API yeModhi yeInkling Isina Kudzorwa

Mamodhi mazhinji pamamodhi gumi nemaviri ari paGedhi redu ndeemavara chete. Mana haasi akadaro, uye pamana iwayo Inkling ine nzvimbo yakatarwa: ndiyo modhi inoona yaunosvika kwairi kana mufananidzo uri chikamu chakaoma chedambudziko asi bhajeti risingasviki pamodhi yepamusoro-soro. Gwaro rino rinofukidza kwakabva hwaro, chakaitwa nechipfupiso cheW4A16-AUTOROUND-REAP kwahuri, zvainoita nezvaisingaiti pasuwo reAPI, uye — nechokwadi — rini modhi yakachipa iri mukatarogu imwe cheteyo iri sarudzo iri nani.

01Inkling chii, uye yakavakwa naani?

Inkling ndiyo modhi yekutanga ine huremu hwakavhurika inobva kuThinking Machines Lab, yakabudiswa musi wa15 Chikunguru 2026. Kusiyana nekubudiswa kwakawanda kunonzi "kwakavhurika", yakauya nehuremu hwakazara hunodhaunirodhwa paHugging Face nekutsigirwa kwezuva rekutanga mutransformers, SGLang, vLLM, nellama.cpp — ndiko kwakaita kuti igone kumhanya panhumbi seyedu zvachose. Nhamba huru, kubva pazvinyorwa zvekuburitswa zveThinking Machines pachadzo:

Ichi chimiro cheMixture-of-Experts yakapararira: madanho edhikodha 66 pahukuru hwakavanzika hwe6,144, maviri ekutanga akazara, makumi matanhatu nemana asara achitakura nyanzvi 256 dzinotumirwa pamwe nenyanzvi 2 dzinogovaniswa, nenyanzvi 6 dzinotumirwa dzichishanda patokeni imwe neimwe. Anenge 4% emapharamita anoshanda patokeni ipi neipi — ndicho chinoita kuti modhi ine mapharamita ari pedyo netiriyoni ikwanise kushandiswa zvachose.

Sarudzo mbiri dzemagadzirirwo hadzijairiki zvekuti dzinofanira kudanwa nezita. Attention yakarukanidzwa — madanho 55 eattention yehwindo rinotsvedza pahwindo rematokeni 512 kupesana nemadanho 11 akazara, muchimiro chinodzokorora che5:1. Uye Inkling haishandisi RoPE: inoshandisa bhaizi renzvimbo rakadzidzwa, rinotsamira pazvinopinda, pamwe nemakonvorushoni mapfupi ecausal pakuisirwa kwekiyi/kukosha nepamativi asara. Zvinyorwa zvemagadzirirwo zvaSebastian Raschka zvinoratidza ese ari maviri sekutsauka kwechokwadi kubva parisipi yakajairika ya2026. Kudzidziswa kwakashandisa Muon pamapharamita emamatrix makuru neAdam pane zvimwe zvose, ipapo kupfuura mamiriyoni makumi matatu ekumhanya kweRL kusina kurongwa mukudzidziswa kwemashure.

Chinokosha kumushandisi weAPI hausi ruzivo rwakadaro. Ndechekuti uhu ihwaro hwechiyero chepamberi, hunogamuchira midziyo yakawanda nenzira yepachahwo hwakadzidziswa pamatokeni matiriyoni makumi mana nemashanu anofukidza mavara, mifananidzo, ruzha nemavhidhiyo — kwete modhi yemavara ine chinamiro chekuona chakazonamirwa gare gare. Mifananidzo inoiswa sezvidimbu zvemapikiseru 40×40 kuburikidza nemupatsanuri wezvidimbu weMLP wemadanho mana wakarongwa uye inogadziriswa pamwe chete nemavara mudhikodha imwe cheteyo. Kunzwisisa zvinoonekwa kwaiva muchinangwa chekudzidziswa kwepakutanga.

02Chii chinorehwa neW4A16-AUTOROUND-REAP?

Nzira nhatu dzakasiyana dzakaunganidzwa muchipfupiso ichocho, uye dzinoita mabasa akasiyana. Zvakakodzera kunzwisisa, nokuti chipfupiso ndicho chinhu chikuru chinosiyanisa modhi yaunodana paAPI yedu nechekipoinzi chakabudiswa neThinking Machines.

REAP — kubvisa nyanzvi dzisingatumirwi nemunhu

REAP iRouter-weighted Expert Activation Pruning, kubva kuCerebras Research (arXiv 2510.13999, rakagamuchirwa kuICLR 2026). MuMoE ine nyanzvi 256 padanho rimwe nerimwe router haigoveri basa zvakaenzana: dzimwe nyanzvi dzinorohwa nguva dzose, dzimwe dzinoshanda zvishoma uye hadzibatsiri zvakawanda kunyange padzinoshanda. REAP inopa nyanzvi imwe neimwe chiyero pazvose zvikoshwa zvegedhi rerouter nezviyero zvekushanda kwenyanzvi — nyanzvi inosarudzwa kakawanda asi ichipa shanduko diki inokosha zvishoma pane zvinoratidzwa nekuverengwa kwekusarudzwa kwayo — ipapo yobvisa dzisingakoshi kamwe chete, quantization isati yaitwa.

Chikonzero chatinoshandisa kubvisa pane kusanganisa nyanzvi kunonzi kwakanaka mazuva ano ndechekuti dziri mbiri dzinosiyana nerudzi rwebasa. Kusanganisa kunoita sekunge kwakakwikwidza pamiyero yekusiyanisa uye kunoputsika pane yekugadzira. Vanyori veREAP vanoshuma kuti kubvisa kunochengetedza mhando zviri nani pamabasa chaiwo anotengerwa API nevanhu — kugadzira kodhi, kudana zvishandiso, kufungisisa kwekutenderera kwakawanda — pazviyero zvekudzvanya zvinopfuura 50%, pamamodhi kubva pa20B kusvika 1T yemapharamita. Kusanganisa kunoparadza unyanzvi hwerouter; kubvisa kunosiya nyanzvi dzasara nekutumira pakati padzo zvakachengetedzeka.

AutoRound — kudzidza kwekutenderedza

AutoRound inzira yeIntel ye quantization yemushure mekudzidzisa (arXiv 2309.05516, yakabudiswa seSignRound). Quantization yakapusa inotenderedza huremu humwe nehumwe kukukosha kwe4-bit kuri pedyo — chinhu chine musoro sechinotangira uye chakaipa pachiyero chikuru, nokuti zvikanganiso zvekutenderedza hazvina kuzvimirira uye zvinoungana kuburikidza nechimiro chemadanho 66. AutoRound inotora sarudzo yekutenderedza sechinhu chekugadzirisa: signed gradient descent inogadzirisa pamwe chete divi rekutenderedza nemiganhu yekucheka huremu pamatanho anenge mazana maviri, ichitora kururama kwekudzidziswa kunoziva quantization pasina kudzidziswazve. Intel inoshuma kunatsiridza kukuru pazvirongwa zvemabhiti mashoma zvakanyanya nemucheto wekukwikwidza paINT4.

W4A16 — kunoenda mabhiti

W4A16 ndicho chimiro chenhamba: huremu hwe4-bit, maactivation e16-bit. Huremu — hunotonga memori mumodhi ye975B — hunodzikira kusvika kuchikamu chinenge chechina chenzvimbo yeBF16, apo masvomhu achisara pa16 bits. Kukamurwa ikoko kunokosha pamodhi inoona chaizvo: zvirongwa zvine maactivation akaitwa quantization zvinokuvadza chaizvo basa renhamba rakatsetseka rinotsamirwa nekunzwisisa mifananidzo — kuverenga rebheru diki reaxis, kuona mutsetse mutete mudhayagiramu, kupatsanura mavara maviri akafanana. Kuchengeta maactivation pa16 bits kunodhura bhandiwidhi uye kunodzosera kururama apo basa rekuona rinozviona.

Kana zvaunganidzwa, zvitatu izvi zvinopa modhi inokwana uye inoramba ichimhanya pasina kuwira kwemhando kwemhando yakaomarara ye3BIT-REAP yatinoshandisa pahwaro hukuru chaizvo. Ndiko kuchinjana kunoratidzwa nemutengo.

03Kuona: API yekuongorora mifananidzo isina kuramba

Inkling ndeimwe yemamodhi mana anoona ari mukatarogu yedu, pamwe neKimi-K3, Kimi-K2.6, neMiniMax-M3. Mhedzisiro yekuona yakabudiswa yemodhi yepasi, kubva pakuburitswa kweThinking Machines, i73.3% paMMMU Pro (Standard 10) ne78.1% pamibvunzo yekufungisisa yeCharXiv — ichikwira kusvika 82.0% kana modhi ichibvumirwa kumhanyisa Python pachati. Idzo inhamba dzerabhoritari pachekipoinzi chisina quantization; dzitore sedenga remhuri, kwete kuyera kwechivakwa chedu.

CharXiv ndiyo yakakodzera kumira pairi, nokuti inoenderana nechikonzero chinoita kuti vanhu vatenge API inoona pane sevhisi yeOCR. Inoedza kufungisisa pamusoro pemachati esayenzi — kuverenga kukosha paaxis, kuenzanisa mitsara miviri, kuona chimiro chichidzoka shure. Kwete "mavara api ari mumufananidzo uyu" asi "chinotaurwa nemufananidzo uyu ndechei, uye chinomira here." Ndizvo zvinoita kuti kutumira mufananidzo kumodhi inofungisisa ye975B kukodzere panzvimbo pemunyori mudiki wetsananguro.

Kutumira mufananidzo ibasa rakajairika rinoenderana neOpenAI — chikamu chezvirimo cheimage_url pa/v1/chat/completions, bhuroko reimage rerudzi rweAnthropic pa/v1/messages, kana chakaenzana pa/v1/responses. Mitauro mitatu yose inosvika pahuremu humwe chete uye mitatu yose inoyerera; sarudza iyo SDK yako yazvino inotoziva. Muenzaniso wakaitwa uri muchikamu 09.

Chikamu chekusadzorwa chinonyanya kukosha pakuona kupfuura pamavara. Chidzitiro chekuramba pasuwo rinoona chakaomarara zvisina kujairika, nokuti chinowanzorongedza mufananidzo modhi isati yafungisisa pamusoro pechikumbiro. Basa riri pamutemo rinobatwa nguva dzose: mufananidzo weronda muhurongwa hwezvinyorwa zvekiriniki, sikirinishoti yeforensic muchinyorwa chechiitiko, furemu kubva mubhaisikopo riri kuongororwa pachinyorwa chekuongorora, sikeni yezvokurapa, mufananidzo wegwaro rine zita. Inkling paGedhi redu haina muongorori akadaro mberi kwayo. Inotarisa zvawatumira yopindura mubvunzo wawabvunza. Ndicho chigadzirwa.

04Zvinobuda zvakarongwa: response_format hongu, json_schema kwete

Iyi ndiyo nzvimbo imwe chete apo Inkling isina unyanzvi hwakaenzana nechikamu chikuru chekatarogu, uye zviri nani kuzvironga pane kuzviwana uri kushandisa.

Unyanzvi	inkling	Kimi-K2.6	MiMo-V2.5	Laguna-S-2.1
Kuona (mifananidzo inopinda)	Hongu	Hongu	Hongu	Kwete
response_format	Hongu	Hongu	Hongu	Kwete
json_schema	Kwete	Hongu	Hongu	Kwete
Kudana zvishandiso	Hongu	Hongu	Hongu	Hongu
Simba rekufungisisa	Hongu	Hongu	Hongu	Hongu

response_format: {"type": "json_object"} inoshanda — unowana JSON ine chimiro chakarurama. Chausingawani kugadzirwa kwakaganhurirwa negiramba pane schema yaunopa, apo musarudzi anodziviswa chaizvo kuburitsa tokeni inotyora chimiro chako. Mazita eminda, kuvapo kwezvinhu muenumu, nekuvapo kwemakiyi anodiwa isarudzo yemodhi, kwete chivimbiso.

Nzira nhatu dzekurarama nazvo, dzakarongwa kubva pane isingavimbiki kusvika pane inovimbika:

1. Isa schema muprompt uongorore. Isa schema muprompt yesisitimu seJSON chaiyo, isa response_format pajson_object, uongorore mhinduro kudivi rako. Edzazve kana kuongorora kwakundikana uchiwedzera meseji yechikanganiso yemuongorori. Yakachipa, inoshanda, inoda bhajeti rekuedzazve.
2. Shandisa chinyorwa chechishandiso seschema yako. Kudana zvishandiso kunotsigirwa, uye chinyorwa chemapharamita chebasa iJSON Schema. Zivisa chishandiso chimwe chete chakatumidzwa zita rekubvisa ruzivo kwaunoda, uchimanikidze, uverenge zvinopinda. Iyi ndiyo inowanzova sarudzo yakasimba kupfuura dzose paInkling uye haikudhurire chimwe.
3. Kamura basa pamamodhi maviri. Rega Inkling iite kuona nekufungisisa mumavara, wozopa mavara iwayo kuMiMo-V2.5-W8A16 pa$0.50 / $2.00 aine json_schema chaiyo pakupfuura kwekurongedza. Kudana kuviri, asi kwechipiri kwakachipa, kwakatarwa pachimiro, uye modhi inofanira kuva yakangwara ndiyo yakaona mufananidzo.

Sarudzo 3 ndicho chimiro chatingakurudzira pahurongwa hupi zvahwo hwehuwandu hukuru hwekubvisa ruzivo. Bhadhara mutengo wepamusoro padanho rinowudiwa chete.

05Mutengo wepamusoro unotenga chii?

Pa$1.42 anopinda / $6.07 anobuda pamatokeni miriyoni, Inkling iri pakati pekatarogu yedu nemaune. Matanhatu emamodhi gumi nemaviri ari pamutengo wepasi we$0.50 / $2.00. Kimi-K3 — modhi yepamusoro-soro, inoonawo — iri pa$3.83 / $19.12. Inkling inopetwa katatu pane yepasi pane zvinobuda uye chikamu chinenge chetatu cheyepamusoro-soro.

Mutengo wezvinobuda, madhora pamatokeni 1M, mitengo yerondedzero yeGedhi reShannon AI, Gunyana 2026. Hurefu hwemutsara hwakaenzana nemutengo; hurefu zvinoreva kudhura.

Heino kuverenga kwechokwadi kwepanokodzera mutengo wepamusoro — nepasingakodzeri.

Basa	Shandisa Inkling here?	Chikonzero
Kufungisisa pamachati / madhayagiramu	Hongu	Nhamba dzeCharXiv neMMMU Pro dzakabudiswa ndipo pane simba guru rehwaro uhu.
Maagent anotungamirwa nemasikirinishoti	Hongu	Kuona pamwe nekudana zvishandiso pamwe nenhoroondo ye262K mumodhi imwe chete, hapana kupfuudzana.
Magwaro marefu ane mifananidzo	Hongu	Mavara nemifananidzo zvinogovana dhikodha imwe chete nehwindo recontext rimwe chete.
Kugadzira mavara chete	Kwete	Mamodhi matanhatu ari pa$0.50 / $2.00 anozviita. Uri kubhadhara nzira yekuona yausingambodani.
Kubvisa ruzivo neschema pahuwandu	Kwete	Hapana json_schema. MiMo-V2.5 yakachipa uye yakaganhurirwa.
Denga repamberi chairo	Kwete	Kimi-K3 ndiyo yepamusoro-soro. Bhadhara ka3.1 kana denga riri chinodzivirira.
Kuona kwakachipa kupfuura kwose	Kwete	MiniMax-M3 inoona pa$0.50 / $2.00. Tanga ipapo, ukwire kana mhando ichizvimanikidza.

Iverenge semutemo wekutumira, kwete kurongwa. Inkling yakakodzera boka rebasa diki kupfuura zvingataurwa nepeji reshambadziro, uye mukati meboka iroro yakaoma kukundwa pamutengo. Kunze kwaro, tinofarira kukutengesera modhi yakachipa tichikuchengeta.

06Hwindo recontext re262K, uye chinonyatsokwana

Mamodhi gumi nemaviri ose ari paGedhi anoshandiswa ane context inopinda yematokeni 262,144 nezvinobuda zvakanyanya zvematokeni 16,384 pachikumbiro chimwe nechimwe. Zvakaenzana nemaune: kubata context hakufaniri kumbova bazi remodhi imwe neimwe mukodhi yako, uye kuchinja inkling kuita MiniMax-M3 mufaira rekurongedza hakufaniri kuchinja chinokwana zvakanyarara.

Ziva mukaha uripo kuenzaniswa nehwaro hwakabudiswa: Thinking Machines inonyora Inkling pamatokeni 1,048,576, uye isu tinopa chikamu che262K chaizvozvo. Kana uine basa chairo rechikumbiro chimwe chete chematokeni miriyoni, hatisi isu suwo rako, uye zviri nani uzive usati wavaka.

Pachiyero pa262K: PDF yeunyanzvi yemapeji mazana mana nemifananidzo yayo, seti yemakodhi yepakati, maawa akati wandei ezvakanyorwa, kana kumhanya kurefu kweagent kunotakura mhedzisiro yezvishandiso makumi nemasikirinishoti mashoma — chimwe nechimwe chinokwana muchikumbiro chimwe chete. Pamodhi inoona, yeuka kuti mifananidzo inodyawo context; sikirinishoti yakawandirana haisi yemahara, uye chikamu chinoyerera mafuremu chinozadza hwindo nekukurumidza kupfuura chemavara chete. Magadzirirwo anoenderana nazvo zvakanaka: chimiro che5:1 chehwindo rinotsvedza kuenda kune rakazara chinoreva kuti madanho mazhinji anotarisa mukati menzvimbo yematokeni 512, madanho akazara achidzosa chimiro chenhambwe refu nguva nenguva.

07Simba rekufungisisa: kubhadhara kufunga pauchakuda chete

Inkling inoburitsa simba rekufungisisa rinodzorwa — chimwe chezvidzoreso zvinobatsira kupfuura zvose mukatarogu. Modhi yekumusoro inogamuchira mazinga kubva panone neminimal achipfuura palow, medium, high, xhigh, kusvika pamax, achiburitswa pamitauro mitatu yose kuburikidza nepharamita yakajairika yereasoning_effort.

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

Ichi chidzoreso chemutengo chine simba, nokuti matokeni ekufungisisa anobhadhariswa semtokeni anobuda pa$6.07 pamiriyoni. Thinking Machines inoshuma kuti paTerminal Bench 2.1 Inkling inosvika pazviyero zvakaenzana neNemotron 3 Ultra pachikamu chinenge chetatu chematokeni — kushandisa matokeni zvakanaka chinangwa chakataurwa chehurongwa hwesimba, kwete chinhu chakangoitika.

Pakuita: mhanyisa low kana minimal pakunyora tsananguro, kurongedza, nekubvisa ruzivo kwakapusa, apo mhinduro iri mumufananidzo uye pasina mutsara wekufungisisa unodiwa. Chengetera high nepamusoro pekufungisisa pamachati, kuenzanisa kwezvinoonekwa kwematanho akawanda, nekutenderera kwemaagent apo danho repakati risiriro rinodhura. Mapoka mazhinji haamboisi pharamita iyi, uye anosiya musiyano wemutengo wezvinobuda unopetwa kanoverengeka pasina kuutora.

08Kuti "isina kudzorwa" inorevei pano — uye zvaisingarevi

Zviri pachena: tinomhanyisa modhi iyi pasina chidzitiro chekuramba uye pasina sefa pane zvinobuda. Hapana muongorori anoverenga prompt yako osarudza kana modhi ichibvumirwa kupindura, uye hapana anoverenga mhinduro isati yayerera. Kana modhi isina chokwadi kana isina ruzivo, inodaro — ndiyo modhi, kwete mutemo. Chausingawani kuramba kusina kufunga panyaya inotorwa nechigadzirwa chevatengi vose sechisingabvumirwi.

Pakuona ndiyo nharo yose yekukosha. Vaongorori vemifananidzo vari mberi kwemaAPI anoona havana kururama, uye chiyero chekuramba kusina musoro chinorwadza zvakanyanya pabasa reunyanzvi rinonyanya kuda unyanzvi uhwu: tsvakurudzo yezvekuchengetedza, mamiriro ekiriniki, hunyanzvi hwenhau neforensic, kuongorora mabhaisikopo neunyanzvi, nezvishandiso zvekudzora zvirimo zvinofanira kutarisa chinhu chazvinodzora. Suwo rinoramba nekuda kwechikamu haringakwanisi kushandira chero chimwe chazvo zvakavimbika.

Zvinoyambirwa zviviri zvechokwadi. Chekutanga, Thinking Machines inoburitsa ongororo dzekuchengetedza dzechivakwa chavo chekutarisa — 78.0% paFORTRESS adversarial, 95.9% paFORTRESS benign, 98.6% paStrongREJECT. Nhamba idzodzo dzinotsanangura chimiro chavo nezvidziviriro zvavo zviripo. Hadzitsananguri zvedu, uye hatidzitauri sekunge dzinodaro. Takabvisa chidzitiro chinoyerwa muchikamu nenhamba idzodzo; ndicho chigadzirwa, uye ndiwo mutoro.

Chechipiri, "hapana chidzitiro chekuramba" hakusi "hapana mitemo." Shannon Lab LLC inoshanda pasi peMutemo Wekushandisa Nemutsigo nemutemo weUS unoshanda. Kubvisa muongorori kunoendesa kutonga kwauri, hakukudzimi. Kana uri kuvaka pane izvi, verenga mutemo usati wakura.

09Kutanga

Base URL imwe cheteyo, kiyi imwe cheteyo, mitauro mitatu. Kana kodhi yako ichitodana OpenAI, Anthropic, kana Responses API, kuchinja mutsara wemodhi kunowanzova kutama kwose.

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="")

Rondedzero yakazara yemapharamita, zvimiro zve/v1/messages ne/v1/responses, zvimiro zvezviitiko zvekuyerera, nenzvimbo yekuedza iri kushanda zviri mumagwaro eAPI. Magwaro emamodhi emamwe gumi nerimwe ari muShannon Research.

Nhamba dzehwaro inhamba dzakabudiswa neThinking Machines Lab dzechekipoinzi cheInkling chisina quantization uye dzinotsanangura denga remhuri, kwete chivakwa chedu chakaitwa quantization. Mitengo nezviratidzo zveunyanzvi ndizvo zvinhu zverondedzero yeGedhi reShannon AI kusvika muna Gunyana 2026.

10Mibvunzo inowanzobvunzwa

---

Zvakatorwa kubva: 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 — zvinyorwa zvemagadzirirwo nezvemiyero zveInkling. Miyero yehwaro yakashumwa nerabhoritari yechekipoinzi chisina quantization.