I-API Ye-MiMo Engavinjelwe Kumabhithi Ayi-8 Okwethembekile

Iningi lamakhathalogu likuthengisela inani lamapharamitha. Yileyo inombolo engalungile yokuthenga ngayo, ngoba yonke imodeli enkulu enezisindo ezivulekile ongayikwazi ngempela isuke icindezelwe ukuze ingene ndawana thize. Umbuzo oqotho akuwona lo othi "yayinkulu kangakanani ngaphambi kokucindezelwa?" kodwa yilo othi "kusasele okungakanani kuyo?" Lesi sihloko simayelana nalowo mbuzo, futhi mayelana nemodeli ohlwini lwethu lapho impendulo ingaphandle nje phose konke.

01Yini ngempela ekunikeza yona i-API ye-MiMo engavinjelwe

I-API ye-MiMo engavinjelwe iyinqama yemodeli eyodwa, u-MiMo-V2.5-Pro-W8A16, efinyeleleka ngezilimi ezintathu ezikhuluma nezisindo ezifanayo eqoqweni lethu lama-GPU: u-/v1/chat/completions, u-/v1/messages no-/v1/responses. Zontathu ziyasakaza futhi zisekela ukubiza amathuluzi nokucabanga.

U-Engavinjelwe unencazelo ekhawulelwe lapha: alukho ungqimba lokwenqaba phambi kwemodeli, asikho isihlungi kokukhishwayo kwayo. Asithatheli phansi, asihlukanisi, futhi asibhali kabusha. Uthola ukuziphatha kwesisekelo uqobo — okusho, kumodeli enezisindo ezivulekile ezingeni locwaningo, ukuthi izibandakanya nocwaningo lokuphepha, izinganekwane zabantu abadala noma ezinodlame, ukucabangela izokwelapha nezomthetho, kanye nezicelo iziphetho zezentengiselwano ezikhonziwe ezizinqaba ngokuzenzakalelayo. Yithuluzi lochwepheshe ngaphansi kwenqubomgomo yethu yokusebenzisa ngokuzibophezela, hhayi ukungabikho kwemithetho.

02Uhlobo lwe-MiMo: ukubheja kuka-Xiaomi ekuqeqeshweni ngaphambili okuqala ngokucabanga

I-MiMo ayiyona into ephuma ku-Llama enegama elisha; iwuhlobo oluhlukile oluvela eqenjini le-MiMo lika-Xiaomi, futhi umqondo walo wokusungulwa ubalulekile ngaphambi kokuba uhlole imodeli. Iqala nge-MiMo-7B (ngo-Ephreli 30, 2025, elinelayisensi ye-MIT), okuthi umbiko wayo wobuchwepheshe uphikise ukuthi amalabhorethri amaningi aphatha ukucabanga njengenkinga yangemuva kokuqeqeshwa ehlonyiwa nge-RL, futhi lokhu kuhlehla emuva. Kunalokho, i-Xiaomi yakha kabusha ukuqeqeshwa ngaphambili ngokusekelwe kulokho: inhlanganisela yezigaba ezintathu kumathokheni ayizigidigidi ezingu-25, isigaba sesibili sisebenzisa cishe u-70% wezibalo nekhodi, isigaba sesithathu sifaka imikhondo yokucabanga eyenziwe kalula kuyilapho selula umongo kusuka ku-8K uye ku-32K — kanye nenjongo ye-Multi-Token Prediction esebenza kabili njengekhanda le-speculative-decoding.

Umphumela kwaba imodeli eyisisekelo engu-7B eyaziphatha njengento enkulu kakhulu. I-MiMo-7B-Base yathola amaphuzu angu-32.9 ku-AIME 2024 lapho i-Qwen2.5-7B ithola u-10.1, i-Gemma-2-9B u-0.0 ne-Llama-3.1-8B u-0.3. Ukuqeqeshwa kwangemuva kwabe sekuqhuba i-RL phezu kwezinkinga zezibalo nekhodi ezingu-130,000 eziqinisekekayo ngomvuzo wekhodi oqhutshwa ubunzima bokuhlolwa, ukubalekela isicupho somvuzo ongaphelele esizinyazayo i-RL yekhodi. Umphumela wafika ngaphambi kuka-o1-mini:

I-AIME 2025, amanani abikwe yi-Xiaomi asuka embikweni wobuchwepheshe we-MiMo (arXiv:2505.07608). I-MiMo-7B-RL nayo ibike u-95.8 ku-MATH-500 no-57.8 ku-LiveCodeBench v5.

Umndeni wanda kunokuphambuka: i-MiMo-VL-7B (ngoJuni 2025) yafaka i-encoder ye-Qwen2.5-ViT kusakhiwo esifanayo, futhi uDisemba 2025 waletha i-MiMo-Audio-7B kanye nokungena kokuqala kwe-MoE, i-MiMo-V2-Flash (angu-309B esewonke / angu-15B asebenzayo, amathokheni ayi-27T, ukunemba okuxutshiwe kwe-FP8). I-MiMo-V2-Pro yalandela ngoNdasa 2026 cishe ku-1T / 42B esebenzayo ngomongo we-1M, ngemuva kwamasonto ku-OpenRouter njenge-"Hunter Alpha". I-MiMo-V2.5 ne-MiMo-V2.5-Pro zafika ndawonye ngo-Ephreli 22, 2026, zombili nge-MIT.

03Ngaphakathi kwe-MiMo-V2.5-Pro: ukuthi isisekelo sibukeka kanjani

Ngokwekhadi lemodeli elishicilelwe yi-Xiaomi, i-MiMo-V2.5-Pro iyimodeli ye-Mixture-of-Experts engaminyene enamapharamitha esewonke ayizigidigidi ezingu-1.02 futhi angu-42 billion asebenzayo kuthokheni ngalinye kuzingqimba ezingu-70 — eyodwa iminyene, ezingu-69 zi-MoE — kanye nochwepheshe abaqondiswayo abangu-384, abangu-8 kuthokheni ngalinye. Usayizi ofihlekile ungu-6144, ngosayizi oluncane wangaphakathi lwe-MoE ongu-2048.

Izinketho ezimbili zibumba indlela ezizwakala ngayo ekusetshenzisweni. Isakhiwo sokunaka siyixubo: izingqimba ezingu-10 zokunaka ngokugcwele nezingu-60 zewindi eliqhubekayo ngewindi elingu-128, kanye nokuchema kwe-attention-sink okufundekayo. I-Xiaomi ibika ukuthi lokhu kunciphisa indawo yokugcina i-KV-cache cishe ngokuphindwe ka-7 uma kuqhathaniswa nokunaka okujwayelekile kuwo wonke amazinga, kuyilapho igcina ikhwalithi yomongo omude — umehluko phakathi kwewindi elithi ngokwezibalo lide nelingu elingakwazi ukuligcwalisa. Iphinde ilethe izingqimba ezintathu ze-Multi-Token Prediction, iphe i-decoder ikhanda le-speculative.

Ukuqeqeshwa kwakungamathokheni ayizigidigidi ezingu-27 ngobude bochungechunge lwemvelo olungu-32K ngokunemba okuxutshiwe kwe-FP8 (E4M3), kwase kuba nokulungiswa okucophelele okuqashwayo, i-RL yama-ejenti ngobukhulu obukhulu kanye nokuncishiswa kwe-multi-teacher on-policy. Gcina lelo qiniso le-FP8 engqondweni; yiwo umbono onamandla kunayo yonke wokukhonza lesi sisekelo emabhithini ayi-8.

I-Xiaomi ibika imodeli ku-57.2% ku-SWE-bench Pro, 72.9 ku-τ3-Bench yokusetshenziswa kwamathuluzi kanye no-48.0% ku-Humanity's Last Exam — amanani omthengisi uqobo emodelini engacindezelwe, ecashunwe njengalokho.

Sikhonza iwindi lamathokheni angu-262,144, hhayi i-1M egcwele yesisekelo — umongo esingawuqinisekisa ngesivinini, kunenombolo eyinhloko eyehla uma kunomthwalo.

04Kusho ukuthini ngokunemba u-W8A16

Yonke imodeli kule khathalogu iphethe isijobelelo sokukalwa kwezinga esichaza ushintsho lwangempela, olulinganisekayo, kunesigaba sokumaketha. Ezethu u-3BIT-REAP, u-W4A16-AUTOROUND-REAP no-W8A16. I-MiMo-V2.5-Pro iyisithathu.

Uphawu luhlukana kabili. U-W8 ukunemba kwesisindo: ipharamitha ngalinye ligcinwa njenge-integer yamabhithi ayi-8 esikhundleni se-float yamabhithi ayi-16, ngesici sesikali esabelwanwayo ngeqembu elincane ukuze ibanga ledatha elizungezayo leqembu ngalinye lisinde. U-A16 ukunemba kokusebenza: amanani ahamba phakathi kwezingqimba ekuqagelweni ahlala emabhithini ayi-16. Izisindo ziyacindezelwa; izibalo azicindezelwa.

Leyo ngxenye yesibili yenza umsebenzi othule, obalulekile. Ukusebenza kunzima kakhulu ukukukala kunezisindo: kuqukethe amashaneli angajwayelekile kakhulu — izici ezimbalwa obukhulu bazo bunciphisa konke okunye kanti futhi ziphethe isabelo esingalingani sokuziphatha kwemodeli. Uma uzicindezela zibe nokunemba okuphansi, ususa kanye lesikhathi ophawu obudinga. Amasu ezisindo kuphela agwema le nkinga ngokushiya indlela yokusebenza ingathintiwe, yingakho umsebenzi wokukalwa kwezinga onzulu uhlangene kuwo.

Ngakho umbuzo osebenzayo ucija kuento eyodwa: malini amabhithi amancane isisindo esingawabekezelela?

05I-W8A16 kunye ne-W4A16 kunye namabhithi amathathu: lapho ukulahlekelwa kuya khona ngempela

Ucwaningo olukhulu kunazo zonke oshicilelwe kulokhu yi-Neural Magic ne-Red Hat ethi "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization — ngaphezu kokuhlolwa okungu-500,000 emndenini we-Llama-3.1 ku-8B, 70B no-405B. Umphumela wabo oyinhloko, ngamanani abo:

Ifomethi	Yini ecindezelwayo	Ukubuyisela okubikiwe kunee-BF16
W8A8-FP (amabhithi ayi-8)	Izisindo + ukusebenza, i-float	~99.75% isilinganiso — "cishe akulahlekelwa lutho"
W8A8-INT (amabhithi ayi-8)	Izisindo + ukusebenza, i-integer	99–100% ku-8B/70B; 98.2–99.3% ku-405B
W4A16-INT (amabhithi amane)	Izisindo kuphela	98–99.4% (8B); 98.6–99.8% (70B)

Amanani abikwe yi-Neural Magic / Red Hat (arXiv:2411.02355) kumamodeli e-Llama-3.1 — ukubukeka kokuthengiselana, hhayi amaphuzu ale modeli.

Uma ufunda ngokuqotho, lokho kubukeka njengombono ophikisayo wokunaka: amabhithi amane abuyisela u-98–99%, ngakho kungani unake? Ngoba lokho kuyizilinganiso zezemfundo ezinezinketho eziningi, eziseduze nethuluzi elibi kunawo wonke elitholakalayo lokuthola umonakalo wokukalwa kwezinga. Umbuzo wesilinganiso unempendulo eyodwa efanele efinyeleleka ngethokheni elilodwa noma amabili, futhi iphutha lokukalwa kwezinga lincane, umsindo oseduze nezero-mean kulogithi ngalinye; embuzweni omfushane onemargini ebanzi, lowo msindo uyamunywa.

Manje bhala amathokheni angu-4,000 edokhumenti yobunjiniyela, gcina izwi kuyo yonke indaba yamathokheni angu-40,000, noma khiphe i-JSON okumele iqinisekiswe emzamweni wokuqala. Umsindo ofanayo uyathathwa amasampula izikhathi eziyizinkulungwane ngochungechunge, isampula ngalinye linquma elilandelayo. Indlela yokwehluleka akuyona impendulo engalungile — kuwukuxega: ukudedela umbambo womyalelo kusukela kumzuliswano wokuqala, umlingiswa oyeka ukuzwakala njengoyena ekungathokheni engu-3,000, ibhuleki yokuvala engaze ifike.

Lapho amabhithi aya khona ngempela

Iphutha lokukalwa kwezinga alisatshalaliswa ngokulingana kuwo wonke amakhono emodeli; ligxile emsileni. Ukuziphatha okunobuningi obukhulu — ulimi lwesiNgisi olulandela umthetho wolimi, isimo somsebenzi we-Python, amaqiniso ayinkulungwane amaphambili emhlabeni — kubhalwa kaninginingi kuzo izisindo eziningi futhi kusinda cishe kunoma yini. Amakhona akalwe ngokunemba, aphansi ngobukhulu okuhlukaniswa kwesisindo abambe izinto ezingavamile: ukusayina okunembile kwe-API engaziwa kangako, isigqi somqondisi othile, umehluko phakathi kwama-CVE amabili aseduzane. Lawo ngamanani igridi eqinile eyiyingilisela kanye.

Isinyathelo esisuka kumabhithi amane siye kwamathathu yilapho lokhu kuyeka ukuba yintela kube usokhalo. Imithetho yokukhula kwe-k-bit inference ka-Dettmers noZettlemoyer yathola ukuthi amabhithi amane iwona wona iphuzu elinyusa ukunemba kwe-zero-shot kubhithi ngalinye lenkumbulo yemodeli, kuyilapho ukuwohloka kubusa ngezansi kwalokho. Amabhithi amane anikeza amazinga angu-16 eqenjini; amathathu anikeza angu-8. Wehlisa ngengxenye ubunjalo begridi esivele iqinile, futhi iphutha lokuyingilisa isisindo ngasinye liphindeka cishe kabili.

Izibalo ezinquma lokhu

Ukuqhathanisa okufanele kwenziwe yizibalo, hhayi isilinganiso. Okusinda ukucindezelwa cishe amapharamitha × amabhithi ngepharamitha — ivolumu yolwazi equkethwe ifayela lesisindo. Thatha izisekelo ezimbili ezicatshangiwe zekhwalithi elinganayo ngepharamitha ngalinye:

• Isisekelo A: amapharamitha angu-400B kumabhithi amathathu → cishe ama-150 GB esisindo esisele.
• Isisekelo B: amapharamitha angu-160B emabhithini ayi-8 → cishe ama-160 GB esisindo esisele.

Ngenani lamapharamitha eqondile, u-A unqoba ngokuphindwe ka-2.5 kunayedwa — inombolo ikhadi lemodeli eliqala ngayo. Ngolwazi olusele zilingana, futhi u-B unenzuzo yesakhiwo ngaphandle kwalokho: izisindo zawo zihlala kugridi encane ngokwanele ukuthi ukuziphatha kwawo emsileni kuqinile, kuyilapho umsila ka-A sewukaliwe waba yingxenye enkulu. U-A wazi izinto eziningi; u-B wazi lokho akwaziyo ngokucophelele kakhulu. Ekukhiqizeni okude, ukuhleleka okunembile, ukulawula isitayela nokuhambisana okuthembekile nomongo omude, thatha u-B.

Izibalo ezichazayo, hhayi isilinganiso: zibonisa ukuthi kungani u-"isisekelo esikhulu" no-"imodeli ekhonzwe kangcono" kuyizimangalo ezahlukene.

06Kungani amathokheni okuqeqeshwa ayizigidigidi ezingu-27 enza ukukalwa kwezinga okuqinile kube kubi kakhudlwana

Umphumela wesibili, ongaqondakali kalula, uthinta ngokuqondile lesi sisekelo. Iphepha lika-Kumar nabanye le-ICLR 2025 elithi Scaling Laws for Precision lithola ukuthi uma imodeli iqeqeshwe idatha eningi, umonakalo wokukalwa kwezinga ngemuva kokuqeqeshwa uba mkhulu kuyo. Ukunemba okuphansi kunciphisa inani lamapharamitha asebenzayo emodeli, futhi lokho kuwohloka kukhula njengoba idatha yokuqeqeshwa ngaphambili imunywa — kuze kube sezingeni lapho idatha eyengeziwe yokuqeqeshwa ngaphambili ingaba yingozi ngokwesibalo uma uhlose ukukala izinga ngokuqinile, ngoba ubeka isakhiwo esicophelele igridi ezosicima.

Imodeli eqeqeshwe ngamathokheni ayi-5T inesakhiwo sesisindo esiqinile, esiphindaphindiwe kanye nesikhala esiseleyo. Imodeli eqeqeshwe ngamathokheni ayi-27T — isabelomali se-MiMo-V2.5-Pro — yasebenzisa leyo datha eyengeziwe ukubhala umehluko ocophelele emananini esisindo anembile kakhulu, okuyikho kanye igridi yamabhithi amathathu engeke ikwazi ukukumela.

Kukhona-ke iqiniso le-FP8 elivela esigabeni 03, umbono ohlanzeke kunayo yonke kulesi sihloko. I-MiMo-V2.5-Pro yaqeqeshwa ngaphambili ngokunemba okuxutshiwe kwe-FP8. Izisindo zayo zabunjwa kusukela ekuqaleni yi-optimizer esebenza cishe ngamabhithi ayisishiyagalombili okunemba. Ngakho-ke ukuyikhonza ngezisindo zamabhithi ayi-8 akuyona umcabango wangemuva olahlekelwayo ohlonywa emodelini yamabhithi ayi-16 — kuseduze nokubuyisela izisindo esimweni sezinombolo ezafunda kuso. Kuncane isakhiwo esicophelele esikhona ngaphansi kwebhithi lesishiyagalombili okumele silahleke. Akufani nalezo zisindo eziphoqelelwa ngezansi ziye kwezintathu.

07I-REAP: kungani ukuthenwa kochwepheshe kuwuhlobo oluhlukile lokulahlekelwa kunokuyingilisa

Ezimbili kwezintathu izigaba ziphethe isijobelelo esithi REAP — umsebenzi ohlukile ngokuphelele kunokukalwa kwezinga. I-REAP (Router-weighted Expert Activation Pruning, Cerebras Research, arXiv:2510.13999, ICLR 2026) ukucindezela ngokuhlangana kanye kwamamodeli e-MoE angaminyene: ihlola ochwepheshe ngamunye ngamanani e-router gate nangezinga lokusebenza kochwepheshe, bese icima abathole amaphuzu aphansi kunazo zonke. I-Cerebras ibika imiphumela eseduze nokungabi nakulahlekelwa ekukhiqizweni kwekhodi nasekubizeni amathuluzi ngisho nasekutheneni u-50% wochwepheshe.

Sisebenzisa i-REAP ngesizathu ekhona ngaso: ezisekelweni ezinkulu kunazo zonke yiyona eyenza zikwazi ukukhonzwa. Kodwa qaphela umehluko ngohlobo. Ukukalwa kwezinga kufiphaza isisindo ngasinye kancane; ukuthena kususa ochwepheshe ngokuphelele. Uma umsebenzi wakho uqondisa kakhulu ekhonweni elalihlala kuchwepheshe okwakungakhethwa kaningi — ulimi lokuhlela olungajwayelekile, isigcawu esikhethekile samagama — akuwohlokile. Kuwukuthi kuphelile, futhi akukho ukuphonsa i-prompt okungakubuyisa. Izilinganiso zihlala zizinhle; icala lakho lomsila kungenzeka ukuthi lingahlali linjalo.

I-MiMo-V2.5-Pro · W8A16 ayiphethe sijobelelo se-REAP. Bonke ochwepheshe abangu-384 bakhona, futhi i-router ikhetha kusukela kusethi egcwele isisekelo esaqeqeshwa ngayo.

Isigaba	Amabhithi esisindo	Ochwepheshe	Okuthengiselayo
W8A16	8	Bonke bakhona	Akukho okwesakhiwo. Isigaba sokwethembeka.
W4A16-AUTOROUND-REAP	4	Kuthenwe	Imininingwane ecophelele yesisindo kanye nochwepheshe abasetshenziswa kancane, uqhathanisa nesisekelo esikhulu kakhulu esisezingeni lentengo elifanayo.
3BIT-REAP	3	Kuthenwe	Ukunemba kwesisindo okubalulekile kanye nochwepheshe abasetshenziswa kancane — intengo yokuqhuba izisekelo ezinkulu kunazo zonke ngesivinini.

Isigaba esiphakathi sifanele ukuzwiwa ngokulingana. I-AutoRound ayiyona ukuyingilisa okulula nje: indlela ka-Intel (arXiv:2309.05516) isebenzisa i-signed gradient descent phezu kwezigaba ezingu-200 ukulungisa izinqumo zokuyingilisa nebanga lokugxumeka ngeqembu lesisindo ngalinye, futhi ibikwa iseduze nokungabi nakulahlekelwa emabhithini amane ezilungiselelweni eziningi, yehlula i-GPTQ, i-AWQ, i-HQQ ne-OmniQuant kuwo wonke amaqoqo amaningi ahlolwayo. Imodeli ye-W4A16-AUTOROUND iyimodeli enhle ngempela yamabhithi amane. Isengumodeli yamabhithi amane.

08Ngubani okumele akhethe isigaba samabhithi ayi-8 — nokuthi ngubani okungafanele

I-MiMo-V2.5-Pro · W8A16 ihlala entengweni ephansi kakhulu yekhathalogu yonke — $0.50 engenayo / $2.00 ephumayo ngethokheni eyisigidi, ibanga elifanayo namamodeli esigaba sokuqala samabhithi amane, futhi cishe yikota yentengo yokungenayo yemodeli ebiza kunazo zonke esiyikhonzayo. Awukhokheli intengo ephakeme ngenxa yokwethembeka; ukhetha uhlangothi oluhlukile lwesabelomali esifanayo.

Yikhethe uma

• Ubude bokukhishwayo bungumsebenzi. Ukubhala kobuchwepheshe okude, izinganekwane ezingangesahluko, imizuliswano yamadokhumenti agcwele — ithokheni ngalinye elikhiqizwayo yileyo umsindo wokucindezela unqwabelanayo kuyo.
• Izwi nokuhambisana kubalulekile. Umsebenzi wabalingiswa, izwi lebhrendi, isimo esigcinwe kuso sonke isikhathi sokusebenza eside. Isitayela sihlala emsileni, futhi umsila yiwona okuqala ukudliwa ukukalwa kwezinga okuqinile.
• Okukhishwayo okuhlelekile kufanele kuqinisekiswe emzamweni wokuqala. Imodeli isekela u-json_schema no-response_format; ukuhambisana nohlelo yisici sokunemba, futhi zonke izikhathi zokuzama futhi ezigwenywayo kuyimali nesikhathi esongiwe.
• Ugcwalisa umongo. Ukuhambisana nemiyalelo kuthokheni engu-200,000 kunzima kunakuthokheni engu-2,000, futhi kuyawohloka ngokunemba kwesisindo — njengoba nomkhondo wokucabanga omude unjalo, lapho amaphutha enqwabelana ngendlela efanayo.
• Umsebenzi ungavinjelwe ngokwemvelo. Ucwaningo lokuphepha, i-red-teaming, izinganekwane ezingahlungiwe, ukuhlaziywa kwezokwelapha noma kwezomthetho okucophelele — lapho impendulo egwemayo imbi kunokungabi nayo.

Khetha okunye uma

• Udinga umbono. Le modeli ingombhalo kuphela; umzalwane wayo i-MiMo-V2.5 · W8A16 unombono ngentengo nesigaba esifanayo.
• Udinga ulwazi olubanzi kunazo zonke lomhlaba ezicelweni ezimfushane. Leso isimo lapho amapharamitha engeziwe esisekelo esikhulu kakhulu samabhithi amathathu enqoba futhi umonakalo waso wokukalwa kwezinga ubonakala kancane.
• Umsebenzi wakho ukuhlukanisa okumfushane ngobuningi obuphezulu. Umonakalo wokukalwa kwezinga awubonakali khona kalula — thenga noma yini eshesha kunazo zonke.

09Ukubiza i-API ye-MiMo engavinjelwe

Uma ikhodi yakho iqondise ku-OpenAI SDK, ukushintsha i-URL eyisisekelo nenqama yemodeli kuyikho konke okudingekayo ukufuduka:

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiMo-V2.5-Pro-W8A16",
    "messages": [
      {"role": "user", "content": "Draft the threat model section."}
    ],
    "reasoning_effort": "high",
    "max_tokens": 16384,
    "stream": true
  }'

U-/v1/messages no-/v1/responses bemukela inqama namakhono afanayo. Uhlu lwamaphalamitha, izinhlelo zokubiza amathuluzi, izibonelo zokukhishwayo okuhlelekile kanye nendawo yokudlala ephilayo kuse-mibhalweni ye-API ye-Shannon AI.

Ikhono	MiMo-V2.5-Pro · W8A16
Iwindi lomongo	amathokheni angu-262,144
Okukhishwayo okuphezulu ngesicelo	amathokheni angu-16,384
Ukusakaza	Yebo — zonke izilimi ezintathu
Ukubiza amathuluzi	Yebo
Ukucabanga / reasoning_effort	Yebo
Okukhishwayo okuhlelekile kwe-json_schema	Yebo
response_format	Yebo
Umbono / ukungenisa izithombe	Cha — sebenzisa i-MiMo-V2.5
Ungqimba lokwenqaba / isihlungi sokukhishwayo	Alukho
Intengo ngamathokheni ayisigidi	$0.50 engenayo / $2.00 ephumayo

10Imibuzo ebuzwa njalo

---

Imithombo: Umbiko wobuchwepheshe we-MiMo, arXiv:2505.07608 · Ikhadi lemodeli le-MiMo-V2.5-Pro (Xiaomi) · "Give Me BF16 or Give Me Death"?, arXiv:2411.02355 · REAP the Experts, arXiv:2510.13999 · AutoRound, arXiv:2309.05516 · Imithetho yokukhula kwe-k-bit inference, arXiv:2212.09720 · Scaling Laws for Precision, Kumar nabanye, ICLR 2025. Wonke amanani ezilinganiso anjengoba ebikwe ngababhali bawo. I-Shannon AI ingumkhiqizo we-Shannon Lab LLC, e-New Mexico, USA.