API yeMiMo Isina Kudzorwa paKutendeseka kwe8-Bit

Makatarogu mazhinji anokutengesera huwandu hwemapharamita. Ndiyo nhamba isiriyo yekutenga nayo, nokuti modhi yega yega huru ine huremu hwakavhurika yaunogona kutenga chaizvo yakadzvanywa kuti ikwane pane imwe nzvimbo. Mubvunzo wechokwadi hausi wekuti "yaiva huru sei isati yadzvanywa?" asi ndewekuti "chikamu chakadii chayo chichiripo?" Chinyorwa chino chiri pamubvunzo iwoyo, uye pamodhi iri mumutsara wedu apo mhinduro iri chikamu chikuru chayo.

01Chaunowana chaizvo paAPI yeMiMo isina kudzorwa

API yeMiMo isina kudzorwa izita remodhi rimwe chete, MiMo-V2.5-Pro-W8A16, rinosvikwa kuburikidza nemitauro mitatu inosvika pahuremu humwe chete pacluster yedu yemaGPU: /v1/chat/completions, /v1/messages ne/v1/responses. Mitatu yose inoyerera uye inotsigira kudana zvishandiso nekufungisisa.

Kusadzorwa kune chirevo chakamanikana pano: hapana chidzitiro chekuramba mberi kwemodhi, hapana sefa pane zvinobuda zvayo. Hatibati, hatirongedzi, hatinyorezve. Unowana maitiro ehwaro pachahwo — izvo pamodhi ine huremu hwakavhurika yechiyero chetsvakurudzo zvinoreva kuti inobata tsvakurudzo yezvekuchengetedza, ngano dzevakuru kana dzine chisimba, kufungisisa kwezvekurapa nezvepamutemo, nezvikumbiro zvinorambwa nesuwo rekutengeserana zvisina kufunga. Chishandiso cheunyanzvi pasi peMutemo wedu Wekushandisa Nemutsigo, kwete kushaikwa kwemitemo.

02Dzinza reMiMo: kutenda kwaXiaomi mukudzidzisa kunotanga nekufungisisa

MiMo haisi chinhu chakatorwa kubva kuLlama chikatumidzwa zita idzva; idzinza rakasiyana rinobva kuchikwata cheMiMo cheXiaomi, uye pfungwa yaro yekuvamba inokosha usati waongorora modhi. Rinotanga neMiMo-7B (30 Kubvumbi 2025, ine rezinesi reMIT), gwaro rayo reunyanzvi richiti marabhoritari mazhinji anotora kufungisisa sedambudziko rekudzidziswa kwemashure rekunamirwa neRL, uye kuti izvozvo zvakadzokera shure. Xiaomi yakarongazve kudzidziswa kwepakutanga ichitenderedza izvozvo panzvimbo pezvo: musanganiswa wemadanho matatu pamatokeni matiriyoni makumi maviri nemashanu, danho rechipiri richimhanya anenge 70% emasvomhu nekodhi, danho rechitatu richipeta mukati tsimba dzekufungisisa dzekunyepedzera richiwedzera context kubva pa8K kuenda pa32K — pamwe nechinangwa cheMulti-Token Prediction chinoshandawo semusoro we speculative decoding.

Chakabuda ndeyekuti modhi yepasi ye7B yakaita sechinhu chihombe zvikuru. MiMo-7B-Base yakawana 32.9 paAIME 2024 apo Qwen2.5-7B yakawana 10.1, Gemma-2-9B 0.0 neLlama-3.1-8B 0.3. Kudzidziswa kwemashure kwakabva kwamhanyisa RL pamatambudziko 130,000 anogona kutariswa emasvomhu nekodhi ine mubairo wekodhi unotungamirwa nekuoma kwemuedzo, ichipukunyuka mumusungo wemibairo mishoma unozunza RL yekodhi. Chakabuda chakamira pamberi peo1-mini:

AIME 2025, nhamba dzakashumwa naXiaomi kubva mugwaro reunyanzvi reMiMo (arXiv:2505.07608). MiMo-7B-RL yakashumawo 95.8 paMATH-500 ne57.8 paLiveCodeBench v5.

Mhuri yakafara pane kutsauka: MiMo-VL-7B (Chikumi 2025) yakaisa muisi weQwen2.5-ViT pahwaro humwe chetehwo, uye Zvita 2025 yakaunza MiMo-Audio-7B pamwe nechinhu chekutanga cheMoE, MiMo-V2-Flash (309B ose / 15B anoshanda, matokeni 27T, FP8 mixed precision). MiMo-V2-Pro yakatevera muna Kurume 2026 pa1T / 42B anoshanda ine context ye1M, mushure memavhiki paOpenRouter se"Hunter Alpha". MiMo-V2.5 neMiMo-V2.5-Pro dzakabuda pamwe chete musi wa22 Kubvumbi 2026, dzose dzine MIT.

03Mukati meMiMo-V2.5-Pro: hwaro hwakaita sei

Maererano nekadhi remodhi rakabudiswa naXiaomi, MiMo-V2.5-Pro imodhi yeMixture-of-Experts yakapararira ine mapharamita matiriyoni 1.02 ose ne mabhiriyoni 42 anoshanda patokeni imwe neimwe pamadanho 70 — rimwe rakazara, 69 eMoE — nenyanzvi 384 dzinotumirwa, 8 patokeni imwe neimwe. Hukuru hwakavanzika i6144, hune hukuru hwepakati hweMoE hwakamanikana hwe2048.

Sarudzo mbiri dzinotara kuti inonzwika sei pakushandiswa. Chimiro cheattention chakasanganiswa: madanho 10 eattention yakazara nemadanho 60 ehwindo rinotsvedza rine hwindo re128, pamwe nebhaizi rinodzidzwa reattention-sink. Xiaomi inoshuma kuti izvi zvinodimbura kuchengetwa kweKV-cache kusvika pedyo neka7 kuenzaniswa neattention yakazara yose apo ichichengeta mhando yecontext refu — musiyano uri pakati pehwindo rakareba pazita nerimwe raunogona kukwanisa kuzadza. Inouyawo nemadanho matatu eMulti-Token Prediction, ichipa dhikodha musoro we speculative.

Kudzidziswa kwakaitwa pamatokeni matiriyoni makumi maviri nemanomwe pakureba kwemutsara kwepachayo kwe32K muFP8 (E4M3) mixed precision, ipapo kugadziriswa kunotungamirwa, RL yemaagent yechiyero chikuru, nekudzidziswa kwekutevedzera kwevadzidzisi vakawanda. Chengeta chinhu ichi cheFP8; ndidzo nharo dzakasimba kupfuura dzose dzekushandisa hwaro uhu pamabhiti masere.

Xiaomi inoshuma modhi pa57.2% paSWE-bench Pro, 72.9 paτ3-Bench pakushandisa zvishandiso ne48.0% paHumanity's Last Exam — nhamba dzemutengesi pachake pamodhi isina kudzvanywa, dzakataurwa dzakadaro.

Tinopa hwindo rematokeni 262,144, kwete 1M yakazara yehwaro — context yatinogona kuvimbisa pakumhanya, pane nhamba huru inodzikira pasi pemutoro.

04Kuti W8A16 inorevei, chaizvo

Modhi yega yega mukatarogu iyi ine chipfupiso che quantization chinotsanangura shanduko chaiyo, inogona kuyerwa pane mhando yeshambadziro. Zvedu ndi3BIT-REAP, W4A16-AUTOROUND-REAP neW8A16. MiMo-V2.5-Pro ndeyechitatu.

Chinyorwa chinokamurwa kaviri. W8 ndiko kururama kwehuremu: pharamita imwe neimwe inochengetwa se integer ye8-bit panzvimbo pefloat ye16-bit, ine chiyero chinogovaniswa paboka diki rimwe nerimwe kuti nzvimbo yeboka rimwe nerimwe irambe iripo. A16 ndiko kururama kwemaactivation: zvikoshwa zvinofamba pakati pemadanho panguva yekushandisa zvinosara pamabhiti gumi nematanhatu. Huremu hunodzvanywa; masvomhu haadzvanywi.

Chikamu chechipiri ichocho chinoita basa rakanyarara, rinokosha. Maactivation akaoma kuita quantization kupfuura huremu: ane makanari ane zvinhu zvisingajairiki — mativi mashoma ane hukuru hunopfuura zvimwe zvose uye anotakura chikamu chikuru chemaitiro emodhi. Aendese pakururama kwakadzikira uye unocheka chaizvo chiratidzo chawaida. Zvirongwa zvehuremu chete zvinonzvenga dambudziko iri nekusiya nzira yemaactivation isina kubatwa, ndicho chikonzero basa rakakosha re quantization richiungana pazviri.

Saka mubvunzo unoshanda unomanikana kuita chinhu chimwe chete: mabhiti mashoma sei anogona kutakurwa nehuremu?

05W8A16 kupesana neW4A16 kupesana ne3-bit: kunoenda kurasikirwa chaiko

Ongororo huru kupfuura dzose yakabudiswa pane izvi ndeyeNeural Magic neRed Hat "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization — inopfuura ongororo 500,000 pamhuri yeLlama-3.1 pa8B, 70B ne405B. Chinobuda chikuru chavo, munhamba dzavo:

Chimiro	Chinodzvanywa	Kudzoserwa kwakashumwa kuenzaniswa neBF16
W8A8-FP (8-bit)	Huremu + maactivation, float	~99.75% pakati — "hakuna chinorasika chaizvo"
W8A8-INT (8-bit)	Huremu + maactivation, integer	99–100% pa8B/70B; 98.2–99.3% pa405B
W4A16-INT (4-bit)	Huremu chete	98–99.4% (8B); 98.6–99.8% (70B)

Nhamba dzakashumwa neNeural Magic / Red Hat (arXiv:2411.02355) pamamodhi eLlama-3.1 — chimiro chekuchinjana, kwete chiyero chemodhi iyi.

Kana zvikaverengwa nechokwadi, izvozvo zvinoita senharo dzinopikisa kuzvikoshesa: 4-bit inodzosera 98–99%, saka sei uchizvitarisa? Nokuti iyo imiyero yezvikoro yemibvunzo yekusarudza, iri pedyo nechishandiso chakaipa kupfuura zvose chekuona kukuvara kwe quantization. Mubvunzo wemuyero une mhinduro imwe chete yakarurama inosvikwa mumatokeni mamwe kana maviri, uye chikanganiso che quantization chidiki, chinenge chisina kutsauka pane logit yega yega; pamubvunzo mupfupi une mukaha mukuru, ruzha irworwo runonwiwa.

Zvino nyora matokeni 4,000 egwaro reunyanzvi, bata izwi rimwe pangano yematokeni 40,000, kana uburitse JSON inofanira kuongororwa pakutanga. Ruzha rumwe chetorwo runotorwa zviuru zvenguva zvakateedzana, chikamu chega chega chichitara chinotevera. Nzira yekukundikana haisi mhinduro isiriyo — ndeyekutsauka: kubata murairo wekutanga kuchisunungurwa, munhu anorega kunzwika saiye kuma tokeni 3,000, bhureketi rekuvhara risingasviki.

Kunoenda mabhiti chaiko

Chikanganiso che quantization hachipararirwi zvakaenzana pahunyanzvi hwemodhi; chinoungana mumuswe. Maitiro ane huremu hukuru — Chirungu chine giramba, chimiro chebasa rePython, zvokwadi chiuru chekutanga pamusoro penyika — akanyorwa kakawanda pahuremu hwakawanda uye anopona pane zvinenge zvose. Mativi akanyatsogadziriswa, ane hukuru hushoma ekugoverwa kwehuremu ndiwo anochengeta zvinhu zvisingawanzoitiki: chimiro chaicho cheAPI isingazivikanwi, chimiro chemunyori akati, musiyano uri pakati pemaCVE maviri ari pedyo. Ndizvo zvikoshwa zvinotenderedzwa zvichisanganiswa negirid rakaomarara.

Danho kubva pamabhiti mana kuenda kumatatu ndipo pazvinorega kuva mutero zvichiita mawere. Mitemo yechiyero yekushandisa mabhiti yaDettmers naZettlemoyer yakawana kuti mabhiti mana ndiwo poinzi inowedzera kururama kwezero-shot pabhiti rimwe nerimwe rememori yemodhi, kudzikira kuchitonga pasi payo. Mabhiti mana anopa mazinga 16 paboka; matatu anopa 8. Uri kudimbura nepakati chiyero cheresolution yegirid rakatoomarara, uye chikanganiso chekutenderedza pahuremu humwe nehumwe chinoda kupetwa kaviri.

Masvomhu anosarudza

Kuenzanisa kwakakodzera kuitwa ndeimasvomhu, kwete muyero. Chinopona pakudzvanya chinenge mapharamita × mabhiti papharamita — hukuru hwefaira rehuremu hunotakura ruzivo. Tora hwaro huviri hwakafungidzirwa hwemhando yakaenzana papharamita imwe neimwe:

• Hwaro A: mapharamita 400B pamabhiti matatu → anenge 150 GB ehuremu hwakachengetwa.
• Hwaro B: mapharamita 160B pamabhiti masere → anenge 160 GB ehuremu hwakachengetwa.

Nehuwandu hwemapharamita chaihwo A inokunda nekaviri nehafu — nhamba inotanga pakadhi remodhi. Neruzivo rwakachengetwa dzakaenzana, uye B ine kunaka kwechimiro pamusoro peizvozvo: huremu hwayo huri pagirid rakatsetseka zvekuti maitiro ayo emuswe akachengetedzeka, apo muswe waA wakaitwa quantization kuenda muchikamu chikuru. A inoziva zvinhu zvakawanda; B inoziva zvainoziva zvakapinza. Pakugadzira kwenguva refu, chimiro chakarurama chaicho, kudzora kwechimiro nekutevedzera kwechokwadi kwecontext refu, tora B.

Masvomhu ekuratidzira, kwete muyero: anoratidza kuti sei "hwaro hukuru" ne"modhi inoshandiswa iri nani" ari matauriro akasiyana.

06Sei matokeni matiriyoni makumi maviri nemanomwe ekudzidzisa achiipisa quantization yakaomarara

Chimwe chinowanikwa chechipiri, chisinganzwisisiki nyore, chinobata hwaro uhu zvakananga. Gwaro reICLR 2025 raKumar nevamwe Scaling Laws for Precision rinowana kuti pakawanda dhata rakadzidziswa modhi, ndipo quantization yemushure mekudzidzisa inonyanya kuikuvadza. Kururama kwakadzikira kunodzikisa huwandu hunoshanda hwemapharamita hwemodhi, uye kudzikira ikoko kunokura nedhata rekudzidziswa kwepakutanga rakanwiwa — kusvika pakuti dhata rekuwedzera rekudzidziswa rinogona kukuvadza kana uchida kuita quantization yakaomarara, nokuti uri kuisa chimiro chakatsetseka chichadzimwa negirid.

Modhi yakadzidziswa pamatokeni 5T ine chimiro chehuremu chakaomarara, chinodzokororwa uye chine nzvimbo yakasara. Modhi yakadzidziswa pamatokeni 27T — bhajeti raMiMo-V2.5-Pro — yakashandisa dhata rekuwedzera iroro ichinyora misiyano yakatsetseka muzvikoshwa zvehuremu zvakanyatsotarwa, ndizvo chaizvo zvisingakwanisi kuratidzwa negirid re3-bit.

Ipapo pane chokwadi cheFP8 chinobva muchikamu 03, nharo dzakachena kupfuura dzose muchinyorwa chino. MiMo-V2.5-Pro yakadzidziswa pakutanga muFP8 mixed precision. Huremu hwayo hwakaumbwa kubva pakutanga neoptimizer yaishanda pakururama kwemabhiti anenge masere. Kuishandisa pahuremu hwe8-bit naizvozvo hakusi kufunga kwemashure kunorasikirwa kwakaiswa pamodhi ye16-bit — kuri pedyo nekudzosera huremu kunzvimbo yenhamba yahwakadzidzirwa. Hapana chimiro chakatsetseka chakawanda pasi pebhiti rechisere chekurasikirwa. Zvimwe chetezvo hazvingataurwi pamusoro pehuremu hwakamanikidzirwa pasi kusvika kumatatu.

07REAP: sei kubvisa nyanzvi kuri rudzi rwekurasikirwa rwakasiyana nekutenderedza

Mbiri pamhando nhatu dzine chipfupiso cheREAP — basa rakasiyana zvachose ne quantization. REAP (Router-weighted Expert Activation Pruning, Cerebras Research, arXiv:2510.13999, ICLR 2026) ndiko kudzvanya kwekamwe chete kwemamodhi eMoE akapararira: inopa nyanzvi imwe neimwe chiyero maererano nezvikoshwa zvegedhi rerouter nechiyero chekushanda kwenyanzvi, ipapo yobvisa dziri pasi. Cerebras inoshuma mhedzisiro isingarasikirwi zvakanyanya pakugadzira kodhi nekudana zvishandiso kunyange pa50% yenyanzvi dzabviswa.

Tinoshandisa REAP nechikonzero chaakaitirwa: pahwaro hukuru kupfuura hwose ndicho chinoita kuti hukwanise kushandiswa. Asi ona musiyano werudzi. Quantization inovhundutsa huremu humwe nehumwe zvishoma; kubvisa kunobvisa nyanzvi zvachose. Kana basa rako richinyanya kutumirwa kuunyanzvi hwaigara munyanzvi isingawanzosarudzwa — mutauro wekunyora purogiramu usinganyanyi kuzivikanwa, mazwi enyanzvi — hahuna kudzikiswa. Hwapera, uye hapana prompt inohudzosera. Maavheji anosara akanaka; nyaya yako yemuswe inogona kusadaro.

MiMo-V2.5-Pro · W8A16 haina chipfupiso cheREAP. Nyanzvi dzose 384 dziripo, uye router inosarudza kubva paseti yakazara yakadzidziswa nayo hwaro.

Mhando	Mabhiti ehuremu	Nyanzvi	Chaunochinjanisa
W8A16	8	Dzose dziripo	Hapana chechimiro. Ndiyo mhando yekutendeseka.
W4A16-AUTOROUND-REAP	4	Dzakabviswa	Ruzivo rwakatsetseka rwehuremu pamwe nenyanzvi dzisinganyanyi kushandiswa, uchipiwa hwaro hukuru zvikuru mubhandi remutengo rimwe chete.
3BIT-REAP	3	Dzakabviswa	Chikamu chikuru cheresolution yehuremu pamwe nenyanzvi dzisinganyanyi kushandiswa — mutengo wekumhanyisa hwaro hukuru chaizvo nekukurumidza.

Mhando yepakati inokodzera kunzwikwa nemazvo. AutoRound haisi kutenderedza kwakapusa: nzira yeIntel (arXiv:2309.05516) inoshandisa signed gradient descent pamatanho anenge mazana maviri kugadzirisa sarudzo dzekutenderedza nemuganhu wekucheka paboka rimwe nerimwe rehuremu, uye inoshumwa seisingarasikirwi zvakanyanya pamabhiti mana mumamiriro mazhinji, ichikunda GPTQ, AWQ, HQQ neOmniQuant muzvirongwa zvakawanda zvakaedzwa. Modhi yeW4A16-AUTOROUND imodhi ye4-bit yakanaka chaizvo. Ichiri modhi ye4-bit.

08Ndiani anofanira kusarudza mhando ye8-bit — uye ndiani asingafaniri

MiMo-V2.5-Pro · W8A16 iri pamutengo wepasi wekatarogu yose — $0.50 anopinda / $2.00 anobuda pamatokeni miriyoni, bhandi rimwe chete nemamodhi e4-bit emhando yekutanga uye chikamu chinenge chechina chemutengo wezvinopinda wemodhi inodhura kupfuura dzose yatinoshandisa. Hausi kubhadhara mutengo wepamusoro nekuda kwekutendeseka; uri kusarudza rimwe divi rebhajeti rimwe chetero.

Isarudze kana

• Kureba kwezvinobuda ndiro basa. Kunyora kweunyanzvi kwenguva refu, ngano dzakareba sezvitsauko, kupfuura kwakazara kwemagwaro — tokeni yega yega inogadzirwa ndipo pane ruzha rwe quantization runowedzerana.
• Izwi nekufanana zvinokosha. Basa rehunhu, chimiro chebhurandi, mutauro unobatwa pamusangano wakareba. Chimiro chinogara mumuswe, uye muswe ndiwo unotanga kudyiwa ne quantization yakaomarara.
• Zvinobuda zvakarongwa zvinofanira kuongororwa pakutanga. Modhi inotsigira json_schema neresponse_format; kuteerera schema chinhu chekururama, uye kuedzazve kwega kwega kwakadzivirirwa imari nekunonoka kwakachengetedzwa.
• Uri kuzadza context. Kuteerera mirairo patokeni 200,000 kwakaoma kupfuura patokeni 2,000, uye kunodzikira nekururama kwehuremu — sezvinoita tsimba refu rekufungisisa, apo zvikanganiso zvinowedzerana zvimwe chetezvo.
• Basa harina kudzorwa nechisikigo. Tsvakurudzo yezvekuchengetedza, mapoka anopikisa, ngano dzisina sefa, ongororo yezvokurapa kana yepamutemo inonetsa — apo mhinduro yakavharwa yakaipa kupfuura kusava nemhinduro.

Sarudza chimwe kana

• Uchida kuona. Modhi iyi ndeyemavara chete; hanzvadzi yayo MiMo-V2.5 · W8A16 inoona pamutengo mumwe chete nemhando imwe cheteyo.
• Uchida ruzivo rwenyika rwakafara kupfuura rwose pamibvunzo mipfupi. Ndiyo nzvimbo apo mapharamita ekuwedzera ehwaro hukuru zvikuru hwe3-bit anokunda uye kukuvara kwe quantization kwahwo kusinganyanyi kuoneka.
• Basa rako nderekurongedza kupfupi kwehuwandu hukuru. Kukuvara kwe quantization hakuoneki ipapo — tenga chero chinokurumidza.

09Kudana API yeMiMo isina kudzorwa

Kana kodhi yako ichinangana neSDK yeOpenAI, kuchinja base URL nezita remodhi ndiko kutama kwose:

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
  }'

/v1/messages ne/v1/responses zvinogamuchira zita rimwe chetero neunyanzvi humwe chetehwo. Rondedzero yemapharamita, maschema ekudana zvishandiso, mienzaniso yezvinobuda zvakarongwa nenzvimbo yekuedza iri kushanda zviri mumagwaro eAPI eShannon AI.

Unyanzvi	MiMo-V2.5-Pro · W8A16
Hwindo recontext	matokeni 262,144
Zvinobuda zvakanyanya pachikumbiro	matokeni 16,384
Kuyerera	Hongu — mitauro mitatu yose
Kudana zvishandiso	Hongu
Kufungisisa / reasoning_effort	Hongu
Zvinobuda zvakarongwa zvejson_schema	Hongu
response_format	Hongu
Kuona / mifananidzo inopinda	Kwete — shandisa MiMo-V2.5
Chidzitiro chekuramba / sefa pane zvinobuda	Hapana
Mutengo pamatokeni 1M	$0.50 anopinda / $2.00 anobuda

10Mibvunzo inowanzobvunzwa

---

Zvakatorwa kubva: gwaro reunyanzvi reMiMo, arXiv:2505.07608 · kadhi remodhi reMiMo-V2.5-Pro (Xiaomi) · "Give Me BF16 or Give Me Death"?, arXiv:2411.02355 · REAP the Experts, arXiv:2510.13999 · AutoRound, arXiv:2309.05516 · mitemo yechiyero yekushandisa mabhiti, arXiv:2212.09720 · Scaling Laws for Precision, Kumar nevamwe, ICLR 2025. Nhamba dzose dzemiyero sezvadzakashumwa nevanyori vadzo. Shannon AI chigadzirwa cheShannon Lab LLC, New Mexico, USA.