IngiaFungua pochi

M3 Ultra yenye GB 96 inaweza kuendesha nini?

Modeli huria 43 kati ya 73 zinaenda vizuri juu yake. Kwa chaguo-msingi GPU inaweza kutumia takriban GB 72 kati ya GB 96 zake, na husoma kumbukumbu kwa 819 GB/s, jambo linaloamua kasi yake ya kuandika. Inapatikana katika: Mac Studio (2025).

Yenye akili zaidi inayoenda vizuriQwen3.8 27B8-bit · 28.7 GB · 13–23 tok/sKubwa zaidi inayoenda vizurigpt-oss-120b4.25-bit asili · 65.3 GB · 137–243 tok/sKwa kuandika codeQwen3 Coder Next6-bit · 65 GB · 154–273 tok/s

Zinaenda vizuri 48

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

ModeliToleoKumbukumbuTokeni/s
Qwen3.8 27B27B · Akili 33.78-bit28.7 GB13–23
Nemotron 3.5 Bolt 74B A9B74B · 9B zinazotumika · Akili 21.46-bit60.1 GB50–90
Qwen3.6 35B A3B35B · 3B zinazotumika · Akili 18.28-bit37.2 GB115–205
Muse Glimmer 30B29.6B · Akili 17.58-bit31.5 GB12–21
Gemma 4 31B30.7B · Akili 14.78-bit32.6 GB11–20
Nemotron 3.5 Lightning 30B A3B30B · 3B zinazotumika · Akili 12.98-bit31.9 GB115–205
Granite 4.2 30B30B · Akili 12.78-bit31.9 GB12–21
MiniCPM5 2B2.52B · Akili 12.58-bit2.7 GB137–243
gpt-oss-120b117B · 5.1B zinazotumika · Akili 11.64.25-bit asili65.3 GB137–243
Qwen3.5 9B9B · Akili 11.28-bit9.6 GB38–68
Granite 4.2 8B8B · Akili 11.18-bit8.5 GB43–77
Qwen3 Coder Next80B · 3B zinazotumika · Akili 9.26-bit65 GB154–273
Granite 4.2 3B3B · Akili 9.18-bit3.2 GB115–205
gpt-oss-20b21B · 3.6B zinazotumika · Akili 94.25-bit asili13.8 GB194–345
Nemotron 3 Nano 30B A3B30B · 3.5B zinazotumika · Akili 8.98-bit31.9 GB100–177
Gemma 4 E4B8B · Akili 8.98-bit8.5 GB77–137
Llama 4 Scout 17B-16E Instruct109B · 17B zinazotumika · Akili 6.54-bit61.3 GB38–68
Ministral 3 8B8.9B · Akili 5.58-bit9.5 GB39–69
Nemotron 3 Nano 4B3.97B · Akili 4.88-bit4.2 GB88–156
GLM-4.5 Air106B · 12B zinazotumika4-bit59.6 GB54–96
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B zinazotumika8-bit32.4 GB105–187
GLM-4.7 Flash30B · 3B zinazotumika8-bit31.9 GB115–205
Qwen3.6 27B27B8-bit28.7 GB13–23
Gemma 4 26B A4B25.2B · 3.8B zinazotumika8-bit26.8 GB92–164
Devstral Small 2 (24B)24B8-bit25.5 GB14–26
Magistral Small 1.2 (2509)24B8-bit25.5 GB14–26
Mistral Small 3.2 (24B)24B8-bit25.5 GB14–26
LFM2.5 24B A2B24B · 2B zinazotumika8-bit25.5 GB176–312
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB23–41
Phi-4 (14B)14B8-bit14.9 GB25–44
Gemma 4 12B Unified12B8-bit12.7 GB29–52
Qwen3 8B8.2B8-bit8.7 GB42–75
Llama 3.1 8B Instruct8B8-bit8.5 GB43–77
Gemma 4 E2B5.1B8-bit5.4 GB154–273
Qwen3.5 4B4B8-bit4.3 GB86–152
Phi-4 mini instruct3.8B8-bit4 GB92–164
Llama 3.2 3B Instruct3.21B8-bit3.4 GB108–193
SmolLM3 3B3B8-bit3.2 GB115–205
LFM2.5 2.6B2.69B8-bit2.9 GB127–226
Qwen3.5 2B2B8-bit2.1 GB176–312
Llama 3.2 1B Instruct1.23B8-bit1.3 GB284–504
Qwen3.5 0.8B0.8B8-bit0.9 GB410–728
Qwen3 0.6B0.6B8-bit0.6 GB614–1092
Qwen3.8 Flash Next180B · 6B zinazotumika · Akili 39.82-bit ubora wa chini56.3 GB194–345
Qwen3.5 122B A10B122B · 10B zinazotumika · Akili 15.63-bit ubora wa chini53.4 GB84–149
Mistral Small 4 (119B)119B · 6.5B zinazotumika · Akili 11.33-bit ubora wa chini52.1 GB132–234
Ornith 1.5 122B122B · 10B zinazotumika3-bit ubora wa chini53.4 GB84–149
Nemotron 3 Super 120B A12B120B · 12B zinazotumika3-bit ubora wa chini52.5 GB70–124

Zinaenda, lakini polepole 3

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

ModeliToleoKumbukumbuTokeni/s
Llama 3.3 70B Instruct70B6-bit56.9 GB6–12
Mistral Medium 3.5 (128B)128B · Akili 14.23-bit ubora wa chini56 GB7–12
Devstral 2 (123B)123B3-bit ubora wa chini53.8 GB7–12

Zinatosha kwa shida 1

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

ModeliToleoKumbukumbuTokeni/s
MiniMax M2.7230B · 10B zinazotumika2-bit ubora wa chini71.9 GB119–211

Kubwa mno 21

Inahitaji kumbukumbu zaidi ya GB 96, hata kwa biti 2.

Mac nyingine

Kila Mac yenye GB 96, ya chipu yoyote.

M3 Ultra yenye kumbukumbu zaidi au kidogo:

Chipu jirani:

Ana kwa ana:

Dhidi ya PC na kadi za michoro:

Ilinganishe na Mac au GPU yoyote · M3 Ultra kwa AI ya ndani ya kifaa · Mac bora kwa LLM za ndani · Kila Mac

Maswali

LLM ipi ni bora kwa M3 Ultra yenye GB 96?
Miongoni mwa modeli zenye Intelligence Index iliyochapishwa, Qwen3.8 27B ina alama ya juu zaidi kati ya zile zinazoenda vizuri hapa: 8-bit, takriban tokeni 13–23 kwa sekunde. Kubwa zaidi inayoenda vizuri ni gpt-oss-120b.
Je, M3 Ultra yenye GB 96 inaweza kuendesha Llama 3.3 70B?
Ndiyo, lakini polepole: 6-bit, kwa takriban tokeni 6–12 kwa sekunde.
gpt-oss-20b ina kasi gani kwenye M3 Ultra?
Takriban tokeni 194–345 kwa sekunde (4.25-bit asili), kutokana na bandwidth ya kumbukumbu ya chipu ya 819 GB/s.

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.

M3 Ultra yenye GB 96 inaendesha LLM zipi? Modeli na kasi