# I-Apple MLX distributed inference: amandla emodeli, isivinini kanye nemingcele yokwethenjwa

Published: 2026-10-18

Ucwaningo lwe-Apple lwangomhla ziyi-15 kuMfumfu lukhombisa i-Muse Spark Open isebenza kuma-Mac Studios amane futhi luchaza ukuthi kungani ukufana kwe-pipeline kanye ne-tensor kunemiphumela ehlukile esivinini sokukhiqiza. Lolu shicilelo lwangomhla ziyi-17 kuMfumfu, olushicilelwe ngomhla ziyi-18 kuMfumfu, luhlola futhi imephu yokufakwa kwe-benchmark ye-Mercor kanye nesibuyekezo sokuphepha se-OpenCode Console sangomhla ziyi-18 kuMfumfu GMT+7 esibhalwe ngokucacile. Izilinganiso zabathengisi, ukufakwa kwe-index ekhethiwe kanye nemibiko yokulahlekelwa umsebenzisi engaqinisekisiwe kugcina izexwayiso zazo ezihlukene.

Canonical: https://thedailydiff.dev/zu/video/2026-10-17-apple-mac-cluster/

## Lokho okuhlanganiswa yile vidiyo

- Isisindo semodeli kanye nendawo yokusebenza yokucabanga kokubili kusebenzisa inkumbulo; isisindo semodeli sodwa singenza amandla adingekayo abukelwe phansi.
- Ukufana kwe-pipeline kusabalalisa izendlalelo ezilandelanayo, kanti ukufana kwe-tensor kuhlukanisa umsebenzi ngaphakathi kwesendlalelo ngasinye futhi kudinga ukuxhumana okuphansi njalo.
- I-Apple ibika amathokheni angaphezu kwama-36 ngomzuzwana kuma-Mac Studios amane e-M5 Ultra anomongo wamathokheni angu-8,192; ukuqhathaniswa kwekhebula kumayelana nokulinda kwethokheni lokuqala kumongo wamathokheni angu-65,536.
- Ukutholakala kuka-Mercor okungama-90% wamahora okusebenza kuchaza ukufakwa okuncane noma okungekho ku-index yayo ekhethiwe ye-benchmark. Akusunguli izinga lokwehluleka kwemodeli elingama-90%.
- Isaziso esihleliwe se-OpenCode sichaza ingxenye yabasebenzisi be-Console abanikezeli bangaphandle abaxhunyiwe. Isaziso somphakathi asisunguli ukwephulwa kwazo zonke izifakiwe zasendaweni.
- Hlola umthwalo wakho wokusebenza wangempela bese ubuyekeza izinsizakalo ezinesiqinisekiso ngaphambi kokuthi uphathe ukusethwa kwemodeli yasendaweni njengesixazululo esiphelele sobumfihlo noma sezindleko.

## Izahluko

- 0:00 Kungani ama-Mac amabili engahlala enesivinini esifanayo?
- 0:41 Kungani imodeli idinga ama-Mac angaphezu koyedwa?
- 1:30 I-Thunderbolt iwenza kanjani ama-Mac asebenzisane?
- 2:19 Yimuphi umsebenzi ama-AI benchmarks awuhlanganisa ngempela?
- 3:11 Ubani i-OpenCode ethi wathinteka ngobusuku obandulelayo?
- 4:02 Yini i-local inference ekubeka ngaphansi kolawulo lwakho?
- 4:40 Ngingayithenga yini i-cluster yama-Mac amane?

## Umbhalo obhaliwe ohunyushwe

Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.

### Kungani ama-Mac amabili engahlala enesivinini esifanayo?

0:00 Ungalindela ukuthi ama-Mac amabili enze i-AI yakho iphindwe kabili ngokushesha. Izilinganiso ze-Apple zikhombisa ukuthi isivinini sokukhiqiza singahlala cishe singashintshiwe. Kule vidiyo, yini etholakala ngama-Mac amane? Yini ama-benchmarks ayiphuthayo? Ithuluzi elilodwa lokubhala ikhodi lixhuma ihadiwe ekuphuleni kobusuku obandulelayo. Sizobuyela kulokho. NgoMgqibelo, ziyi-17 kuMfumfu, lokhu yi-The Daily Diff. I-MLX uhlaka lwe-Apple lomthombo ovulekile lwe-machine learning ku-Apple

0:22 silicon. Ukuhlolwa kwayo okusha kusebenzisa i-Meta's Muse Spark Open kuma-Mac Studios amane. Isebenzisa inkumbulo ehlanganisiwe, ehanjiswana yi-CPU ne-GPU. I-Apple yakushicilela lokhu ngomhla ziyi-15 kuMfumfu. I-Mercor yamepha umsebenzi ama-AI benchmarks angaulinganisi kangako. Sihlanganisa ushicilelo lwangomhla ziyi-17 kuMfumfu, ngesibuyekezo sobusuku obandulelayo ngomhla ziyi-18 kuMfumfu mayelana ne-OpenCode Console.

### Kungani imodeli idinga ama-Mac angaphezu koyedwa?

0:41 Qala ngengxenye engalula. Ukuhlolwa kwe-Apple's Muse Spark Open kusebenzisa cishe amakhulu amahlanu amagigabhythi wemodeli yamabhithi amane kuphela. I-Mac Studio eyodwa ifinyelela kumakhulu amahlanu nambili amagigabhythi, futhi i-inference idinga inkumbulo yokusebenza nayo. Izisindo cishe zigcwalisa isikhwama ngaphambi kokuba upakishe isokisi elilodwa. Imishini eminingi yenza imodeli ilingane. Ukufana kwe-pipeline kunqamula imodeli ngezendlalelo zayo. I-Mac A yenza izendlalelo zangaphambili, bese idlulisela umphumela ophakathi ku-Mac B ngezendlalelo zangemuva.

1:05 Ngethokheni elilodwa elikhiqizwayo, lezo zinyathelo zisekwenzeka ngokulandelana. Izilinganiso ze-Apple zikhombisa ukuthi isivinini sokukhulula sihlala sifana kakhulu nemishini eyodwa. Uqashe umpheki wesibili, futhi wamnika isangweji elifanayo. Ku-Hacker News, omunye unjiniyela wabuza ukuthi kungani ithebula le-Apple lisebenzise imodeli yamabhithi amane ye-Meta eduze kwemodeli ye-Qwen ngokunemba kwayo kwangempela. Ithebula le-Apple lisho amafomethi, ngakho funda izinyawo ngaphambi kokuthenga. Ama-benchmark e-slide deck ahlala engahlulwa. Ukufana kwe-tensor kunqamula isendlalelo ngasinye esikhundleni salokho. Ama-Mac amaningi asebenza ngezingxenye zesendlalelo esifanayo ndawonye,

1:29 bese ehlanganisa imiphumela yawo.

### I-Thunderbolt iwenza kanjani ama-Mac asebenzisane?

1:32 Lokho kungashesha ukukhiqiza, kodwa imishini ixhumana ngokuphindaphindiwe. I-Apple isebenzisa ukufinyelela inkumbulo kude okuphansi kwe-latency nge-Thunderbolt. Isinyathelo esinesibindi somsebenzi wekhebula. Umphumela oyinhloko unamathokheni angaphezu kwamashumi amathathu nesithupha ngomzuzwana kusuka kuma-M5 Ultra Mac Studios amane, anamacala angaba yizinkulungwane eziyisishiyagalombili. I-Apple yakulinganisa ngokwayo. Izaziso ezinde zishintsha indaba. Kwesinye isivivinyo somongo omkhulu, izintambo ezimbili ngesixhumanisi zinciphise isikhathi sokulinda ithokheni lokuqala

1:55 cishe ngamasekhondi ayishumi. Imodeli yasendaweni ikunikeza umngcele owusizo nxazonke zokukhuthaza, inqobo nje uma ukusetha kwakho kwangempela kugcina kulendawo. I-Apple ayinikezanga intengo yokulingana yonke ngokumelene nama-API aphethwe. Lokho kuncike ekusetshenzisweni nasekeleni oyidingayo. Umshini ophendula imibuzo ngezikhathi ezithile uchitha isikhathi esiningi uba umhlobiso wedeski obiza kakhulu. Ukukhetha leyo modeli kudinga ukuhlolwa komsebenzi wakho.

2:17 IMercor yamepha ama-benchmarks angamashumi amabili nesishiyagalolunye ohlelweni lwentela lwaseMelika. I-index yayo ihlanganisa cishe izinkulungwane eziyisishiyagalolunye zemisebenzi kunezinkulungwane eziyishumi nesishiyagalolunye.

### Yimuphi umsebenzi ama-AI benchmarks awuhlanganisa ngempela?

2:21 Imephu ilinganisela ukuthi amaphesenti angamashumi ayisishiyagalolunye amahora okusebenza anokufakwa okuncane noma okungekho kwe-benchmark. Leso yisikhala esikhulu esingenalutho ephepheni lokuhlolwa. Imephu kaMercor ilinganisa ukufakwa. Umsebenzi ongenalutho unemisebenzi engahlolwanga kahle; amandla emodeli ahlala eyimibuzo evulekile. Ukumepha kusebenzise imodeli ye-AI, ngokubuyekezwa komuntu engxenyeni yamasaziso kaMercor. Ababuyekezi bamukele cishe izingxenye ezintathu kwezine zezinketho zayo zokuqala. Uhlelo lokuhlela ludinga uhlelo lwalo lokuhlela,

2:49 ngokwemvelo. Ngakho-ke uma ukhathazekile ngomsebenzi wakho, i-leaderboard ayikwazi ukuqeda impikiswano. Uma uthenga umsizi wokubhala ikhodi, dlala umsebenzi wangempela ngowakho amasheke okwamukela. Umsebenzi owusizo uhlanganisa izinhlangothi ezingemnandi ngemuva kwedemo. Akekho okhushulwa ngenxa yokuthi i-chatbot yaqeda ngokuzethemba i-spreadsheet engalungile. Okusiletha kusibuyekezo sobusuku obandulelayo. I-OpenCode ithi ukwephulwa kokuphepha kwathinta amaphesenti angu-1.4 wabasebenzisi be-Console. Isaziso sayo esihleliwe sithinta ingxenye enabanikezeli bangaphandle

### Ubani i-OpenCode ethi wathinteka ngobusuku obandulelayo?

3:13 abaxhunyiwe, futhi ithi abasebenzisi abathintekile bathunyelwe i-imeyili. Lokhu ikubiza ngokuthi ukuhlasela okukhulu kokuhlanganiswa, futhi ithi isehlakalo siyaqhubeka. Imbangela yangempela yomphakathi ayikafiki. Izimpendulo zihlanganisa abantu ababika ukuthi isikweletu somhlinzeki sanyamalala, okuyimibiko yabo yezehlakalo. Isaziso asisunguli ukulahlekelwa noma isinxephezelo somuntu ngamunye. I-Apple yayisanda kukhombisa ukuthi ungayixhuma kanjani i-OpenCode ku-local inference, futhi manje igama layo liseduze nesixwayiso sokuphepha.

3:38 Igama lomkhiqizo lihlanganisa imingcele yokwethemba eminingana ehlukene. Abasebenzisi be-Console kufanele bafunde isaziso esisemthethweni futhi bahlole ukusetshenziswa komhlinzeki. Uma uthintekile, landela imiyalelo yokulungisa futhi ubuyekeze iziqinisekiso ezixhunyiwe. Lesi saziso sisinikeza umkhawulo othile we-Console ophethwe. Alinikezi ubufakazi lapha bokwephulwa kwazo zonke izifakiwe zasendaweni. Nasi isixhumanisi. Isihloko se-Apple siqamba i-OpenCode njengeklayenti leseva yayo yemodeli yasendaweni.

### Yini i-local inference ekubeka ngaphansi kolawulo lwakho?

4:02 I-interface yokubhala ikhodi efanayo ingahlala phambi kwengqalasizinda ehlukile kakhulu. Lokho kuphendula ukuthi yini ethengwa ama-Mac amane. Amandla, nomngcele we-inference ongawuphatha. Izinsizakalo ezizungeze lowo mngcele zisafanele ukubuyekezwa kwazo. Umtethelo wami wokuthenga ulula. Fakaza ukuthi imodeli iyafaneleka, linganisa umsebenzi wonke, futhi uhlole yonke insizakalo ephethe isiqinisekiso. Umphumela we-Apple wenza i-inference enkulu yasendaweni ibe elula kakhulu, kanti ukuhlolwa kwe-pipeline kuchaza ukuthi kungani ukwengeza ihadiwe kungakhungathekisa.

4:23 Ukuhlolwa kwakho kokwamukela kufanele kube seduze kwe-oda lokuthenga. Uma uthanda ukufunda lokhu kunokuzwa ngikusho, i-diff ifika ebhokisini lakho leposi njalo ekuseni, mahhala ku-daily diff dot dev, isixhumanisi ngezansi. Ngakho-ke isinqumo sanamuhla si-NEEDS REVIEW. Ngingalinganisa umthwalo wami wokusebenza wangempela ngaphambi kokuthenga i-cluster.

### Ngingayithenga yini i-cluster yama-Mac amane?

4:40 Bhalisa, shaya insimbi, futhi ungitshele kumazwana uma ngabe ungakushaya ngokuhlukile. Futhi lokho yi-diff yanamuhla. Ngingu-Niko wase-Axrisi. Hlanganisa ngendlela efanele. Ngingu-Niko wase-Axrisi. Hlanganisa ngendlela efanele.

## Imithombo

- [Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio](https://machinelearning.apple.com/research/mlx-distributed-inference) — Apple Machine Learning Research
- [A benchmark map of the US economy](https://www.mercor.com/research/labor-index/benchmark-map/) — Mercor
- [Benchmark Map methodology and expert review](https://www.mercor.com/research/labor-index/benchmark-map/methodology/) — Mercor
- [Edited security-breach notice with Console scope](https://x.com/opencode/status/2111519637049319818) — OpenCode
- [Reader questions mixed benchmark precision formats](https://news.ycombinator.com/item?id=50119761) — Hacker News · peri-cl (reader opinion)
- [User reports vanished provider credit (unverified loss claim)](https://x.com/KingOKings/status/2111517723109015607) — Amruth Pillai on X