFungua pochi

LLM bora kwa Mac ya GB 48.

Modeli huria 38 kati ya 73 zinaenda vizuri ndani ya GB 48. Kwa chaguo-msingi GPU inaweza kutumia takriban GB 36 kati yake. Kasi hutegemea chipu: kuanzia 273 GB/s kwenye M4 Pro hadi 614 GB/s kwenye M5 Max.

Yenye akili zaidi inayoenda vizuriQwen3.8 27B8-bit · 28.7 GB · 4–17 tok/sKubwa zaidi inayoenda vizuriQwen3.6 35B A3B6-bit · 28.4 GB · 51–205 tok/sKwa kuandika codeQwen3 Coder 30B A3B Instruct8-bit · 32.4 GB · 35–140 tok/s

Mac zenye GB 48

Chipu 5 kuanzia M3 hadi M5 huja na GB 48. Fungua yoyote kuona orodha na kasi zake.

Zinaenda vizuri 41

Inatosha na kubakiza nafasi ya context, kwa tokeni 10 au zaidi kwa sekunde kwenye M5 Max.

ModeliToleoKumbukumbuM4 ProM5 Max
Qwen3.8 27B27B · Akili 33.78-bit28.7 GB4–810–17
Qwen3.6 35B A3B35B · 3B zinazotumika · Akili 18.26-bit28.4 GB51–91115–205
Muse Glimmer 30B29.6B · Akili 17.56-bit24.1 GB4–711–20
Gemma 4 31B30.7B · Akili 14.76-bit24.9 GB4–711–20
Nemotron 3.5 Lightning 30B A3B30B · 3B zinazotumika · Akili 12.98-bit31.9 GB38–6886–154
Granite 4.2 30B30B · Akili 12.76-bit24.4 GB4–711–20
MiniCPM5 2B2.52B · Akili 12.58-bit2.7 GB45–81102–182
Qwen3.5 9B9B · Akili 11.28-bit9.6 GB13–2329–51
Granite 4.2 8B8B · Akili 11.18-bit8.5 GB14–2633–58
Granite 4.2 3B3B · Akili 9.18-bit3.2 GB38–6886–154
gpt-oss-20b21B · 3.6B zinazotumika · Akili 94.25-bit asili13.8 GB65–115145–259
Nemotron 3 Nano 30B A3B30B · 3.5B zinazotumika · Akili 8.98-bit31.9 GB33–5975–133
Gemma 4 E4B8B · Akili 8.98-bit8.5 GB26–4658–102
Ministral 3 8B8.9B · Akili 5.58-bit9.5 GB13–2329–52
Nemotron 3 Nano 4B3.97B · Akili 4.88-bit4.2 GB29–5266–117
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B zinazotumika8-bit32.4 GB35–6279–140
GLM-4.7 Flash30B · 3B zinazotumika8-bit31.9 GB38–6886–154
Qwen3.6 27B27B8-bit28.7 GB4–810–17
Gemma 4 26B A4B25.2B · 3.8B zinazotumika8-bit26.8 GB31–5569–123
Devstral Small 2 (24B)24B8-bit25.5 GB5–911–19
Magistral Small 1.2 (2509)24B8-bit25.5 GB5–911–19
Mistral Small 3.2 (24B)24B8-bit25.5 GB5–911–19
LFM2.5 24B A2B24B · 2B zinazotumika8-bit25.5 GB59–104132–234
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB10–1817–31
Phi-4 (14B)14B8-bit14.9 GB11–1919–33
Gemma 4 12B Unified12B8-bit12.7 GB10–1722–39
Qwen3 8B8.2B8-bit8.7 GB14–2532–56
Llama 3.1 8B Instruct8B8-bit8.5 GB14–2633–58
Gemma 4 E2B5.1B8-bit5.4 GB51–91115–205
Qwen3.5 4B4B8-bit4.3 GB29–5164–114
Phi-4 mini instruct3.8B8-bit4 GB31–5569–123
Llama 3.2 3B Instruct3.21B8-bit3.4 GB36–6481–144
SmolLM3 3B3B8-bit3.2 GB38–6886–154
LFM2.5 2.6B2.69B8-bit2.9 GB42–7595–169
Qwen3.5 2B2B8-bit2.1 GB59–104132–234
Llama 3.2 1B Instruct1.23B8-bit1.3 GB95–168213–378
Qwen3.5 0.8B0.8B8-bit0.9 GB137–243307–546
Qwen3 0.6B0.6B8-bit0.6 GB205–364461–819
Nemotron 3.5 Bolt 74B A9B74B · 9B zinazotumika · Akili 21.43-bit ubora wa chini32.4 GB32–5671–126
Qwen3 Coder Next80B · 3B zinazotumika · Akili 9.22-bit ubora wa chini25 GB137–243307–546
GLM-4.5 Air106B · 12B zinazotumika2-bit ubora wa chini33.1 GB32–5773–129

Zinaenda, lakini polepole 1

Inatosha, lakini chini ya tokeni 10 kwa sekunde: sawa kwa kazi za kundi, polepole kwa mazungumzo.

ModeliToleoKumbukumbuM4 ProM5 Max
Llama 3.3 70B Instruct70B3-bit ubora wa chini30.6 GB4–79–16

Zinatosha kwa shida 8

Inahitaji karibu kumbukumbu yote: ongeza kikomo cha kumbukumbu ya GPU na weka context fupi.

ModeliToleoKumbukumbuM4 ProM5 Max
Qwen3.5 122B A10B122B · 10B zinazotumika · Akili 15.62-bit ubora wa chini38.1 GB40–7089–158
Mistral Medium 3.5 (128B)128B · Akili 14.22-bit ubora wa chini40 GB3–57–12
gpt-oss-120b117B · 5.1B zinazotumika · Akili 11.62-bit ubora wa chini36.6 GB77–137173–307
Mistral Small 4 (119B)119B · 6.5B zinazotumika · Akili 11.32-bit ubora wa chini37.2 GB61–109138–246
Llama 4 Scout 17B-16E Instruct109B · 17B zinazotumika · Akili 6.52-bit ubora wa chini34.1 GB23–4152–93
Devstral 2 (123B)123B2-bit ubora wa chini38.4 GB3–67–13
Ornith 1.5 122B122B · 10B zinazotumika2-bit ubora wa chini38.1 GB40–7089–158
Nemotron 3 Super 120B A12B120B · 12B zinazotumika2-bit ubora wa chini37.5 GB32–5773–129

Kubwa mno 23

Inahitaji kumbukumbu zaidi ya GB 48, hata kwa biti 2.

Maswali

LLM ipi ni bora kwa Mac yenye GB 48?
Miongoni mwa modeli zenye Intelligence Index iliyochapishwa, Qwen3.8 27B ina alama ya juu zaidi kati ya zile zinazoenda vizuri ndani ya GB 48 (8-bit). Kubwa zaidi inayoenda vizuri ni Qwen3.6 35B A3B. Kwa code: Qwen3 Coder 30B A3B Instruct.
Je, GB 48 zinatosha kwa AI ya ndani kwenye Mac?
Modeli huria 38 kati ya 73 zilizo hapa zinaenda vizuri ndani yake kwa biti 4 au zaidi. Kwa chaguo-msingi macOS huiruhusu GPU kutumia takriban GB 36 kati yake; modeli na context yake lazima zitoshe humo.
Je, Mac ya GB 48 inaweza kuendesha gpt-oss-20b?
Ndiyo, 4.25-bit asili, kwa takriban tokeni 65–115 kwa sekunde kwenye M4 Pro na 145–259 kwenye M5 Max.

Ukubwa mwingine: 8 GB · 16 GB · 18 GB · 24 GB · 32 GB · 36 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB

Ziendeshe kwa MLX Harness (bure, open source) au mlx-lm ya Apple ndani ya Terminal: hatua kwa hatua.

Jinsi inavyokokotolewa

Kumbukumbu: vigezo vya modeli kwa biti 8, 6, 4, 3 au 2 kwa kila weight pamoja na group scales za MLX (nusu biti kwa kila weight), au muundo wake asili, pamoja na nafasi ya context juu yake. Kwa chaguo-msingi macOS huiruhusu GPU kutumia takriban 75% ya unified memory. Kasi: bandwidth ya kumbukumbu ikigawanywa kwa baiti zinazosomwa kwa kila tokeni ndiyo kikomo cha juu kwa mazungumzo moja; katika benchmarks za umma za llama.cpp kwenye Apple Silicon uendeshaji halisi hufikia takriban 45% hadi 80% yake, na huo ndio upana unaoonyeshwa. Haya ni makadirio, si vipimo: pima Mac yako mwenyewe kwenye MLX Bench. Takwimu za modeli ni kutoka kwa kadi rasmi ya kila modeli, takwimu za chipu ni kutoka Tech Specs za Apple. Imekaguliwa 2026-10-19. Data ileile ya mlx.app/compare.

LLM bora za ndani kwa Mac ya GB 48 (M3 hadi M5)