I-API ye-MiMo Engacinezelwanga Ethembeka kwiiBhithi eziSibhozo

Iikhathalogu ezininzi zikuthengisela inani leeparamitha. Lelo linani elingalunganga ekuthengeni ngalo, kuba yonke imodeli enkulu enobunzima obuvulekileyo onokuyifumana ngokwenene icinezelwe ukuze ilingane kwenye indawo. Umbuzo onyanisekileyo asingu "ibinkulu kangakanani phambi kokuba icudiswe?" koko ngu "kusekhona kangakanani kuyo?" Eli nqaku limalunga naloo mbuzo, nangemodeli kuluhlu lwethu apho impendulo iphantse ibe yiyo yonke.

01Ikunika ntoni ngokwenene i-API ye-MiMo engacinezelwanga

I-API ye-MiMo engacinezelwanga sisazisi semodeli esinye, i-MiMo-V2.5-Pro-W8A16, esifikelelekayo ngeelwimi ezintathu ezifikelela kubunzima obunye kwiqoqo lethu leeGPU: i-/v1/chat/completions, i-/v1/messages ne-/v1/responses. Zontathu ziyasasaza yaye zixhasa ukubiza izixhobo nokuqiqa.

Ukungacinezelwa kunentsingiselo emxinwa apha: akukho maleko wokwala phambi kwemodeli, akukho sihluzo kwimveliso yayo. Asithimbi, asahluli ngeendidi, asibhali kwakhona. Ufumana ukuziphatha kwesiseko ngokwaso — into ethetha kwimodeli enobunzima obuvulekileyo yodidi lophando ukuba ijongana nophando lokhuseleko, intsomi yabantu abadala okanye yobundlobongela, ukuqiqa kwezonyango nezomthetho, kunye nezicelo ii-endpoint zorhwebo ezibanjelweyo ezizala ngokungenangqondo. Sisixhobo sobungcali phantsi kwePolisi yethu yoSetyenziso oluNoxanduva, ingekuko ukungabikho kwemigaqo.

02Umnombo we-MiMo: ukubheja kwe-Xiaomi kuqeqesho lwangaphambili olubeka ukuqiqa kuqala

I-MiMo ayisiyiyo into ephuma kwi-Llama enegama elitsha; ngumnombo owahlukileyo ovela kwiqela le-MiMo le-Xiaomi, kwaye umbono waso osisiseko ubalulekile phambi kokuba uvavanye imodeli. Uqala nge-MiMo-7B (30 Epreli 2025, inelayisensi ye-MIT), ingxelo yayo yobuchwepheshe exoxa ukuba iilebhu ezininzi ziphatha ukuqiqa njengengxaki yasemva koqeqesho emayixotyelelwe nge-RL, nokuba oku kubuyela umva. I-Xiaomi yakha ngokutsha uqeqesho lwangaphambili ngokujikeleza kona endaweni yoko: umxube wamanqanaba amathathu kwiitoken ezisithriliyoni ezingama-25, inqanaba lesibini liqhuba imathematika nekhowudi malunga nama-70%, inqanaba lesithathu lisonga imikhondo yokuqiqa eyenziweyo ngoxa landisa umxholo ukusuka kwi-8K ukuya kwi-32K — kunye nenjongo yoKuqikelela iiToken eziNinzi esebenza njengentloko yokukhupha okuqikelelayo.

Isiphumo yayiyimodeli esisiseko ye-7B eyayiziphatha njengento enkulu kakhulu. I-MiMo-7B-Base yafumana ama-32.9 kwi-AIME 2024 apho i-Qwen2.5-7B yafumana ama-10.1, i-Gemma-2-9B ama-0.0 ne-Llama-3.1-8B ama-0.3. Emva koqeqesho kwaqhutywa i-RL kwiingxaki zemathematika nezekhowudi ezingama-130,000 ezinokuqinisekiswa ngomvuzo wekhowudi oqhutywa bubunzima bovavanyo, kuphunyuka kumgibe womvuzo osasazekileyo owonakalisa i-RL yekhowudi. Isiphumo safika phambi kwe-o1-mini:

I-AIME 2025, amanani axelwe yi-Xiaomi kwingxelo yobuchwepheshe ye-MiMo (arXiv:2505.07608). I-MiMo-7B-RL ikwaxele ama-95.8 kwi-MATH-500 nama-57.8 kwi-LiveCodeBench v5.

Usapho lwaya lwanda endaweni yokuphambuka: i-MiMo-VL-7B (Juni 2025) yabeka i-encoder ye-Qwen2.5-ViT kumqolo omnye, kwaye uDisemba 2025 wazisa i-MiMo-Audio-7B kunye nokungena kokuqala kwe-MoE, i-MiMo-V2-Flash (309B zizonke / 15B ezisebenzayo, iitoken ezingama-27T, ukuchaneka okuxutyiweyo kwe-FP8). I-MiMo-V2-Pro yalandela ngoMatshi 2026 kwi-1T malunga / 42B ezisebenzayo nomxholo weetoken ezisigidi, emva kweeveki kwi-OpenRouter njenge-"Hunter Alpha". I-MiMo-V2.5 ne-MiMo-V2.5-Pro zahamba kunye nge-22 Epreli 2026, zombini ngaphantsi kwe-MIT.

03Ngaphakathi kwe-MiMo-V2.5-Pro: sikhangeleka njani isiseko

Ngokwesikhokelo seModeli esipapashiweyo se-Xiaomi, i-MiMo-V2.5-Pro yimodeli ye-Mixture-of-Experts esasazekileyo eneeparamitha ezisithriliyoni esinye nezimbini kwekhulu zizonke nezingamashumi amane anesibini ezigidigidi ezisebenzayo kwitoken nganye kwiimaleko ezingama-70 — enye ishinyene, ezingama-69 ziyi-MoE — kunye neengcali ezingama-384 ezirawutwayo, ezisibhozo kwitoken nganye. Ubungakanani obufihlakeleyo bungama-6144, nobungakanani obuphakathi be-MoE obumxinwa obungama-2048.

Izigqibo ezibini zimilisa indlela evakala ngayo ekusebenziseni. Isitaki sengqwalasela sidibene: iimaleko ezilishumi zengqwalasela epheleleyo neemaleko ezingama-60 zefestile etyibilikayo ngefestile ye-128, kunye netyekelo elinokufundwa lokuzinza kwengqwalasela. I-Xiaomi ixela ukuba oku kunciphisa ugcino lwe-KV-cache lusondele kwizihlandlo eziyi-7 xa kuthelekiswa nengqwalasela yehlabathi ipheleleyo ngoxa igcina umgangatho womxholo omde — ngumahluko phakathi kwefestile ende ngamagama nefestile onokuyizalisa ngexabiso. Ikwahambisa iimaleko ezintathu zokuQikelela iiToken eziNinzi, zinika i-decoder intloko yokuqikelela.

Uqeqesho lwaba ziitoken ezisithriliyoni ezingama-27 kubude bolandelelwano lwemvelo lwe-32K kukuchaneka okuxutyiweyo kwe-FP8 (E4M3), emva koko kwalandela ukulungiswa okubekwe iliso, i-RL ye-agentic yobungakanani obukhulu kunye nokucuthwa kootitshala abaninzi ngomgaqo-nkqubo. Yibambe leyo nkcukacha ye-FP8; yeyona ngxoxo inamandla yokuhambisa esi siseko kwiibhithi ezisibhozo.

I-Xiaomi ixela imodeli kuma-57.2% kwi-SWE-bench Pro, kuma-72.9 kwi-τ3-Bench ngokusetyenziswa kwezixhobo nakuma-48.0% kwi-Humanity's Last Exam — amanani omthengisi ngokwakhe kwimodeli engacinezelwanga, acatshulwa njengoko.

Sihambisa ifestile yeetoken ezingama-262,144, ingekuko iitoken ezisigidi ezipheleleyo zesiseko — umxholo esinokuwuqinisekisa ngesantya, endaweni yenani eliphambili elehla phantsi komthwalo.

04Ithetha ntoni ngokuchanekileyo i-W8A16

Yonke imodeli kule khathalogu ithwala impumezelo yokucutha echaza uguqulo lokwenene, olulinganisekayo endaweni yenqanaba lokuthengisa. Ezethu yi-3BIT-REAP, i-W4A16-AUTOROUND-REAP ne-W8A16. I-MiMo-V2.5-Pro yeyesithathu.

Le nkcazo yahlulwe kabini. I-W8 kukuchaneka kobunzima: iparamitha nganye igcinwa njengenani elipheleleyo leebhithi ezisibhozo endaweni yenani elidadayo leebhithi ezilishumi elinesithandathu, ngefektha yokulinganisa eyabelwanayo kwiqela ngalinye elincinci ukuze uluhlu lwenguqu lwalapha lweqela ngalinye lusinde. I-A16 kukuchaneka kokusebenza: amaxabiso ahamba phakathi kweemaleko ngexesha lokuqikelela ahlala kwiibhithi ezilishumi elinesithandathu. Ubunzima buyacinezelwa; izibalo azicinezelwa.

Isiqingatha sesibini senza umsebenzi othe cwaka, obalulekileyo. Ukusebenza kunzima kakhulu ukukucutha kunobunzima: kuqulethe iitshaneli ezingaqhelekanga ezigqithisileyo — imilinganiselo embalwa ebungakanani bayo bugqiba konke okunye nethwala isabelo esingalinganiyo sokuziphatha kwemodeli. Zityhale ezo kukuchaneka okusezantsi uze unqumle kanye umqondiso obuwufuna. Amacebo obunzima kuphela ayayiphepha le ngxaki ngokushiya indlela yokusebenza ingachukunyiswa, nto leyo esisizathu sokuba umsebenzi wokucutha onzulu udibane kuwo.

Ngoko umbuzo osebenzayo uncipha ube yinto enye: zingaphi iibhithi ezimbalwa ubunzima obunokuzinyamezela?

05I-W8A16 ngokuchasene ne-W4A16 nakwiibhithi ezintathu: iya phi ngokwenene ilahleko

Olona phando lukhulu lupapashiweyo ngale nto lolwe-Neural Magic ne-Red Hat "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization — ngaphezu kovavanyo oluzizigidi ezingama-500,000 kusapho lwe-Llama-3.1 kuma-8B, 70B ne-405B. Isiphumo sabo esiphambili, ngamanani abo:

Ifomathi	Yintoni ecinezelweyo	Ukubuyiselwa okuxeliweyo ngokuchasene ne-BF16
W8A8-FP (iibhithi ezisibhozo)	Ubunzima + ukusebenza, okudadayo	~99.75% ngokuphakathi — "phantse akukho lahleko"
W8A8-INT (iibhithi ezisibhozo)	Ubunzima + ukusebenza, amanani apheleleyo	99–100% kwi-8B/70B; 98.2–99.3% kwi-405B
W4A16-INT (iibhithi ezine)	Ubunzima kuphela	98–99.4% (8B); 98.6–99.8% (70B)

Amanani axelwe yi-Neural Magic / Red Hat (arXiv:2411.02355) kwiimodeli ze-Llama-3.1 — ngumilo wentengiselwano, ingelilo inqaku lale modeli.

Xa kufundwa ngokunyanisekileyo, oko kubonakala njengengxoxo echasene nokukhathala: iibhithi ezine zibuyisa ama-98–99%, ngoko kutheni sinike ingqwalasela? Kuba ezo ziibalo-mgangatho zezifundo zokukhetha kwezininzi, ezikufuphi kwesona sixhobo sibi sikhoyo sokufumanisa umonakalo wokucutha. Umbuzo wobalo-mgangatho unempendulo enye echanekileyo efikelelekayo kwitoken enye okanye ezimbini, kwaye impazamo yokucutha incinci, iyingxolo ephantse ibe zero-mean kuwo wonke umqondiso; kumbuzo omfutshane onomda obanzi, loo ngxolo iyafunxwa.

Ngoku bhala iitoken ezingama-4,000 zoxwebhu lobuchwepheshe, ubambe ilizwi kwibali leetoken ezingama-40,000, okanye ukhuphe i-JSON ekufuneka iqinisekiseke okokuqala. Ingxolo enye ithathwa amawaka amaxesha ngokulandelelana, isampuli nganye imisela elandelayo. Indlela yokusilela ayisiyompendulo engachanekanga — kukuphambuka: ukukhululeka kokubamba umyalelo ukususela kumjikelo wokuqala, umlinganiswa oyeka ukuvakala njengaye malunga netoken ye-3,000, ibrakethi yokuvala engasoze ifike.

Ziya phi ngokwenene iibhithi

Impazamo yokucutha ayisasazwa ngokulinganayo kwizakhono zemodeli; igxila kumsila. Ukuziphatha kobunzima obuphezulu — isiNgesi esinegrama, umilo womsebenzi we-Python, iinyaniso eziliwaka eziphambili ngehlabathi — kukhowudwe ngokuphindaphindeneyo kubunzima obuninzi kwaye kusinda phantse kuyo yonke into. Iikona zokusasazeka kobunzima ezitunwe ngokuchanekileyo, ezibungakanani obuncinci zibamba izinto ezinqabileyo: uphawu oluchanekileyo lwe-API engaziwayo, isingqi sombhali othile, umahluko phakathi kwee-CVE ezimbini ezimelene. Lawo ngamaxabiso igridi engacokisekanga ejikelezela kunye.

Inyathelo elisuka kwiibhithi ezine ukuya kwezintathu kulapho oku kuyeka ukuba yirhafu kube liliwa. Imithetho yobungakanani bokuqikelela kweebhithi ezingu-k ka-Dettmers no-Zettlemoyer yafumanisa iibhithi ezine njengendawo yokwandisa ukuchaneka kwesitshotsho esingekho kwibhithi nganye yenkumbulo yemodeli, ngokwehla okulawulayo ngezantsi kwayo. Iibhithi ezine zinika amanqanaba ali-16 kwiqela ngalinye; ezintathu zinika asibhozo. Wenza isiqingatha sesisombululo segridi esele ingacokisekanga, kwaye impazamo yokujikeleza kubunzima ngabunye iphantse iphindwe kabini.

Izibalo ezikugqibayo

Uthelekiso olufanele ukwenziwa zizibalo, ingekuko ubalo-mgangatho. Into esindayo kukucinezelwa iphantse ibe ziiparamitha × iibhithi kwiparamitha nganye — umthamo othwala ulwazi wefayile yobunzima. Thatha iziseko ezibini ezingacingelwayo zomgangatho olinganayo kwiparamitha nganye:

• Isiseko A: iiparamitha ezingama-400B kwiibhithi ezintathu → malunga nee-GB ezili-150 zobunzima obugciniweyo.
• Isiseko B: iiparamitha ezingama-160B kwiibhithi ezisibhozo → malunga nee-GB ezili-160 zobunzima obugciniweyo.

Ngenani leeparamitha elikrwada u-A uphumelela ngesibini nesiqingatha kwesinye — inani isikhokelo seModeli esiqala ngalo. Ngolwazi olugciniweyo ziyalingana, kwaye u-B unenzuzo yesakhiwo ngaphezulu: ubunzima bakhe buhlala kwigridi ecokisekileyo ngokwaneleyo ukuze ukuziphatha komsila wakhe kuphelele, ngoxa umsila ka-A ucuthwe waya kubuninzi. U-A wazi izinto ezininzi; u-B wazi into ayaziyo ngokucacileyo ngakumbi. Ngokuvelisa okude, ifomathi echanekileyo, ulawulo lwesimbo nokubambelela okuthembekileyo kumxholo omde, thatha u-B.

Zizibalo zomzekelo, ingekuko ubalo-mgangatho: zibonisa ukuba "isiseko esikhulu" kunye "nemodeli ehanjiswa ngcono" ngamabango ahlukeneyo.

06Kutheni iitoken ezisithriliyoni ezingama-27 zoqeqesho zisenza ukucutha okunamandla kube kubi ngakumbi

Isiphumo sesibini, esingacacanga ngokulula, sithe ngqo kwesi siseko. Iphepha le-ICLR 2025 lika-Kumar nabanye Scaling Laws for Precision lifumanisa ukuba okukhona imodeli iqeqeshwe kwidatha eninzi, kokukhona ukucutha kwasemva koqeqesho kwenza umonakalo omkhulu kuyo. Ukuchaneka okusezantsi kunciphisa inani leeparamitha ezisebenzayo zemodeli, kwaye oko kwehla kukhula ngedatha yoqeqesho lwangaphambili efunxiweyo — kude kangangokuba idatha eyongezelelweyo yoqeqesho lwangaphambili inokuba yingozi ekugqibeleni ukuba unenjongo yokucutha ngamandla, kuba ubeka ubume obucokisekileyo igridi eza kubucima.

Imodeli eqeqeshwe kwiitoken ezingama-5T inobume bobunzima obungacokisekanga, obuphindaphindeneyo nendawo eshiyekileyo. Imodeli eqeqeshwe kwiitoken ezingama-27T — uhlahlo-lwabiwo-mali lwe-MiMo-V2.5-Pro — yachitha loo datha yongezelelweyo ibhala umahluko ocokisekileyo kumaxabiso obunzima achanekileyo ngakumbi, into kanye igridi yeebhithi ezintathu engenako ukuyimela.

Emva koko kukho inyaniso ye-FP8 evela kwicandelo 03, eyona ngxoxo icoceke kweli nqaku. I-MiMo-V2.5-Pro yaqeqeshwa kwangaphambili kukuchaneka okuxutyiweyo kwe-FP8. Ubunzima bayo bamilwa kwasekuqaleni yi-optimizer eyayisebenza malunga neebhithi ezisibhozo zokuchaneka. Ukuyihambisa kubunzima beebhithi ezisibhozo ngoko akusiyocinga lasemva elilahlekelwayo elinyanzeliswe kwimodeli yeebhithi ezilishumi elinesithandathu — kukufuphi ekubuyiseleni ubunzima kwimeko yamanani obafundwa kuyo. Bubume obucokisekileyo obuncinci obukhoyo ngezantsi kwebhithi yesibhozo obungalahlekayo. Loo nto ayinakuthethwa ngobo bunzima bunyanzelwa buye kwezintathu.

07REAP: kutheni ukuthenwa kweengcali kuhlobo olwahlukileyo lwelahleko kunokujikeleza

Amabini kula manqanaba mathathu athwala impumezelo ye-REAP — umsebenzi owahlukileyo ngokupheleleyo kukucutha. I-REAP (Router-weighted Expert Activation Pruning, Cerebras Research, arXiv:2510.13999, ICLR 2026) kukucinezela kwesitshotsho esinye kwiimodeli ze-MoE ezisasazekileyo: inika inqaku ingcali nganye ngamaxabiso egeyithi yerawuta nomlinganiselo wokusebenza kwengcali, ize icime ezinamanqaku asezantsi. I-Cerebras ixela iziphumo ezingaphantse zibe nalahleko ekuveliseni ikhowudi nasekubizeni izixhobo nakuma-50% eengcali ezithenwe.

Sisebenzisa i-REAP ngenxa yesizathu sokubakho kwayo: kwezona ziseko zinkulu yiyo eyenza zibe nako ukuhanjiswa. Kodwa qaphela umahluko wodidi. Ukucutha kufiphaza ubunzima ngabunye kancinci; ukuthena kususa iingcali ngokupheleleyo. Ukuba umsebenzi wakho urawuta kakhulu kwisakhono ebesihlala kwingcali ekhethwa kanqabe — ulwimi lwenkqubo oluncinci, isigama esikhethekileyo — asehlanga. Sishiyile, kwaye akukho kuyalela okusibuyisayo. Iziqingatha zihlala zilungile; ityala lakho lomsila lisenokungahlali lilungile.

I-MiMo-V2.5-Pro · W8A16 ayithwali mpumezelo ye-REAP. Zonke iingcali ezingama-384 zikhona, kwaye irawuta ikhetha kwiseti epheleleyo isiseko esaqeqeshwa ngayo.

Inqanaba	Iibhithi zobunzima	Iingcali	Utshintshisa ntoni
W8A16	8	Zonke zikhona	Akukho nto yesakhiwo. Linqanaba lokuthembeka.
W4A16-AUTOROUND-REAP	4	Zithenwe	Iinkcukacha ezicokisekileyo zobunzima kunye neengcali ezinomsinga osezantsi, endaweni yesiseko esikhulu kakhulu kwibhendi enye yexabiso.
3BIT-REAP	3	Zithenwe	Isisombululo esikhulu sobunzima kunye neengcali ezinomsinga osezantsi — ixabiso lokuqhuba ezona ziseko zinkulu ngesantya.

Inqanaba eliphakathi lifanele ukuviwa ngokufanelekileyo. I-AutoRound ayisikho ukujikeleza okungenabuchule: indlela ye-Intel (arXiv:2309.05516) isebenzisa ukwehla kwesantya okunophawu kumanyathelo amalunga nangama-200 ukutuna izigqibo zokujikeleza kunye noluhlu lokunqumla kwiqela ngalinye lobunzima, kwaye ixelwa njengephantse ingabi nalahleko kwiibhithi ezine kwiimeko ezininzi, ibetha i-GPTQ, i-AWQ, i-HQQ ne-OmniQuant kwiicwangciso ezininzi ezivavanyiweyo. Imodeli ye-W4A16-AUTOROUND yimodeli elungileyo ngokwenene yeebhithi ezine. Isengumodeli yeebhithi ezine noko.

08Ngubani ofanele ukukhetha inqanaba leebhithi ezisibhozo — nongafanelekanga

I-MiMo-V2.5-Pro · W8A16 ihlala kwixabiso elisezantsi lekhathalogu iphela — $0.50 engenayo / $2.00 ephumayo kwitoken enye yezigidi, ibhendi enye neemodeli zenqanaba lokungena zeebhithi ezine kwaye isisiqingatha sesine malunga sexabiso lezingenayo lemodeli ebiza kakhulu esiyihambisayo. Awuhlawuleli ixabiso eliphezulu ngokuthembeka; ukhetha umzila owahlukileyo wohlahlo-lwabiwo-mali olunye.

Yikhethe xa

• Ubude bemveliso ngumsebenzi. Ukubhala kobuchwepheshe okude, intsomi yobude besahluko, udlulo lwamaxwebhu apheleleyo — itoken nganye eveliswayo yenye apho ingxolo yokucinezela iqokelela khona.
• Ilizwi nokuhambelana kubalulekile. Umsebenzi womlinganiswa, isandi sophawu lorhwebo, isimo esibanjwa kwiseshoni ende. Isimbo sihlala kumsila, kwaye umsila yinto ukucutha okungacokisekanga okuyitya kuqala.
• Imveliso ecwangcisiweyo kufuneka iqinisekiseke okokuqala. Imodeli ixhasa i-json_schema ne-response_format; ukubambelela kwisikimu yimpawu yokuchaneka, kwaye ukuzama kwakhona ngakunye okuphetshiweyo yimali nexesha lokulinda elongiweyo.
• Uzalisa umxholo. Ukubambelela kumyalelo kwitoken ye-200,000 kunzima kunakwitoken ye-2,000, kwaye kuyehla ngokuchaneka kobunzima — kunjalo nangomkhondo omde wokuqiqa, apho iimpazamo ziqokelela ngendlela efanayo.
• Umsebenzi awucinezelwanga ngokwemvelo. Uphando lokhuseleko, ukusebenza kweqela elibomvu, intsomi engahluzwanga, uhlalutyo lonyango okanye lwezomthetho olubuthathaka — apho impendulo ethandabuzayo imbi ngaphezu kokungabikho kwayo.

Khetha enye into xa

• Ufuna umbono. Le modeli ngumbhalo kuphela; udade wayo i-MiMo-V2.5 · W8A16 inombono kwixabiso elinye nakwinqanaba elinye.
• Ufuna ulwazi lwehlabathi olubanzi kunazo zonke kwimibuzo emifutshane. Yiloo meko apho iiparamitha ezongezelelweyo zesiseko esikhulu kakhulu seebhithi ezintathu ziphumelela khona kwaye umonakalo wazo wokucutha ubonakala kancinci.
• Umsebenzi wakho kukwahlula ngeendidi okumfutshane komthamo omkhulu. Umonakalo wokucutha uphantse ungabonakali apho — thenga nayiphi na ekhawuleza kunazo zonke.

09Ukubiza i-API ye-MiMo engacinezelwanga

Ukuba ikhowudi yakho ijolise kwi-SDK ye-OpenAI, ukutshintsha i-URL esisiseko nesazisi semodeli lolo lonke ufuduko:

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
  }'

I-/v1/messages ne-/v1/responses zamkela isazisi esinye nezakhono ezinye. Ireferensi yeeparamitha, izikimu zokubiza izixhobo, imizekelo yemveliso ecwangcisiweyo nendawo yokudlala ephilayo kukwi-amaxwebhu e-API ye-Shannon AI.

Isakhono	MiMo-V2.5-Pro · W8A16
Ifestile yomxholo	iitoken ezingama-262,144
Imveliso ephezulu kwisicelo ngasinye	iitoken ezingama-16,384
Ukusasaza	Ewe — kuzo zontathu iilwimi
Ukubiza izixhobo	Ewe
Ukuqiqa / reasoning_effort	Ewe
Imveliso ecwangcisiweyo ye-json_schema	Ewe
response_format	Ewe
Umbono / umfanekiso ongenayo	Hayi — sebenzisa i-MiMo-V2.5
Umaleko wokwala / isihluzo semveliso	Akukho
Ixabiso kwitoken enye yezigidi	$0.50 engenayo / $2.00 ephumayo

10Imibuzo ebuzwa rhoqo

---

Imithombo: Ingxelo yobuchwepheshe ye-MiMo, arXiv:2505.07608 · Isikhokelo seModeli ye-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 yobungakanani bokuqikelela kweebhithi ezingu-k, arXiv:2212.09720 · Scaling Laws for Precision, Kumar nabanye, ICLR 2025. Onke amanani obalo-mgangatho anjengoko exelwe ngababhali bawo. I-Shannon AI yimveliso ye-Shannon Lab LLC, New Mexico, USA.