I-API Yemodeli Ye-Inkling Engavinjelwe

Iningi lamamodeli ayishumi nambili esangweni lethu angombhalo kuphela. Amane awenjalo, futhi phakathi kwalawo mane i-Inkling ibamba isikhundla esithile: imodeli yombono oyifunayo lapho isithombe siyingxenye enzima yenkinga kodwa isabelomali singafiki emodelini eyinhloko. Lo mhlahlandlela ukhuluma ngokuthi isisekelo sivelaphi, ukuthi isijobelelo esithi W4A16-AUTOROUND-REAP senzeni kuso, ukuthi izokwenzani nokuthi ngeke yenzeni esangweni le-API, futhi — ngeqiniso — ukuthi yinini lapho imodeli eshibhile kuleyo khathalogu ifanelekile kakhulu.

01Iyini i-Inkling, futhi ngubani oyakhile?

I-Inkling iyimodeli yokuqala enezisindo ezivulekile evela ku-Thinking Machines Lab, eshicilelwe ngoJulayi 15, 2026. Ngokungafani neningi lokukhishwa okubekwa njengo-"okuvulekile," yafika nezisindo ezigcwele ezingalanjwa ku-Hugging Face kanye nokusekelwa kwosuku-lokuqala kuma-transformers, i-SGLang, i-vLLM, ne-llama.cpp — nayindlela eyaba khona ngayo isebenza engqalasizindeni efana neyethu. Namba eziyinhloko, ezivela kokushicilelwe okwakhiwe yi-Thinking Machines uqobo:

Lokho ukuhleleka kwe-Mixture-of-Experts okungaminyene: izingqimba ze-decoder ezingu-66 ngosayizi ofihlekile ongu-6,144, ezimbili zokuqala ziminyene, ezingu-64 ezisele ziphethe ochwepheshe abaqondiswayo abangu-256 kanye nochwepheshe abelanwayo abangu-2, kanye nochwepheshe abaqondiswayo abangu-6 abasebenzayo kuthokheni ngalinye. Cishe u-4% wamapharamitha uyasebenza kunoma yiliphi ithokheni elithile — nakuyikho okwenza imodeli eseduze nesigidigidi samapharamitha ibe nokwenzeka ukuyikhonza nakanjani.

Kunezinketho ezimbili zokwakhiwa ezingajwayelekile ngokwanele ukuba ziqanjwe. Ukunaka kuxutshiwe — izingqimba ezingu-55 zokunaka kwewindi eliqhubekayo phezu kwewindi lamathokheni angu-512 mayelana nezingqimba ezijwayelekile ezingu-11, ngephethini ephindaphindwayo engu-5:1. Futhi i-Inkling ayisebenzisi i-RoPE: isebenzisa ukuchema kwesikhundla esihambisanayo esifundiwe, esincike kokungenayo, eceleni kwama-convolution afushane abangelayo kuma-projection e-key/value nasemagatsheni asele. Amanothi okwakhiwa ka-Sebastian Raschka aveza kokubili njengokuphambuka kwangempela kusuka enqubweni ejwayelekile ka-2026. Ukuqeqeshwa kusebenzise i-Muon kumapharamitha amakhulu e-matrix ne-Adam kokuseleyo, kwase kuba nokuqhutshwa kwe-RL okungahambisani ngesikhathi okungaphezu kwezigidi ezingu-30 ekuqeqeshweni okulandelayo.

Iphuzu komsebenzisi we-API akulona ulwazi olungebona olubalulekile. Yikuthi lesi isisekelo esisezingeni eliphakeme, esinezindlela eziningi ngemvelo esiqeqeshwe ngamathokheni ayizigidigidi ezingu-45 asakaza umbhalo, izithombe, umsindo namavidiyo — hhayi imodeli yombhalo enomshuweleti wombono ohlonyiwe kamuva. Izithombe zibhalwa njengamaphetshana e-pixel angu-40×40 ngokusebenzisa i-patchifier ye-MLP enezingqimba ezine ezilandelanayo bese ziqhutshwa kanyekanye nombhalo ku-decoder efanayo. Ukuqonda ngokubona bekukuhlose kokuqeqeshwa ngaphambili.

02Kusho ukuthini u-W4A16-AUTOROUND-REAP?

Amasu amathathu ahlukene abekwe ndawonye kulesi sijobelelo, futhi enza imisebenzi ehlukene. Kubalulekile ukukuqonda, ngoba isijobelelo ungumehluko omkhulu kunazo zonke phakathi kwemodeli oyibiza ku-API yethu nechekhiphoyinti eshicilelwe yi-Thinking Machines.

I-REAP — ukuthena ochwepheshe okungekho muntu obaqondisayo

I-REAP imele u-Router-weighted Expert Activation Pruning, evela ku-Cerebras Research (arXiv 2510.13999, eyamukelwa ku-ICLR 2026). Ku-MoE enochwepheshe abangu-256 kungqimba ngalinye, i-router ayisabalalisi umsebenzi ngokulinganayo: abanye ochwepheshe bafinyelelwa njalo, abanye basebenza kancane futhi bafaka kancane lapho besebenza. I-REAP ihlola ochwepheshe ngamunye kokubili ngamanani e-router gate nangezinga lokusebenza kochwepheshe — ochwepheshe okhethwa kaningi kodwa ofaka izinguquko ezincane ubaluleke kancane kunalokho inani lokukhethwa kwakhe elikuphakamisayo — futhi ilahle labo abangabalulekile kancane ngokuhlangana kanye, ngaphambi kokukalwa kwezinga.

Isizathu sokuthi sisebenzise ukuthena kunokwenza ukuhlanganiswa kochwepheshe okuthandwa kakhudlwana ukuthi lokhu kokubili kuyahlukana ngohlobo lomsebenzi. Ukuhlanganiswa kubukeka kuncintisana kahle ezilinganisweni ezihlukanisayo kodwa kuwe phansi kwezikhiqizayo. Ababhali be-REAP babika ukuthi ukuthena kugcina ikhwalithi kangcono kanye emisebenzini abantu abathenga i-API ngayo — ukuhlanganiswa kwekhodi, ukubiza amathuluzi, ukucabanga kwemizuliswano eminingi — ngezilinganiso zokucindezela ezingaphezu kuka-50%, kuwo wonke amamodeli kusukela kumapharamitha angu-20B kuya ku-1T. Ukuhlanganiswa kucekela phansi ubuchwepheshe be-router obuqondisiwe; ukuthena kushiya ochwepheshe abasindile nokuqondiswa phakathi kwabo kungathintiwe.

I-AutoRound — ukufunda lapho okumele kuyingiliswe khona

I-AutoRound yindlela yokukala izinga ngemuva kokuqeqeshwa ka-Intel (arXiv 2309.05516, eshicilelwe njenge-SignRound). Ukukalwa kwezinga okulula kuyingilisa isisindo ngasinye kunani elingu-amabhithi amane eliseduze kunawo wonke — okuyisimo esizenzakalelayo esifanele kodwa esibi uma kwenziwa ngokubanzi, ngoba amaphutha okuyingilisa awazimele futhi ayanqwabelana kuso sonke isakhiwo esingqimba ezingu-66. I-AutoRound iphatha isinqumo sokuyingilisa njengento okumele ilungiswe: i-signed gradient descent ilungisa kanyekanye indlela yokuyingilisa nezilinganiso zokugxumeka isisindo phezu kwezigaba ezingamakhulu ambalwa, ibolekwe ukunemba kokuqeqeshwa okuqaphela ukukalwa kwezinga ngaphandle kokuqeqeshwa kabusha. I-Intel ibika inzuzo enkulu kunazo zonke ezilungiselelweni zamabhithi amancane kakhulu nelinye ilinganiselo elincintisanayo ku-INT4.

I-W4A16 — lapho amabhithi ayakhona khona

I-W4A16 ifomethi yezinombolo: izisindo zamabhithi amane, ukusebenza kwamabhithi ayi-16. Izisindo — ezibusa inkumbulo kwimodeli engu-975B — zincipha zibe cishe yikota yendawo eyayithathwa yi-BF16, kuyilapho izibalo zihlala kumabhithi ayi-16. Lokho kuhlukaniswa kubaluleke ikakhulukazi kumodeli yombono: amasu okukala izinga lokusebenza alimaza kanye umsebenzi wezinombolo ocophelele ukuqonda kwesithombe okuncike kuwo — ukufunda ilebula encane ye-axis, ukususa umugqa omncinyane kwidayagremu, ukuhlukanisa amabala amabili afanayo. Ukugcina ukusebenza emabhithini ayi-16 kubiza indawo yokudlulisa kodwa kubuyisa ukunemba lapho umsebenzi wombono uwubona.

Uma kubekwe ndawonye, lokho kokuthathu kunikeza imodeli engena futhi ehlale ishesha ngaphandle kokuwa kwekhwalithi okusheshayo esitebhisi esiqinile esithi 3BIT-REAP esisisebenzisa ezisekelweni ezinkulu kunazo zonke. Lokho ukuthengiselana intengo ekubonisayo.

03Umbono: i-API yokuhlaziya izithombe engenqabi

I-Inkling ingenye yamamodeli amane anombono ekhathalogu yethu, eceleni kwe-Kimi-K3, i-Kimi-K2.6, ne-MiniMax-M3. Imiphumela yombono eshicilelwe yemodeli eyisisekelo, evela ekukhishweni kwe-Thinking Machines, i-73.3% ku-MMMU Pro (Standard 10) ne-78.1% kwemibuzo yokucabanga ye-CharXiv — enyukela ku-82.0% lapho imodeli ivunyelwe ukuqhuba i-Python macala eshadi. Lawo mangani awalabho be-labhorethri uqobo kwechekhiphoyinti elingakaliwe; awuphathe njengophahla lomndeni, hhayi njengesilinganiso sokwakhiwa kwethu.

I-CharXiv iyiyona efanele ukugxiliswa kuyo, ngoba ihambisana nesizathu sokuthi kungani abantu bethenga i-API yombono kunesevisi ye-OCR. Ihlola ukucabanga phezu kwamashadi esayensi — ukufunda inani ku-axis, ukuqhathanisa uchungechunge oluphindiwe, ukuqaphela indlela eguqukayo. Hhayi "yimuphi umbhalo osesithombeni" kodwa "kusho ukuthini leli shadi, futhi ingabe lokho okuthiwayo kuqinisile." Yilokho okufanelanisa ukuthumela isithombe kumodeli yokucabanga engu-975B kunokuya kumshicileli omncane.

Ukuthumela isithombe ngumsebenzi ojwayelekile ohambisana ne-OpenAI — ingxenye yokuqukethwe ye-image_url ku-/v1/chat/completions, ibhulokhi yesithombe ngendlela ye-Anthropic ku-/v1/messages, noma okulinganayo ku-/v1/responses. Zontathu izilimi zifinyelela kuzo zonke izisindo ezifanayo futhi zonke ziyasakaza; khetha noma iyiphi eye-SDK yakho esesikhona esikhulumayo. Isibonelo esisebenzayo sikwisigaba 09.

Ingxenye engavinjelwe ibaluleke kakhulu embonweni kunombhalweni. Ungqimba lokwenqaba kusiphetho sombono lubuthuntu ngendlela engajwayelekile, ngoba ngokuvamile luhlukanisa isithombe ngaphambi kokuba imodeli icabange ngesicelo. Umsebenzi osemthethweni ubanjwa njalo: isithombe senxeba kupayipi lamanothi omtholampilo, isithombe-skrini sokuhlolwa kwezobufakazi embikweni wesigameko, ifreyimu evela efilimini ehlaziywa ombhalweni wobuhlakani, ukuskena kwezokwelapha, isithombe sedokhumenti esiqukethe igama ngephutha. I-Inkling esangweni lethu ayinaso isihlukanisi esinjalo phambi kwayo. Ibheka lokho okuthumele bese iphendula umbuzo owubuzile. Lokho kungumkhiqizo.

04Okukhishwayo okuhlelekile: u-response_format yebo, u-json_schema cha

Yilapha kuphela lapho i-Inkling ingenamandla amancane kunemiqoka yekhathalogu, futhi kungcono ukukuhlelela kunokukuthola usukwenza umsebenzi wangempela.

Ikhono	inkling	Kimi-K2.6	MiMo-V2.5	Laguna-S-2.1
Umbono (ukungenisa izithombe)	Yebo	Yebo	Yebo	Cha
response_format	Yebo	Yebo	Yebo	Cha
json_schema	Cha	Yebo	Yebo	Cha
Ukubiza amathuluzi	Yebo	Yebo	Yebo	Yebo
Umzamo wokucabanga	Yebo	Yebo	Yebo	Yebo

U-response_format: {"type": "json_object"} uyasebenza — uthola i-JSON esemthethweni ngokwesakhiwo. Okungathole yikho ukukhishwa okukhawulelwe uhlelo ngokumelene nohlelo owanikeza lona, lapho i-sampler ivinjelwe ngokoqobo ekukhipheni ithokheni ephula ukwakheka kwakho. Amagama emikhawulo, ubulungu bwe-enum, kanye nokuba khona kwezikhiye ezidingekayo kuyisinqumo semodeli, hhayi isiqinisekiso.

Izindlela ezintathu zokubhekana nalokho, ngokwelinganiso lokuthembeka elikhuphukayo:

1. Faka uhlelo ku-prompt bese uqinisekisa. Faka uhlelo ku-system prompt njenge-JSON eqondile, misa u-response_format ube ngu-json_object, bese uqinisekisa impendulo ohlangothini lwakho. Zama futhi uma ukuqinisekisa kwehluleka, unamathisele umlayezo wephutha lomqinisekisi. Kushibhile, kuyasebenza, kudinga isabelo sokuzama futhi.
2. Sebenzisa incazelo yethuluzi njengohlelo. Ukubiza amathuluzi kuyasekelwa, futhi incazelo yepharamitha yomsebenzi iyi-JSON Schema. Memezela ithuluzi elilodwa eliqanjwe ngokwenani olifunayo, ulincindezele, bese ufunda izimpikiswano. Le ngokuvamile iyona ndlela enamandla kunazo zonke ku-Inkling futhi ayikubizi lutho olwengeziwe.
3. Hlukanisa umsebenzi phakathi kwamamodeli amabili. Yenza i-Inkling yenze umbono nokucabanga ngombhalo ojwayelekile, bese udlulisela lowo mbhalo ku-MiMo-V2.5-W8A16 ku-$0.50 / $2.00 ngo-json_schema wangempela womzuliswano wokwakha uhlelo. Amakholi amabili, kodwa elesibili lishibhile, linesimo esiqinisekile, futhi imodeli okumele ihlakaniphile yileyo eyabona isithombe.

Inketho 3 iyiphethini esiyincoma kunoma iyiphi ipayipi yokukhipha ulwazi ngobuningi obuphezulu. Khokhela intengo ephakeme kuphela isinyathelo esiyidingayo.

05Yini ekhokhelwa yintengo ephakeme?

Ku-$1.42 engenayo / $6.07 ephumayo ngethokheni eyisigidi, i-Inkling ihlala ngamabomu maphakathi nekhathalogu yethu. Ayisithupha kwayishumi nambili amamodeli asezingeni eliphansi elingu-$0.50 / $2.00. I-Kimi-K3 — eyinhloko, enombono nayo — ingu-$3.83 / $19.12. I-Inkling icishe kuphindwe kathathu izinga eliphansi kokukhishwayo futhi cishe ingxenye eyodwa kwezintathu yeyinhloko.

Intengo yokukhishwayo, ngo-USD ngamathokheni ayisigidi, amanani ohlu lwe-Shannon AI Gateway, Septhemba 2026. Ububanzi bebha bulingana nentengo; ende kakhulu ibiza kakhulu.

Nasi ukufunda okuqotho kokuthi leyo ntengo ephakeme izuza kuphi — nokuthi ayizuzi kuphi.

Umsebenzi	Sebenzisa i-Inkling?	Kungani
Ukucabanga ngamashadi / amadayagremu	Yebo	Amanani ashicilelwe e-CharXiv ne-MMMU Pro yilapho leli sisekelo linamandla kakhulu khona.
Ama-ejenti aqhutshwa izithombe-skrini	Yebo	Umbono kanye nokubiza amathuluzi kanye nomlando we-262K kumodeli eyodwa, ngaphandle kokudlulisela.
Amadokhumenti amade anemidwebo	Yebo	Umbhalo nezithombe kwabelana nge-decoder eyodwa newindi lomongo elilodwa.
Ukukhiqiza umbhalo ojwayelekile	Cha	Amamodeli ayisithupha ku-$0.50 / $2.00 enza lokhu. Ukhokhela indlela yombono ongayibizi nhlobo.
Ukukhishwa kohlelo ngobuningi	Cha	Alikho i-json_schema. I-MiMo-V2.5 ishibhile futhi ikhawulelwe.
Uphahla oluphakeme ngempela	Cha	I-Kimi-K3 iyinhloko. Khokha okuphindwe ka-3.1 uma uphahla luyisithiyo.
Umbono oshibhile kunazo zonke	Cha	I-MiniMax-M3 inombono ku-$0.50 / $2.00. Qala lapho, khuphukele phezulu uma ikhwalithi iyakuphoqa.

Kufunde njengomthetho wokuqondisa, hhayi njengokuhlelwa ngezinga. I-Inkling ilungele ibanga elincane lomsebenzi kunalokho ikhasi lokumaketha elingasho, futhi ngaphakathi kwalelo banga kunzima ukuyeqa ngentengo. Ngaphandle kwalo, singathanda ukukuthengisela imodeli eshibhile futhi sikugcine.

06Iwindi lomongo we-262K, nokuthi yini engena ngempela

Wonke amamodeli ayishumi nambili esangweni akhonzwa ngomongo wokungenayo ongamathokheni angu-262,144 nokukhishwayo okuphezulu okungamathokheni angu-16,384 ngesicelo ngasinye. Kulinganiswe ngamabomu: ukuphathwa komongo akufanele neze kube ligatsha ngamodeli ngayinye ekhodini yakho, futhi ukushintshanisa u-inkling nge-MiniMax-M3 ku-config file akufanele kushintshe ngokuthula lokho okungenayo.

Qaphela isikhala ekuqhathaniswa nesisekelo esishicilelwe: i-Thinking Machines ibhala i-Inkling ku-1,048,576 amathokheni, futhi thina sikhonza incenye engu-262K yalokho. Uma unomsebenzi wangempela ongamathokheni ayisigidi ngesicelo esisodwa, asisona isiphetho sawo, futhi kungcono ukuthi ukwazi ngaphambi kokuba wakhe.

Ukuze ubone isilinganiso ku-262K: i-PDF yobunjiniyela enamakhasi angu-400 kanye nemidwebo yayo, isizinda sekhodi esiphakathi nendawo, amahora amaningana ombhalo wengxoxo, noma isikhathi eside sokusebenza kwe-ejenti sithwele imiphumela yamathuluzi engamashumi ambalwa kanye nezithombe-skrini ezimbalwa — kungena ngasinye esicelweni esisodwa. Kumodeli yombono, khumbula ukuthi izithombe nazo zisebenzisa umongo; isithombe-skrini esiminyene asikhokhelwa mahhala, futhi isikhathi sokusebenza esisakaza amafreyimu sigcwalisa iwindi ngokushesha kunesingumbhalo kuphela. Ukwakhiwa kufanele lokhu kahle: iphethini engu-5:1 ephakathi kwewindi eliqhubekayo nendawo ejwayelekile isho ukuthi izingqimba eziningi zinaka phakathi kwendawo ye-512-token, kanti izingqimba ezijwayelekile zihlwitha ngokuphindaphindiwe isakhiwo sebanga elide zilibuyisele.

07Umzamo wokucabanga: ukukhokhela ukucabanga kuphela lapho ukudinga

I-Inkling iveza umzamo wokucabanga olawulekayo — enye yezinkinobho eziwusizo kakhulu ekhathalogu. Imodeli yasekuqaleni yamukela amazinga kusuka ku-none ne-minimal kuya ku-low, medium, high, xhigh, kuze kufike ku-max, aveza kuzo zontathu izilimi ngepharamitha ejwayelekile ethi reasoning_effort.

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

Lokhu kuyisilawuli sezindleko esinamandla ngempela, ngoba amathokheni okucabanga akhokhiswa njengamathokheni aphumayo ku-$6.07 ngesigidi. I-Thinking Machines ibika ukuthi ku-Terminal Bench 2.1 i-Inkling ifinyelela amaphuzu alingana ne-Nemotron 3 Ultra ngamathokheni angu-ingxenye eyodwa kwezintathu — ukusebenza kahle kwamathokheni kuyinjongo yokwakhiwa eshiwo yohlelo lomzamo, hhayi ingozi.

Ekusebenzeni: qhuba u-low noma u-minimal ekubhaleni izihloko, ekuhlukaniseni, nasekukhipheni ulwazi okuqondile, lapho impendulo isesithombeni futhi kungadingekile uchungechunge lokucabangela. Gcinela u-high nangaphezulu ekucabangeni ngamashadi, ekuqhathaniseni ngokubona okunezinyathelo eziningi, nasemizuliswaneni ye-ejenti lapho isinyathelo esiphakathi esingalungile sibiza kakhulu. Amaqembu amaningi awuze awumise nakanye lelo phalamitha, futhi ashiya umehluko wezindleko zokukhishwayo ophindwe izikhathi eziningi etafuleni.

08Kusho ukuthini u-"okungavinjelwe" lapha — nokuthi akusho ukuthini

Ngokuqondile: siqhuba le modeli ngaphandle kongqimba lokwenqaba nangaphandle kokuhlungwa kokukhishwayo. Alikho isihlukanisi esifunda i-prompt yakho bese sinquma ukuthi ingabe imodeli ingaphendula, futhi asikho esifunda impendulo ngaphambi kokuba isakaze. Uma imodeli inganqumile noma ingenalo ulwazi, iyakusho lokho — leyo imodeli, hhayi inqubomgomo. Okungeke ukuthole yikwenqaba ngokuzenzakalelayo esihlokweni umkhiqizo wabathengi abavamile owubheka njengongavunyelwe.

Embonweni lokhu kuyiwo wonke umhlahlandlela wenani. Izihlukanisi zezithombe ezingaphambi kwama-API ombono azinembile, futhi izinga lamaphutha okucabangela into engekho kanjalo liwela kanzima kakhulu emsebenzini wobuchwepheshe odinga kakhulu leli khono: ucwaningo lokuphepha, izimo zomtholampilo, ubuntatheli nezobufakazi, ukugxekwa kwamafilimu nobuciko, amathuluzi okulawula okuqukethwe okumele abheke into ewulawulayo. Isiphetho esenqaba ngohlobo asikwazi ukukhonza noma yikuphi kwalokho ngokuthembekile.

Izexwayiso ezimbili eziqotho. Okokuqala, i-Thinking Machines ishicilela ukuhlolwa kokuphepha kokukhonzwa kwabo okuyisibonelo — u-78.0% ku-FORTRESS adversarial, u-95.9% ku-FORTRESS benign, u-98.6% ku-StrongREJECT. Lawo manani achaza uhlaka lwabo nezivikelo zabo zikhona. Awachazi olwethu, futhi ngeke sikubale sengathi kunjalo. Sisusile ungqimba lawo manani aluhlola ingxenye; lokho kungumkhiqizo, nomthwalo wesibopho.

Okwesibili, u-"ngaphandle kongqimba lokwenqaba" akusho u-"ngaphandle kwemithetho." I-Shannon Lab LLC isebenza ngaphansi kwe-Nqubomgomo Yokusebenzisa Ngokuzibophezela nomthetho kaHulumeni wase-US osebenzayo. Ukususa isihlukanisi kudlulisela isahlulelo kuwe, akusicimi. Uma wakha ngale ndlela, funda inqubomgomo ngaphambi kokukhulisa.

09Ukuqala

I-URL efanayo eyisisekelo, ukhiye ofanayo, izilimi ezintathu. Uma ikhodi yakho isivele ibiza i-OpenAI, i-Anthropic, noma i-Responses API, ukushintsha inqama yemodeli ngokuvamile kuyikho konke okudingekayo ukufuduka.

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="")

Uhlu olugcwele lwamaphalamitha, izimo ze-/v1/messages ne-/v1/responses, amafomethi ezenzakalo zokusakaza, kanye nendawo yokudlala ephilayo kuse-khasini lemibhalo ye-API. Imihlahlandlela yamamodeli ayishumi nanye esele iku-Shannon Research.

Amanani esisekelo angamanani ashicilelwe yi-Thinking Machines Lab wechekhiphoyinti le-Inkling elingakaliwe futhi achaza uphahla lomndeni, hhayi ukwakhiwa kwethu okukaliwe. Amanani entengo namafulegi amakhono angamanani ohlu lwe-Shannon AI Gateway kusukela ngoSepthemba 2026.

10Imibuzo ebuzwa njalo

---

Imithombo: Thinking Machines Lab — Iyethula i-Inkling · Hugging Face — Wamukela i-Inkling · REAP the Experts (arXiv 2510.13999) · Optimize Weight Rounding via Signed Gradient Descent (arXiv 2309.05516) · Sebastian Raschka — Amanothi wokwakhiwa nokulinganiswa kwe-Inkling. Ama-benchmark esisekelo abikwe yi-labhorethri wechekhiphoyinti elingakaliwe.