API na Samfurin Inkling Marar Tantancewa

Yawancin samfura goma sha biyu a ƙofar shigarmu rubutu kaɗai ne. Huɗu ba haka ba ne, kuma a cikin waɗannan huɗun Inkling yana riƙe takamaiman matsayi: samfurin gani da za ka kai gare shi lokacin da hoton shi ne ɓangare mai wuya na matsalar amma kasafin kuɗi bai kai ga jagora ba. Wannan katin ya rufe inda tushen ya fito, abin da ƙarin W4A16-AUTOROUND-REAP ya yi masa, abin da zai yi da abin da ba zai yi ba a filin API, kuma — da gaskiya — lokacin da samfuri mai rahusa a katalogi guda shi ne zaɓi mafi kyau.

01Menene Inkling, kuma wa ya gina shi?

Inkling shi ne samfurin farko mai buɗaɗɗen nauyi daga Thinking Machines Lab, an wallafa shi a 15 ga Yuli, 2026. Ba kamar yawancin sakin da aka sanya a matsayin "buɗaɗɗe" ba, ya fito da cikakken nauyi da za a iya saukewa a Hugging Face da goyon baya tun rana ta sifili a transformers, SGLang, vLLM, da llama.cpp — wanda shi ne ya sa ya zama mai yiwuwar gudana a kan tsarin kamar namu kwata-kwata. Manyan lambobin, daga kayan sakin Thinking Machines da kansu:

Tsarin Mixture-of-Experts mai raguwa ne: shimfidun ɓoye-ɓoye 66 a girman ɓoye na 6,144, biyun farko masu yawa, sauran 64 suna ɗauke da masana 256 da ake turawa gare su tare da masana 2 da ake rabawa, da masana 6 da ake turawa gare su masu kunnuwa a kowane token. Kusan kashi 4% na sigogi suke kunna a kan kowane token — wanda shi ne ya sa samfuri mai kusan sigogi tiriliyan ya zama mai yiwuwar hidima kwata-kwata.

Zaɓuɓɓukan gine-gine biyu ba a saba da su ba har sun cancanci a ambace su. Kulawa mai haɗewa ce — shimfidu 55 na kulawa ta taga mai zamewa a kan taga na tokens 512 idan aka kwatanta da shimfidu 11 na gaba ɗaya, a tsari mai maimaituwa na 5:1. Kuma Inkling ba ya amfani da RoPE: yana amfani da son zuciyar matsayi mai dangantaka da aka koya, mai dogaro ga shigarwa, tare da gajerun juyi na sanadi a kan turawar maɓalli/daraja da rassan saura. Bayanan gine-gine na Sebastian Raschka sun yi alama a kan duka biyun a matsayin ɓacewa na gaskiya daga girke-girken 2026 na yau da kullum. Horo ya yi amfani da Muon ga manyan sigogin matrix da Adam ga sauran, sannan sama da miliyan 30 na gudun RL marasa jituwa a horo na baya.

Batun ga mai amfani da API ba ƙaramin bayanin ba ne. Shi ne cewa wannan tushe ne mai girman gaba, mai yawan hanyoyi ta asali da aka horar a kan tokens tiriliyan 45 da suka mamaye rubutu, hotuna, sauti da bidiyo — ba samfurin rubutu da aka manne masa mai daidaita gani daga baya ba. Ana ɓoye hotuna a matsayin faci na pixel 40×40 ta mai yanka faci na MLP mai matakai huɗu kuma ana sarrafa su tare da rubutu a cikin mai ɓoye-ɓoye guda. Fahimtar gani tana cikin manufar horo na farko.

02Menene ma'anar W4A16-AUTOROUND-REAP?

Dabaru uku daban an tara su a cikin wannan ƙarin, kuma suna yin aikace-aikace daban. Ya cancanci a fahimta, saboda ƙarin shi ne bambanci mafi girma tsakanin samfurin da kake kira a API ɗinmu da wurin ajiyar da Thinking Machines ya wallafa.

REAP — datsa masanan da babu wanda yake turawa gare su

REAP shi ne Router-weighted Expert Activation Pruning, daga Cerebras Research (arXiv 2510.13999, an karɓe shi zuwa ICLR 2026). A cikin MoE mai masana 256 a kowace shimfida router ba ya rarraba aiki daidai: ana buga wasu masana koyaushe, wasu suna kunna da wuya kuma suna bayar da kaɗan idan suka yi. REAP yana ba kowane masani maki a kan duka darajojin ƙofar router da ma'aunin kunnuwan masanin — masanin da ake zaɓarsa sau da yawa amma yake bayar da gyara masu ƙaramin girma ba shi da muhimmanci fiye da yadda adadin zaɓarsa yake nunawa — kuma yana gogewa waɗanda suke da ƙarancin muhimmanci a harbi guda, kafin matsi.

Dalilin da ya sa muke amfani da datsawa maimakon haɗa masana mai shahara shi ne cewa su biyun sun bambanta bisa nau'in aikin. Haɗawa tana kama da mai gasa a kan ma'aunan rarrabewa kuma tana rushewa a kan na samarwa. Marubutan REAP sun bayar da rahoton cewa datsawa tana kiyaye inganci da kyau daidai a kan ayyukan da mutane suke sayen API don su — haɗa lamba, kiran kayan aiki, tunani mai juyi da dama — a girman matsi sama da 50%, a kan samfura daga 20B zuwa sigogi 1T. Haɗawa tana lalata ƙwarewar router; datsawa tana barin masanan da suka tsira da turawa a tsakaninsu ba tare da lalacewa ba.

AutoRound — koyan inda za a zagaya

AutoRound ita ce hanyar matsi bayan horo ta Intel (arXiv 2309.05516, an wallafa ta a matsayin SignRound). Matsi mai sauƙi yana juya kowane nauyi zuwa darajar 4-bit da za a iya wakilta mafi kusa — abin da ya dace ta asali kuma mara kyau a girma, saboda kurakuran zagaye ba masu zaman kansu ba ne kuma suna taruwa ta tarin shimfidu 66. AutoRound yana ɗaukar hukuncin zagaye a matsayin abin da za a inganta: gangarawar gradient mai alama tana daidaita alkiblar zagaye da iyakokin yanke nauyi tare a kan ƴan matakai ɗari biyu, tana aro daidaiton horo mai lura da matsi ba tare da sake horo ba. Intel ya bayar da rahoton ribar da ta fi girma a saitunan ƙananan bits masu tsanani da gefen gasa a INT4.

W4A16 — inda bits suke tafiya

W4A16 shi ne tsarin lissafi: nauyi na 4-bit, kunnuwa na 16-bit. Nauyin — wanda yake mamaye ƙwaƙwalwa a samfuri mai 975B — yana raguwa zuwa kusan kashi ɗaya cikin huɗu na sawun BF16, alhali lissafin yana zama a 16 bits. Wannan rabon yana da muhimmanci ga samfurin gani musamman: tsare-tsaren da suka matsa kunnuwa suna cutar da daidai aikin lissafi mai daki-daki da fahimtar hoto ta dogara a kai — karanta ƙaramin lakabin gatari, warware layi mai laushi a cikin taswira, raba inuwa biyu masu kama da juna. Riƙe kunnuwa a 16 bits yana kashe faɗin hanya kuma yana sayo daidaito inda aikin gani yake lura.

Idan aka tara ukun, suna ba da samfurin da yake shiga kuma yake ci gaba da sauri ba tare da faɗuwar ingancin matakin 3BIT-REAP mai tsanani da muke amfani da shi a kan manyan tushe ba. Wannan shi ne cinikin da farashin yake nunawa.

03Gani: API na nazarin hoto mara ƙi

Inkling yana ɗaya daga cikin samfura huɗu masu iya gani a katalogin mu, tare da Kimi-K3, Kimi-K2.6, da MiniMax-M3. Sakamakon gani da aka wallafa na samfurin asali, daga sakin Thinking Machines, su ne 73.3% a MMMU Pro (Standard 10) da 78.1% a tambayoyin tunani na CharXiv — suna hawa zuwa 82.0% idan aka bar samfurin ya gudanar da Python a kan zanen. Waɗannan lambobin dakin binciken ne ga wurin ajiyar da ba a matsa ba; ka ɗauke su a matsayin rufin dangin, ba aunan ginin mu ba.

CharXiv shi ne wanda ya cancanci a tsaya a kai, saboda yana danganta da dalilin da ya sa mutane suke sayen API na gani maimakon hidimar OCR. Yana gwada tunani a kan zanen kimiyya — karanta daraja daga gatari, kwatanta jerin biyu, lura da juyawar yanayi. Ba "wane rubutu ne a cikin wannan hoton" ba amma "me wannan siffar take iƙirari, kuma shin iƙirarin yana riƙewa." Wannan shi ne abin da yake ba da hujjar aika hoto zuwa samfurin tunani mai 975B maimakon ƙaramin mai bayanin hoto.

Aika hoto aiki ne na yau da kullum mai jituwa da OpenAI — ɓangaren abun ciki na image_url a kan /v1/chat/completions, toshen image na salon Anthropic a kan /v1/messages, ko makamancinsa a kan /v1/responses. Yaruka ukun suna buga nauyi guda kuma ukun suna yin streaming; ka zaɓi wanda SDK ɗinka na yanzu ya riga ya sani. Misali da aka yi yana cikin sashe na 09.

Ɓangaren marar tantancewa ya fi muhimmanci ga gani fiye da ga rubutu. Shimfidar ƙi a kan ƙarshen hanyar gani ba ta da hankali sosai, saboda yawanci tana rarraba hoton kafin samfurin ya yi tunani a kan buƙatar. Aiki halastacce yana faɗawa cikin tarko koyaushe: hoton rauni a cikin hanyar aikin bayanan asibiti, hoton allo na shaida a cikin rubutun abin da ya faru, firam daga fim ɗin da ake nazari don rubutun suka, hoton dubawa na likita, hoton takarda da ya faru yana ɗauke da suna. Inkling a ƙofar shigarmu ba shi da irin wannan mai rarrabewa a gabansa. Yana duban abin da ka aika kuma yana amsa tambayar da ka yi. Wannan shi ne kayan.

04Fitarwa mai tsari: response_format eh, json_schema a'a

Wannan shi ne wuri guda inda Inkling ya fi ƙarancin iyawa daga yawancin katalogin, kuma ya fi kyau a tsara masa maimakon a gano shi a cikin samarwa.

Iyawa	inkling	Kimi-K2.6	MiMo-V2.5	Laguna-S-2.1
Gani (shigar hoto)	Eh	Eh	Eh	A'a
response_format	Eh	Eh	Eh	A'a
json_schema	A'a	Eh	Eh	A'a
Kiran kayan aiki	Eh	Eh	Eh	Eh
Ƙoƙarin tunani	Eh	Eh	Eh	Eh

response_format: {"type": "json_object"} yana aiki — kana samun JSON mai inganci a tsari. Abin da ba ka samu shi ne ɓoye-ɓoye da nahawu ta ƙuntata a kan schema da ka bayar, inda aka hana mai ɗauka a zahiri fitar da token da ya karya tsarinka. Sunayen filaye, kasancewa a cikin enum, da kasancewar maɓallan da ake buƙata zaɓin samfurin ne, ba tabbaci ba.

Hanyoyi uku na zama da hakan, a tsari mai hawa na aminci:

1. Ka sanya schema a umarni ka tabbatar. Ka sanya schema a cikin umarnin tsari a matsayin JSON kai tsaye, ka saita response_format zuwa json_object, kuma ka tabbatar da amsar a ɓangarenka. Ka sake gwadawa idan tabbatarwa ta gaza tare da ƙara saƙon kuskuren mai tabbatarwa. Mai rahusa, yana aiki, yana buƙatar kasafin sake gwadawa.
2. Ka yi amfani da ma'anar kayan aiki a matsayin schema. Ana goyon bayan kiran kayan aiki, kuma ma'anar sigar aiki JSON Schema ce. Ka bayyana kayan aiki guda da aka ba shi sunan cirar bayanan da kake so, ka tilasta shi, kuma ka karanta hujjojin. Wannan yakan zama zaɓi mafi ƙarfi a kan Inkling kuma ba ya kashe maka wani ƙari.
3. Ka raba aikin a kan samfura biyu. Ka bar Inkling ya yi gani da tunani a rubutu, sannan ka ba da wannan rubutun ga MiMo-V2.5-W8A16 a $0.50 / $2.00 tare da json_schema na gaskiya ga wucewar tsarawa. Kiraye-kiraye biyu, amma na biyun mai rahusa ne, mai tabbaci a siffa, kuma samfurin da yake buƙatar zama mai wayo shi ne wanda ya ga hoton.

Zaɓi na 3 shi ne tsarin da za mu ba da shawara ga duk wata hanyar cirar bayanai mai yawa. Ka biya ƙarin kuɗin kawai ga matakin da yake buƙatarsa.

05Menene ƙarin farashin yake saya?

A $1.42 shiga / $6.07 fita kowane tokens miliyan, Inkling yana zaune da gangan a tsakiyar katalogin mu. Shida daga cikin samfura goma sha biyu suna a ƙasan $0.50 / $2.00. Kimi-K3 — jagoran, wanda shi ma yana iya gani — yana $3.83 / $19.12. Inkling kusan ninki uku ne na ƙasan a fitarwa kuma kusan kashi ɗaya cikin uku na jagoran.

Farashin fitarwa, dala kowane tokens 1M, farashin jeri na Ƙofar Shiga ta Shannon AI, Satumba 2026. Faɗin sanda yana daidai da farashi; mafi tsawo ya fi tsada.

Ga karatun gaskiya a kan inda wannan ƙarin yake cin abincinsa — da inda ba ya ci.

Aiki	A yi amfani da Inkling?	Dalili
Tunani a kan zane / taswira	Eh	Lambobin CharXiv da MMMU Pro da aka wallafa su ne abin da wannan tushen ya fi ƙarfi a kai.
Wakilan da hoton allo ke tuƙa su	Eh	Gani tare da kiran kayan aiki tare da tarihi na 262K a samfuri guda, babu miƙawa.
Dogayen takardu masu siffofi	Eh	Rubutu da hotuna suna raba mai ɓoye-ɓoye guda da taga mahalli guda.
Samar da rubutu kawai	A'a	Samfura shida a $0.50 / $2.00 suna yin haka. Kana biyan kuɗin hanyar gani da ba ka taɓa kira ba.
Cirar bayanai ta schema mai yawa	A'a	Babu json_schema. MiMo-V2.5 ya fi rahusa kuma an ƙuntata shi.
Cikakken rufin gaba	A'a	Kimi-K3 shi ne jagoran. Ka biya ninki 3.1 idan rufin shi ne ƙuntatawar.
Gani mafi rahusa da zai yiwu	A'a	MiniMax-M3 yana da gani a $0.50 / $2.00. Ka fara daga can, ka haura idan inganci ya tilasta.

Ka karanta wannan a matsayin dokar turawa, ba matsayi ba. Inkling ya dace da fage mai kunkuntar aiki fiye da yadda shafin talla zai iƙirari, kuma a cikin wannan fagen yana da wuyar dokewa a farashi. A wajensa, mun fi so mu sayar maka da samfuri mai rahusa mu kuma riƙe ka.

06Taga mahalli na 262K, da abin da yake shiga da gaske

Ana hidimar duka samfura goma sha biyu a ƙofar shiga da mahalli na shigarwa na tokens 262,144 da matsakaicin fitarwa na tokens 16,384 a kowace buƙata. Iri ɗaya da gangan: sarrafa mahalli bai kamata ya taɓa zama reshe na kowane samfuri a lambarka ba, kuma musanya inkling da MiniMax-M3 a fayil ɗin saiti bai kamata ya canja abin da yake shiga a shiru ba.

Ka lura da giɓin idan aka kwatanta da tushen da aka wallafa: Thinking Machines yana rubuta Inkling a tokens 1,048,576, kuma muna hidimar yanki na 262K daga wannan. Idan kana da aikin buƙata guda na tokens miliyan na gaskiya, ba mu ne ƙarshen hanyar hakan ba, kuma ya fi kyau ka sani kafin ka gina.

Ga girman a 262K: PDF na fasaha mai shafuka 400 tare da siffofinsa, tarin lamba mai matsakaicin girma, awanni da dama na rubutun magana, ko dogon zaman wakilai da yake ɗauke da dozin na sakamakon kayan aiki da ƴan hotunan allo — kowanne yana shiga a buƙata guda. A kan samfurin gani, ka tuna hotuna suna cin mahalli ma; hoton allo mai yawa ba kyauta ba ne, kuma zaman da yake yin streaming na firam yana cika taga da sauri fiye da na rubutu kaɗai. Gine-ginen ya dace da wannan sosai: tsarin taga-mai-zamewa zuwa gaba ɗaya na 5:1 yana nufin yawancin shimfidu suna kulawa a cikin unguwar tokens 512, tare da shimfidun gaba ɗaya suna jan tsarin dogon zango baya lokaci-lokaci.

07Ƙoƙarin tunani: biyan kuɗin tunani kawai lokacin da kake buƙatarsa

Inkling yana buɗe ƙoƙarin tunani da za a iya sarrafawa — ɗaya daga cikin maɓallan da suka fi amfani a katalogin. Samfurin da ke sama yana karɓar matakai daga none da minimal ta low, medium, high, xhigh, har zuwa max, an fito da su a kan yaruka ukun ta sigar reasoning_effort ta yau da kullum.

{
  "model": "inkling-W4A16-AUTOROUND-REAP",
  "reasoning_effort": "low",
  "messages": [ ... ]
}

Wannan maɓallin kuɗi ne mai haƙora, saboda ana caje tokens na tunani a matsayin tokens na fitarwa a $6.07 kowane miliyan. Thinking Machines ya bayar da rahoton cewa a Terminal Bench 2.1 Inkling yana kai maki kwatankwacin Nemotron 3 Ultra a kusan kashi ɗaya cikin uku na tokens — ingancin tokens manufar ƙira ce da aka bayyana ta tsarin ƙoƙari, ba haɗari ba.

A aikace: ka gudanar da low ko minimal don bayanin hoto, rarrabewa, da cirar bayanai kai tsaye, inda amsar tana cikin hoton kuma babu buƙatar sarkar tunani. Ka ajiye high da sama don tunani a kan zane, kwatancin gani mai matakai da dama, da juyin wakilai inda matakin tsakiya mara kyau yake da tsada. Yawancin ƙungiyoyi ba sa saita sigar kwata-kwata, kuma suna barin bambancin kuɗin fitarwa na ninki da dama a kan tebur.

08Menene ma'anar "marar tantancewa" a nan — kuma menene ba ya nufi

A takamaime: muna gudanar da wannan samfurin da babu shimfidar ƙi kuma babu matata a kan fitarwa. Babu mai rarrabewa da yake karanta umarninka yana yanke shawarar ko samfurin zai iya amsawa, kuma babu wanda yake karanta amsar kafin ta yi streaming. Idan samfurin bai da tabbaci ko ba shi da ilimin, yana faɗa — wannan samfurin ne, ba manufa ba. Abin da ba za ka samu ba shi ne ƙi ta atomatik a kan batun da kayan masu amfani na gaba ɗaya suke ɗauka a matsayin haramtacce.

Ga gani wannan shi ne duka darajar da ake bayarwa. Masu rarraba hoto a gaban API na gani ba su da daidaito, kuma adadin kuskuren tabbatarwa yana sauka mafi tsanani a kan aikin ƙwararru da ya fi buƙatar iyawar: binciken tsaro, mahallin asibiti, aikin jarida da shaida, sukar fim da fasaha, kayan aikin tantancewa da dole ne su duba abin da suke tantancewa. Ƙarshen hanyar da take ƙi bisa rukuni ba za ta iya hidimar ko ɗaya daga cikinsu da aminci ba.

Gargaɗi biyu na gaskiya. Na farko, Thinking Machines yana wallafa aunan tsaro ga turarsu ta tunani — 78.0% a FORTRESS adversarial, 95.9% a FORTRESS benign, 98.6% a StrongREJECT. Waɗannan lambobin suna kwatanta tarin nasu tare da shingen nasu a wurin. Ba sa kwatanta namu, kuma ba za mu ambace su kamar suna yi ba. Mun cire shimfidar da waɗannan lambobin suke auna wani ɓangare; wannan shi ne kayan, kuma alhakin.

Na biyu, "babu shimfidar ƙi" ba "babu dokoki" ba ne. Shannon Lab LLC yana aiki ƙarƙashin Manufar Amfani da Alhaki da dokar Amurka da ta shafe shi. Cire mai rarrabewa yana motsa hukunci zuwa gare ka, ba ya goge shi. Idan kana gina a kan wannan, ka karanta manufar kafin ka faɗaɗa.

09Farawa

URL na asali guda, maɓalli guda, yaruka uku. Idan lambarka tana riga tana kiran OpenAI, Anthropic, ko API na Responses, canja kirtanin samfurin yakan zama duka ƙaurar.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.shannon-ai.com/v1",
    api_key=SHANNON_API_KEY,
)

stream = client.chat.completions.create(
    model="inkling-W4A16-AUTOROUND-REAP",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "What does this figure actually show?"},
        {"type": "image_url", "image_url": {"url": image_data_url}},
    ]}],
    reasoning_effort="high",
    max_tokens=4000,
    stream=True,
)

for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="")

Cikakken bayanin sigogi, siffofin /v1/messages da /v1/responses, tsarin abubuwan streaming, da wurin gwaji kai tsaye suna cikin takardun API. Katunan samfuri ga sauran samfura goma sha ɗaya suna cikin Shannon Research.

Lambobin tushe lambobin Thinking Machines Lab ne da aka wallafa ga wurin ajiyar Inkling da ba a matsa ba kuma suna kwatanta rufin dangin, ba ginin mu da aka matsa ba. Farashi da alamun iyawa darajojin jeri ne na Ƙofar Shiga ta Shannon AI tun Satumba 2026.

10Tambayoyin da ake yawan yi

---

Majiyoyi: Thinking Machines Lab — Introducing Inkling · Hugging Face — Welcome Inkling · REAP the Experts (arXiv 2510.13999) · Optimize Weight Rounding via Signed Gradient Descent (arXiv 2309.05516) · Sebastian Raschka — bayanan gine-gine da ma'auni na Inkling. Dakin bincike ne ya bayar da rahoton ma'aunan tushe ga wurin ajiyar da ba a matsa ba.