API na MiMo Marar Tantancewa a Aminci na 8-Bit

Yawancin katalogi suna sayar maka da adadin sigogi. Wannan ba lambar da ya kamata ka yi siyayya a kanta ba ce, saboda duk wani babban samfuri mai buɗaɗɗen nauyi da za ka iya samu da gaske an matsa shi don ya shiga wani wuri. Tambaya ta gaskiya ba ita ce "girmansa nawa kafin a matse shi?" ba amma "nawa daga cikinsa ne har yanzu yake nan?" Wannan labarin yana game da wannan tambayar, kuma game da samfurin a jerinmu inda amsar take kusan duka.

01Abin da API na MiMo marar tantancewa yake ba ka da gaske

API na MiMo marar tantancewa alamar samfuri guda ce, MiMo-V2.5-Pro-W8A16, ana kai gare ta ta yaruka uku da suke kai ga nauyi guda a kan gungun GPU namu: /v1/chat/completions, /v1/messages da /v1/responses. Ukun suna yin streaming kuma suna goyon bayan kiran kayan aiki da tunani.

Marar tantancewa yana da kunkuntar ma'ana a nan: babu shimfidar ƙi a gaban samfurin, babu matata a kan fitarwarsa. Ba ma tsayarwa, rarrabewa ko sake rubutawa. Kana samun ɗabi'ar tushen da kansa — wanda ga samfuri mai buɗaɗɗen nauyi na matakin bincike yana nufin yana shiga cikin binciken tsaro, almara ta manya ko ta tashin hankali, tunani na likita da na shari'a, da buƙatun da ƙarshen hanya na kasuwanci suke ƙi ta atomatik. Kayan aiki ne na ƙwararru ƙarƙashin Manufarmu ta Amfani da Alhaki, ba rashin dokoki ba.

02Zuriyar MiMo: fatan Xiaomi a kan horo na farko mai fifita tunani

MiMo ba abin da aka samo daga Llama da sabon suna ba ne; zuriya ce ta daban daga ƙungiyar MiMo ta Xiaomi, kuma ra'ayin da ya kafa ta yana da muhimmanci kafin ka auna samfurin. Ya buɗe da MiMo-7B (30 ga Afrilu, 2025, mai lasisin MIT), wanda rahotonsa na fasaha yake jayayya cewa yawancin dakunan bincike suna ɗaukar tunani a matsayin matsalar horo na baya da za a manne da RL, kuma cewa wannan a baya ne. Xiaomi ya sake tsara horo na farko a kansa maimakon haka: haɗi mai matakai uku a kan tokens tiriliyan 25, mataki na biyu yana gudanar da kusan kashi 70% lissafi da lamba, mataki na uku yana naɗe alamun tunani na roba yayin da yake faɗaɗa mahalli daga 8K zuwa 32K — tare da manufar Hasashen Tokens da Yawa da take kuma aiki a matsayin kan ɓoye-ɓoye na hasashe.

Sakamakon ya kasance samfurin asali mai 7B da yake ɗabi'a kamar wani abu mafi girma sosai. MiMo-7B-Base ya samu 32.9 a AIME 2024 inda Qwen2.5-7B ya samu 10.1, Gemma-2-9B 0.0 da Llama-3.1-8B 0.3. Sannan horo na baya ya gudanar da RL a kan matsalolin lissafi da lamba 130,000 da za a iya tabbatarwa tare da ladan lamba da wahalar gwaji ke tuƙa, yana tserewa tarkon ladan da ba a cika ba wanda yake rushe RL na lamba. Sakamakon ya sauka a gaban o1-mini:

AIME 2025, lambobin da Xiaomi ya bayar da rahotonsu daga rahoton fasaha na MiMo (arXiv:2505.07608). MiMo-7B-RL ya kuma bayar da rahoton 95.8 a MATH-500 da 57.8 a LiveCodeBench v5.

Dangin ya faɗaɗa maimakon ya kauce: MiMo-VL-7B (Yuni 2025) ya sanya mai ɓoyewa na Qwen2.5-ViT a kan kashin baya guda, kuma Disamba 2025 ya kawo MiMo-Audio-7B tare da shigarwar MoE ta farko, MiMo-V2-Flash (jimla 309B / 15B masu kunnuwa, tokens 27T, daidaito gauraye na FP8). MiMo-V2-Pro ya biyo baya a Maris 2026 a kusan 1T / 42B masu kunnuwa da mahalli na 1M, bayan makonni a OpenRouter a matsayin "Hunter Alpha". MiMo-V2.5 da MiMo-V2.5-Pro sun fito tare a 22 ga Afrilu, 2026, duka biyun MIT.

03Cikin MiMo-V2.5-Pro: yadda tushen yake

Bisa katin samfuri da Xiaomi ya wallafa, MiMo-V2.5-Pro samfurin Mixture-of-Experts mai raguwa ne mai jimlar sigogi tiriliyan 1.02 da biliyan 42 masu kunnuwa a kowane token a kan shimfidu 70 — guda mai yawa, 69 na MoE — da masana 384 da ake turawa gare su, 8 a kowane token. Girman ɓoye 6144 ne, da ƙaramin girman tsakiya na MoE na 2048.

Zaɓuɓɓuka biyu suna siffanta yadda yake ji a amfani. Tarin kulawa haɗe ne: shimfidu 10 na cikakkiyar kulawa da shimfidu 60 na taga mai zamewa da taga na 128, tare da son zuciyar attention-sink da za a iya koya. Xiaomi ya bayar da rahoton cewa wannan yana yanke ajiyar ma'ajiyar KV kusan sau 7 idan aka kwatanta da kulawa ta gaba ɗaya alhali yana riƙe ingancin dogon mahalli — bambanci tsakanin taga da take da tsawo a suna da wadda za ka iya cika ta. Yana kuma fito da shimfidu uku na Hasashen Tokens da Yawa, yana ba mai ɓoye-ɓoye kan hasashe.

Horo ya kasance tokens tiriliyan 27 a tsawon jeri na asali na 32K a daidaito gauraye na FP8 (E4M3), sannan tace mai kulawa, RL na wakilai mai girma da distillation na on-policy mai malamai da dama. Ka riƙe wannan cikakken bayanin FP8; shi ne hujja guda mafi ƙarfi ta hidimar wannan tushen a 8 bits.

Xiaomi ya bayar da rahoton samfurin a 57.2% a SWE-bench Pro, 72.9 a τ3-Bench ga amfani da kayan aiki da 48.0% a Humanity's Last Exam — lambobin mai kera da kansa a kan samfurin da ba a matsa ba, an ambace su haka.

Muna hidimar taga na tokens 262,144, ba cikakken 1M na tushen ba — mahallin da za mu iya tabbatarwa a sauri, maimakon babbar lamba da take raguwa a ƙarƙashin nauyi.

04Menene ma'anar W8A16, daidai

Kowane samfuri a wannan katalogin yana ɗauke da ƙarin matsi da yake kwatanta ainihin canji da za a iya aunawa maimakon matakin talla. Namu su ne 3BIT-REAP, W4A16-AUTOROUND-REAP da W8A16. MiMo-V2.5-Pro shi ne na uku.

Rubutun yana rabuwa gida biyu. W8 shi ne daidaiton nauyi: ana adana kowace siga a matsayin lamba mai 8-bit maimakon float mai 16-bit, tare da abin ninkawa da ake rabawa a kowace ƙaramar ƙungiya don fagen motsi na kowace ƙungiya ya tsira. A16 shi ne daidaiton kunnuwa: darajojin da suke gudana tsakanin shimfidu a lokacin tunani suna zama a 16 bits. Ana matsa nauyi; ba a matsa lissafin ba.

Rabin na biyu yana yin aiki a shiru amma mai muhimmanci. Kunnuwa sun fi wuyar matsawa fiye da nauyi: suna ɗauke da tashoshi masu tsananin bambanci — ƴan girma waɗanda girmansu ya zarce duk sauran kuma waɗanda suke ɗauke da kaso mai yawa na ɗabi'ar samfurin. Ka tura waɗannan zuwa ƙaramin daidaito kuma za ka yanke daidai alamar da kake buƙata. Tsare-tsaren nauyi-kaɗai suna guje wa matsalar ta barin hanyar kunnuwa, wanda shi ne dalilin da ya sa aikin matsi na gaskiya ya haɗu a kansu.

Don haka tambayar aiki ta matse zuwa abu guda: bits nawa ne mafi ƙaranci da nauyi zai iya jurewa?

05W8A16 da W4A16 da 3-bit: inda asarar take tafiya da gaske

Binciken da aka wallafa mafi girma a kan wannan shi ne na Neural Magic da Red Hat, "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization — sama da auna 500,000 a kan dangin Llama-3.1 a 8B, 70B da 405B. Babban sakamakonsu, a lambobinsu:

Tsari	Abin da aka matsa	Dawowar da aka bayar da rahoto idan aka kwatanta da BF16
W8A8-FP (8-bit)	Nauyi + kunnuwa, float	kusan 99.75% matsakaici — "kusan babu asara"
W8A8-INT (8-bit)	Nauyi + kunnuwa, lamba	99–100% a 8B/70B; 98.2–99.3% a 405B
W4A16-INT (4-bit)	Nauyi kaɗai	98–99.4% (8B); 98.6–99.8% (70B)

Lambobin da Neural Magic / Red Hat suka bayar da rahotonsu (arXiv:2411.02355) a kan samfuran Llama-3.1 — siffar cinikin, ba makin wannan samfurin ba.

Idan aka karanta da gaskiya, wannan yana kama da hujja a kan damuwa: 4-bit yana dawo da 98–99%, don me za a kula? Saboda waɗannan ma'aunan zaɓi da yawa na ilimi ne, kusan mafi munin kayan aiki da ake da shi don gano barnar matsi. Tambayar ma'auni tana da amsa daidai guda da za a iya kaiwa a token ɗaya ko biyu, kuma kuskuren matsi ƙarami ne, kusan hayaniya mai matsakaicin sifili a kan kowane logit; a kan gajeriyar tambaya mai faɗin gefe, ana shanye wannan hayaniyar.

Yanzu ka rubuta tokens 4,000 na takardar fasaha, ka riƙe murya a kan labari na tokens 40,000, ko ka fitar da JSON da dole ne ya inganta a karo na farko. Ana ɗaukar hayaniya guda dubbai sau a jere, kowane ɗauka yana daidaita na gaba. Hanyar gazawar ba amsa mara kyau ba ce — kaucewa ce: riƙon umarnin daga juyi na farko yana sassauta, hali da ya daina jin kamar kansa a kusa da token 3,000, ƙarshen ɗaure da bai taɓa zuwa ba.

Inda bits suke tafiya da gaske

Ba a rarraba kuskuren matsi daidai a kan iyawar samfurin ba; yana taruwa a wutsiya. Ɗabi'un masu yawan nauyi — Ingilishi mai nahawu, siffar aikin Python, manyan gaskiya dubu game da duniya — an ɓoye su a maimaice a kan nauyi da dama kuma suna tsira kusan komai. Kusurwoyin rabon nauyi da aka daidaita daidai, masu ƙaramin girma suna riƙe abubuwa da ba a saba ba: ainihin sa hannun API mai duhu, sautin wani takamaiman marubuci, bambanci tsakanin CVE biyu da suke maƙwabta. Waɗannan su ne darajojin da grid mai kauri yake juya su cikin juna.

Matakin daga bits 4 zuwa 3 shi ne inda wannan ya daina zama haraji ya zama tsaunuka. Dokokin girman tunani na k-bit na Dettmers da Zettlemoyer sun gano cewa 4-bit shi ne matsayin da yake ƙara daidaiton harbi-sifili a kowane bit na ƙwaƙwalwar samfuri, tare da raguwa ta mamaye a ƙasansa. Bits huɗu suna ba da matakai 16 a kowace ƙungiya; uku suna ba da 8. Kana rabin ƙudurin grid da tuni yake da kauri, kuma kuskuren zagaye a kowane nauyi kusan yana ninkuwa.

Lissafin da yake yanke hukunci

Kwatancin da ya cancanci a yi lissafi ne, ba ma'auni ba. Abin da yake tsira daga matsi kusan sigogi × bits kowace siga ne — girman fayil ɗin nauyi mai ɗauke da bayanai. Ka ɗauki tushe biyu na hasashe masu inganci daidai a kowace siga:

• Tushe A: sigogi 400B a 3 bits → kusan GB 150 na nauyin da aka riƙe.
• Tushe B: sigogi 160B a 8 bits → kusan GB 160 na nauyin da aka riƙe.

Bisa adadin sigogi tsantsa A ya yi nasara biyu da rabi zuwa ɗaya — lambar da katin samfurin ya jagoranta da ita. Bisa bayanan da aka riƙe suna daidai, kuma B yana da fa'idar tsari bugu da ƙari: nauyinsa yana zaune a kan grid mai laushi har ɗabi'ar wutsiyarsa ba ta lalace ba, alhali an matsa wutsiyar A cikin ɓangaren yawa. A ya san abubuwa da yawa; B ya san abin da ya sani da kaifi. Ga samarwa mai tsawo, tsari daidai, sarrafa salo da bin dogon mahalli da aminci, ka ɗauki B.

Lissafi na misali, ba ma'auni ba: yana nuna dalilin da ya sa "tushe mafi girma" da "samfurin da aka fi hidimarsa da kyau" iƙirari ne daban.

06Dalilin da ya sa tokens tiriliyan 27 na horo suke sa matsi mai tsanani ya fi muni

Sakamako na biyu, mara sauƙin fahimta, yana shafar wannan tushen kai tsaye. Takardar Kumar da abokansa ta ICLR 2025 Scaling Laws for Precision ta gano cewa yawan bayanan da aka horar da samfuri a kansu, shi ne yawan barnar da matsi bayan horo yake yi masa. Ƙaramin daidaito yana rage adadin sigogi masu tasiri na samfuri, kuma wannan raguwar tana girma da bayanan horo na farko da aka sha — har ya kai ga ƙarin bayanan horo na farko na iya zama cutarwa idan kana nufin matsi mai tsanani, saboda kana kwanta da tsari mai laushi da grid zai goge.

Samfurin da aka horar a kan tokens 5T yana da tsarin nauyi mai kauri, mai maimaituwa da sarari. Samfurin da aka horar a kan tokens 27T — kasafin MiMo-V2.5-Pro — ya kashe wannan ƙarin bayanan yana rubuta bambance-bambance masu laushi cikin darajojin nauyi masu daidaito, wanda shi ne ainihin abin da grid na 3-bit ba zai iya wakilta ba.

Sannan akwai gaskiyar FP8 daga sashe na 03, hujja mafi tsafta a wannan labarin. An horar da MiMo-V2.5-Pro a farko a daidaito gauraye na FP8. Mai inganta da yake aiki a kusan bits takwas na daidaito ne ya siffanta nauyinsa tun daga farko. Don haka hidimarsa a nauyi na 8-bit ba tunani na baya mai asara da aka ɗora a kan samfuri na 16-bit ba ne — yana kusa da mayar da nauyin zuwa yanayin lissafi da aka koye su a ciki. Ƙaramin tsari mai laushi ne yake wanzuwa a ƙasan bit na takwas da za a rasa. Ba za a iya faɗin haka ba ga waɗannan nauyin da aka tura zuwa uku.

07REAP: dalilin da ya sa datsa masana asara ce ta wani nau'i daban da zagaye

Biyu daga cikin matakan ukun suna ɗauke da ƙarin REAP — aiki ne na rukuni daban gaba ɗaya da matsi. REAP (Router-weighted Expert Activation Pruning, Cerebras Research, arXiv:2510.13999, ICLR 2026) matsi ne na harbi guda ga samfuran MoE masu raguwa: yana ba kowane masani maki bisa darajojin ƙofar router da ma'aunin kunnuwan masanin, sannan ya goge waɗanda suka fi ƙaranci. Cerebras ya bayar da rahoton sakamako kusan babu asara a kan samar da lamba da kiran kayan aiki ko da a datsawar kashi 50% na masana.

Muna amfani da REAP saboda dalilin da yake nan don shi: a kan manyan tushe shi ne ya sa su zama masu yiwuwar hidima. Amma ka lura da bambancin nau'i. Matsi yana rikita kowane nauyi kaɗan; datsawa tana cire masana gaba ɗaya. Idan aikinka yana turawa sosai zuwa iyawar da take zaune a cikin masani da ba a fi zaɓarsa ba — ƙaramin harshen shirye-shirye, ƙamus na musamman — ba a rage shi ba. Ya tafi, kuma babu umarnin da zai dawo da shi. Matsakaita suna zama masu kyau; lamarinka na wutsiya na iya kada ya zama.

MiMo-V2.5-Pro · W8A16 ba ya ɗauke da ƙarin REAP. Duka masana 384 suna nan, kuma router yana zaɓa daga cikakken tarin da aka horar da tushen da shi.

Mataki	Bits na nauyi	Masana	Abin da ka sayar
W8A16	8	Duka suna nan	Babu wani abu na tsari. Matakin aminci.
W4A16-AUTOROUND-REAP	4	An datse	Cikakken bayanin nauyi tare da masana masu ƙarancin zirga-zirga, don musanya da tushe mafi girma sosai a fagen farashi guda.
3BIT-REAP	3	An datse	Ƙudurin nauyi mai yawa tare da masana masu ƙarancin zirga-zirga — kuɗin gudanar da manyan tushe a sauri.

Matakin tsakiya ya cancanci a saurare shi da adalci. AutoRound ba zagaye mai sauƙi ba ne: hanyar Intel (arXiv:2309.05516) tana amfani da gangarawar gradient mai alama a kan kusan matakai 200 don daidaita hukunce-hukuncen zagaye da fagen yanke a kowace ƙungiyar nauyi, kuma an bayar da rahotonta kusan babu asara a 4 bits a yawancin saitunan, tana dokewa GPTQ, AWQ, HQQ da OmniQuant a kan yawancin tsare-tsaren da aka gwada. Samfurin W4A16-AUTOROUND samfurin 4-bit ne mai kyau da gaske. Har yanzu samfurin 4-bit ne.

08Wa ya kamata ya zaɓi matakin 8-bit — kuma wa bai kamata ba

MiMo-V2.5-Pro · W8A16 yana zaune a farashi mafi ƙasa na duka katalogin — $0.50 shiga / $2.00 fita kowane tokens miliyan, fage guda da samfuran 4-bit na matakin shiga kuma kusan kashi ɗaya cikin huɗu na farashin shigarwa na samfurin da ya fi tsada da muke hidimarsa. Ba kana biyan ƙari don aminci ba; kana zaɓar wani gatari daban na kasafi guda.

Ka zaɓe shi lokacin da

• Tsawon fitarwa shi ne aikin. Rubutun fasaha mai tsawo, almara mai tsawon babi, cikakkun wucewar takardu — kowane token da aka samar shi ne inda hayaniyar matsi take taruwa.
• Murya da daidaituwa suna da muhimmanci. Aikin hali, sautin alama, salon da aka riƙe a kan dogon zama. Salo yana zaune a wutsiya, kuma wutsiya ita ce abin da matsi mai kauri yake ci da farko.
• Fitarwa mai tsari dole ta inganta a karo na farko. Samfurin yana goyon bayan json_schema da response_format; bin schema siffa ce ta daidaito, kuma kowace sake gwadawa da aka guje wa kuɗi ne da jinkiri da aka adana.
• Kana cika mahallin. Bin umarni a token 200,000 ya fi wuya fiye da a token 2,000, kuma yana raguwa da daidaiton nauyi — haka kuma dogon alamar tunani, inda kurakurai suke taruwa haka.
• Aikin marar tantancewa ne ta ɗabi'arsa. Binciken tsaro, red-teaming, almara mara matata, nazarin likita ko na shari'a mai laushi — inda amsar da aka kauce mata ta fi ta rashin komai muni.

Ka zaɓi wani abu daban lokacin da

• Kana buƙatar gani. Wannan samfurin rubutu kaɗai ne; ɗan'uwansa MiMo-V2.5 · W8A16 yana iya gani a farashi da mataki guda.
• Kana buƙatar ilimin duniya mafi faɗi a kan gajerun tambayoyi. Wannan shi ne yanayin da ƙarin sigogin tushe mai 3-bit mafi girma sosai suke yin nasara kuma barnar matsinsa ba ta bayyana ba.
• Aikinka rarrabewa ce ta gajarce mai yawa. Barnar matsi kusan ba a ganinta a nan — ka sayi duk abin da ya fi sauri.

09Kiran API na MiMo marar tantancewa

Idan lambarka tana nufin SDK na OpenAI, canja URL na asali da alamar samfuri shi ne duka ƙaurar:

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 da /v1/responses suna karɓar alama da iyawa guda. Bayanin sigogi, schema na kiran kayan aiki, misalan fitarwa mai tsari da wurin gwaji kai tsaye suna cikin takardun API na Shannon AI.

Iyawa	MiMo-V2.5-Pro · W8A16
Taga mahalli	tokens 262,144
Matsakaicin fitarwa a kowace buƙata	tokens 16,384
Streaming	Eh — yaruka ukun
Kiran kayan aiki	Eh
Tunani / reasoning_effort	Eh
Fitarwa mai tsari ta json_schema	Eh
response_format	Eh
Gani / shigar hoto	A'a — ka yi amfani da MiMo-V2.5
Shimfidar ƙi / matatar fitarwa	Babu
Farashi kowane tokens 1M	$0.50 shiga / $2.00 fita

10Tambayoyin da ake yawan yi

---

Majiyoyi: Rahoton fasaha na MiMo, arXiv:2505.07608 · Katin samfuri na 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 · Dokokin girman tunani na k-bit, arXiv:2212.09720 · Scaling Laws for Precision, Kumar da abokansa, ICLR 2025. Duk lambobin ma'auni kamar yadda marubutansu suka bayar da rahotonsu. Shannon AI kayan Shannon Lab LLC ne, New Mexico, Amurka.