I-API yeModeli ye-Inkling Engacinezelwanga

Uninzi lweemodeli ezilishumi elinambini kwiSango lethu zezombhalo kuphela. Ezine azinjalo, kwaye phakathi kwezo zine i-Inkling ibambe indawo ethile: yimodeli yombono oyifikelelayo xa umfanekiso ingoyona nxalenye inzima yengxaki kodwa uhlahlo-lwabiwo-mali lungafikeleli kweyona iphambili. Esi sikhokelo sigubungela apho isiseko sivela khona, into impumezelo ye-W4A16-AUTOROUND-REAP eyenzileyo kuso, into eza kuyenza nengasayi kuyenza kumphezulu we-API, kunye — ngokunyanisekileyo — nexesha imodeli etshiphu kwikhathalogu enye ilukhetho olungcono ngalo.

01Yintoni i-Inkling, kwaye yakhiwe ngubani?

I-Inkling yimodeli yokuqala enobunzima obuvulekileyo evela kwi-Thinking Machines Lab, epapashwe nge-15 Julayi 2026. Ngokungafaniyo nokukhutshwa okuninzi okubekwa "njengokuvulekileyo", yeza nobunzima obupheleleyo obunokukhutshelwa kwi-Hugging Face nenkxaso yosuku lokuqala kwi-transformers, i-SGLang, i-vLLM ne-llama.cpp — nto leyo eyenza yakwazi ukuqhutywa kuluhlaka olunjengolwethu kwaphela. Amanani aphambili, avela kwizinto zokukhutshwa ze-Thinking Machines ngokwazo:

Bubume be-Mixture-of-Experts esasazekileyo obo: iimaleko ze-decoder ezingama-66 kubungakanani obufihlakeleyo obungama-6,144, ezimbini zokuqala zishinyene, ezingama-64 eziseleyo zithwala iingcali ezingama-256 ezirawutwayo kunye neengcali ezimbini ezabelwanayo, ngeengcali ezintandathu ezirawutwayo ezisebenzayo kwitoken nganye. Malunga nama-4% eeparamitha adubulayo kuyo nayiphi na itoken enikiweyo — nto leyo eyenza imodeli ephantse ibe neeparamitha ezisithriliyoni ikwazi ukuhanjiswa kwaphela.

Izigqibo ezibini zokwakha zingaqhelekanga ngokwaneleyo ukuba zibizwe. Ingqwalasela iphithikeziwe — iimaleko ezingama-55 zengqwalasela yefestile etyibilikayo kwifestile yeetoken ezingama-512 ngokuchasene neemaleko zehlabathi ezili-11, kwipatheni ephindaphindayo ye-5:1. Kwaye i-Inkling ayisebenzisi i-RoPE: isebenzisa ityekelo lendawo ehambelanayo elifundiweyo, elixhomekeke kwizinto ezingenayo, ecaleni kweekhonvolyushini ezimfutshane zonobangela kwiiprojekshini zesitshixo/ixabiso nakumasebe aseleyo. Amanqaku esakhiwo ka-Sebastian Raschka aphawula omabini njengokwahluka kokwenene kwiresiphi eqhelekileyo yowama-2026. Uqeqesho lwasebenzisa i-Muon kwiiparamitha zematriki enkulu ne-Adam kwezinye, emva koko izigqibo ze-RL ezingangqubaniyo ezingaphezu kwezigidi ezingama-30 emva koqeqesho.

Inqaku kumthengi we-API ayizizo iindaba nje. Kukuba esi sisiseko sobungakanani bomphambili, esineendlela ezininzi zemvelo esaqeqeshwa kwiitoken ezisithriliyoni ezingama-45 ezibandakanya umbhalo, imifanekiso, isandi neevidiyo — ingekuko imodeli yombhalo enesandiso sombono esixotyelelwe kamva. Imifanekiso ikhowudwa njengeziqwenga ze-40×40 ze-pixel ngomahluli weziqwenga we-MLP wamaleko mane onamanqanaba yaye icutshungulwa kunye nombhalo kwi-decoder enye. Ukuqonda okubonwayo kwakukwinjongo yoqeqesho lwangaphambili.

02Ithetha ntoni i-W4A16-AUTOROUND-REAP?

Iindlela ezintathu ezahlukeneyo zibekwe phezu komnye kuloo mpumezelo, kwaye zenza imisebenzi eyahlukeneyo. Kufanelekile ukuqonda, kuba impumezelo ngoyena mahluko mkhulu phakathi kwemodeli oyibizayo kwi-API yethu nendawo yokugcina i-Thinking Machines eyayipapashayo.

REAP — ukuthena iingcali ekungekho mntu urawuta kuzo

I-REAP yi-Router-weighted Expert Activation Pruning, evela kwi-Cerebras Research (arXiv 2510.13999, eyamkelwa kwi-ICLR 2026). Kwi-MoE eneengcali ezingama-256 kwimaleko nganye irawuta ayiwabi umsebenzi ngokulinganayo: ezinye iingcali zichukunyiswa rhoqo, ezinye zidubula kanqabe kwaye zifake isandla esincinci xa zisenza njalo. I-REAP inika inqaku ingcali nganye kuwo omabini amaxabiso egeyithi yerawuta nemilinganiselo yokusebenza kwengcali — ingcali ekhethwa rhoqo kodwa efaka izilungiso zobungakanani obuncinci ibaluleke ngaphantsi kokuba inani lokukhethwa kwayo libonisa — ize ilahle ezinokubaluleka okusezantsi kwisitshotsho esinye, phambi kokucutha.

Isizathu sokuba sisebenzise ukuthena endaweni yokudibanisa iingcali okuthandwa ngakumbi kukuba ezi zimbini ziyahluka ngohlobo lomsebenzi. Ukudibanisa kubonakala kukhuphisana kwiibalo-mgangatho zokwahlula kwaye kuyaphuka kwezokuvelisa. Ababhali be-REAP baxela ukuba ukuthena kugcina umgangatho ngcono kanye kwimisebenzi abantu abathenga i-API ngenxa yayo — ukwakhiwa kwekhowudi, ukubiza izixhobo, ukuqiqa kwemijikelo emininzi — kwimilinganiselo yokucinezela edlula kuma-50%, kwiimodeli ezisuka kuma-20B ukuya kwi-1T yeeparamitha. Ukudibanisa kutshabalalisa ubuchule berawuta; ukuthena kushiya iingcali ezisindileyo nokurawuta phakathi kwazo kuphelele.

AutoRound — ukufunda apho kufuneka kujikelezwe khona

I-AutoRound yindlela ye-Intel yokucutha emva koqeqesho (arXiv 2309.05516, epapashwe njenge-SignRound). Ukucutha okungenabuchule kujikelezela ubunzima ngabunye kwelona xabiso lifuphi leebhithi ezine elinokumelwa — into engagqibekanga enengqiqo nembi kubungakanani obukhulu, kuba iimpazamo zokujikeleza azizimelanga kwaye ziyaqokelela kwisitaki seemaleko ezingama-66. I-AutoRound iphatha isigqibo sokujikeleza njengento emayilungiselelwe: ukwehla kwesantya okunophawu kutuna kunye icala lokujikeleza neenqanaba zokunqumla ubunzima kumanyathelo amakhulu amabini, ithatha ukuchaneka koqeqesho oluqonda ukucutha ngaphandle koqeqesho olutsha. I-Intel ixela ukuphucuka okukhulu kunazo zonke kwiisetingi zeebhithi ezisezantsi kakhulu kunye nokukhuphisana kwi-INT4.

W4A16 — apho iibhithi ziya khona

I-W4A16 yifomathi yamanani: ubunzima beebhithi ezine, ukusebenza kweebhithi ezilishumi elinesithandathu. Ubunzima — obulawula inkumbulo kwimodeli ye-975B — bunciphela buye kwisiqingatha sesine sobungakanani be-BF16, ngoxa izibalo zihlala kwiibhithi ezilishumi elinesithandathu. Loo mahluko ubalulekile kwimodeli yombono ngokukodwa: amacebo acutha ukusebenza onakalisa kanye umsebenzi wamanani ocokisekileyo ukuqonda umfanekiso okuxhomekeke kuwo — ukufunda ileyibhile encinci ye-eksisi, ukusombulula umgca omncinci kumzobo, ukwahlula imibala emibini efanayo. Ukugcina ukusebenza kwiibhithi ezilishumi elinesithandathu kubiza ububanzi kwaye kubuyisa ukuchaneka apho umsebenzi wombono ukuqaphelayo.

Xa zibekwe phezu komnye, ezi zintathu zinika imodeli elinganayo ehlala ikhawuleza ngaphandle kokuwa komgangatho kwenqanaba elinamandla le-3BIT-REAP esilisebenzisa kwezona ziseko zinkulu. Yiyo leyo intengiselwano ebonakaliswa lixabiso.

03Umbono: i-API yohlalutyo lomfanekiso engenakwala

I-Inkling yenye yeemodeli ezine ezinombono kwikhathalogu yethu, ecaleni kwe-Kimi-K3, i-Kimi-K2.6, ne-MiniMax-M3. Iziphumo zombono ezipapashiweyo zemodeli esisiseko, ezivela ekukhutshweni kwe-Thinking Machines, ngama-73.3% kwi-MMMU Pro (Standard 10) nama-78.1% kwimibuzo yokuqiqa ye-CharXiv — inyuka iye kuma-82.0% xa imodeli ivunyelwa ukuqhuba i-Python ngokuchasene netshati. Lawo ngamanani elebhu ngokwayo endaweni yokugcina engacutshwanga; wathathe njengophahla losapho, ingekuko ulinganiselo lwesakhiwo sethu.

I-CharXiv yeyona ifanele ukugxilwa kuyo, kuba ihambelana nesizathu sokuba abantu bathenge i-API yombono endaweni yenkonzo ye-OCR. Ivavanya ukuqiqa phezu kweetshati zesayensi — ukufunda ixabiso kwi-eksisi, ukuthelekisa amaqela amabini, ukuqaphela ukujika kwendlela. Ayisiyi "nguwuphi umbhalo okulo mfanekiso" kodwa "ibanga ntoni lo mzobo, kwaye ingaba ibango liyahlala na". Yiyo leyo ithethelela ukuthumela umfanekiso kwimodeli yokuqiqa ye-975B endaweni yombhali wezihloko omncinci.

Ukuthumela umfanekiso ngumsebenzi oqhelekileyo ohambelana ne-OpenAI — inxalenye yesiqulatho ye-image_url kwi-/v1/chat/completions, ibhloko ye-image yohlobo lwe-Anthropic kwi-/v1/messages, okanye into elinganayo kwi-/v1/responses. Zontathu iilwimi zibetha kubunzima obunye kwaye zontathu ziyasasaza; khetha nayiphi na esele ithethwa yi-SDK yakho ekhoyo. Umzekelo osebenzileyo ukwicandelo 09.

Inxalenye engacinezelwanga ibaluleke ngakumbi kumbono kunakumbhalo. Umaleko wokwala kwi-endpoint yombono ungqale ngokungaqhelekanga, kuba udla ngokwahlula umfanekiso ngeendidi phambi kokuba imodeli iqiqe ngesicelo. Umsebenzi osemthethweni ubanjwa rhoqo: ifoto yenxeba kumgca wokusebenza wamanqaku eklinikhi, umfanekiso wesikrini woqhaqho kwingxelo yesiganeko, isikrini kwifilimu ehlalutywayo ngenqaku lokugxeka, isikeni sonyango, umfanekiso woxwebhu othi ube nelinye igama. I-Inkling kwiSango lethu ayinasahluli-ndidi sinjalo phambi kwayo. Ijonga into oyithumeleyo ize iphendule umbuzo owubuzileyo. Yiyo leyo imveliso.

04Imveliso ecwangcisiweyo: i-response_format ewe, i-json_schema hayi

Yindawo enye le apho i-Inkling inesakhono esingaphantsi kunekhathalogu, kwaye kungcono ukuyicwangcisela endaweni yokuyifumana kwimveliso.

Isakhono	inkling	Kimi-K2.6	MiMo-V2.5	Laguna-S-2.1
Umbono (umfanekiso ongenayo)	Ewe	Ewe	Ewe	Hayi
response_format	Ewe	Ewe	Ewe	Hayi
json_schema	Hayi	Ewe	Ewe	Hayi
Ukubiza izixhobo	Ewe	Ewe	Ewe	Ewe
Amandla okuqiqa	Ewe	Ewe	Ewe	Ewe

I-response_format: {"type": "json_object"} iyasebenza — ufumana i-JSON esemthethweni ngokwesivakalisi. Into ongayifumaniyo kukukhutshwa okuthintelwe ngesigrama ngokuchasene nesikimu osinikayo, apho isampuli ithintelwa ngokoqobo ekukhupheni itoken eyaphula ubume bakho. Amagama eendawo, ubulungu be-enum, nobukho bezitshixo ezifunekayo lukhetho lwemodeli, ingesiso isiqinisekiso.

Ziintathu iindlela zokuphila noko, ngokulandelelana kokunyuka kokuthembeka:

1. Yalela isikimu uze uqinisekise. Beka isikimu kumyalelo wenkqubo njenge-JSON yoqobo, umisele i-response_format ibe yi-json_object, uze uqinisekise impendulo ngakwelakho icala. Zama kwakhona ekusileleni koqinisekiso ngomyalezo wempazamo womqinisekisi wongeziwe. Kutshiphu, kuyasebenza, kufuna uhlahlo-lwabiwo-mali lokuzama kwakhona.
2. Sebenzisa inkcazelo yesixhobo njengesikimu. Ukubiza izixhobo kuyaxhaswa, kwaye inkcazelo yeparamitha yomsebenzi yiJSON Schema. Xela isixhobo esinye esibizwa ngokukhupha okufunayo, usinyanzele, uze ufunde iingxoxo. Oku kudla ngokuba lolona khetho lunamandla kwi-Inkling kwaye alukubizi nto eyongezelelweyo.
3. Yahlula umsebenzi phakathi kweemodeli ezimbini. Yenza i-Inkling yenze umbono nokuqiqa ngombhalo, uze unike loo mbhalo kwi-MiMo-V2.5-W8A16 nge-$0.50 / $2.00 nge-json_schema yokwenene ngodlulo lokucwangcisa. Imibizo emibini, kodwa owesibini utshiphu, umilo wawo umisiwe, kwaye imodeli ekufuneka ikrelekrele yiloo nto ibone umfanekiso.

Ukhetho lwesithathu yipatheni ebesiya kuyincomela nawuphi na umgca wokusebenza wokukhupha womthamo omkhulu. Hlawula ixabiso eliphezulu kwinyathelo elilifunayo kuphela.

05Lithenga ntoni ixabiso eliphezulu?

Nge-$1.42 engenayo / $6.07 ephumayo kwitoken enye yezigidi, i-Inkling ihlala ngabom embindini wekhathalogu yethu. Ezintandathu kwiimodeli ezilishumi elinambini zisemgangathweni osezantsi we-$0.50 / $2.00. I-Kimi-K3 — ephambili, nayo inombono — yi-$3.83 / $19.12. I-Inkling iphindwe kathathu malunga kumgangatho osezantsi kwezimphumayo kwaye isisithathu malunga sephambili.

Ixabiso lemveliso, ngeedola zase-US kwitoken enye yezigidi, imilinganiselo yoluhlu lweSango le-Shannon AI, Septemba 2026. Ububanzi bomqolo buhambelana nexabiso; obude bubiza ngakumbi.

Nantsi ingxelo enyanisekileyo yendawo elo xabiso eliphezulu elizizuza kuyo — nendawo elingazizuzi kuyo.

Umsebenzi	Sebenzisa i-Inkling?	Kutheni
Ukuqiqa ngeetshati / imizobo	Ewe	Amanani e-CharXiv ne-MMMU Pro apapashiweyo yiyo into esi siseko sinamandla kuyo.
Iimeli eziqhutywa yimifanekiso yesikrini	Ewe	Umbono nokubiza izixhobo nembali ye-262K kwimodeli enye, akukho kunikezelana.
Amaxwebhu amade anemizobo	Ewe	Umbhalo nemifanekiso zabelana nge-decoder enye nefestile yomxholo enye.
Ukuvelisa umbhalo olula	Hayi	Iimodeli ezintandathu nge-$0.50 / $2.00 ziyayenza le nto. Uhlawulela indlela yombono ongayibizi.
Ukukhupha ngesikimu ngobuninzi	Hayi	Akukho json_schema. I-MiMo-V2.5 itshiphu kwaye ithintelwe.
Uphahla lomphambili olupheleleyo	Hayi	I-Kimi-K3 yeyona iphambili. Hlawula izihlandlo eziyi-3.1 ukuba uphahla lusithintelo.
Umbono otshiphu kunazo zonke	Hayi	I-MiniMax-M3 inombono nge-$0.50 / $2.00. Qala apho, unyuke ukuba umgangatho ukunyanzela.

Yifunde loo nto njengomthetho wokurawuta, ingekuko njengoluhlu lokuhlela. I-Inkling ichanekile kubanzi obumxinwa ngakumbi bomsebenzi kunokuba iphepha lokuthengisa belingabanga, kwaye ngaphakathi kobo banzi kunzima ukuyibetha ngexabiso. Ngaphandle kwabo, besiya kuthanda ukukuthengisela imodeli etshiphu size sikugcine.

06Ifestile yomxholo ye-262K, nento elingana ngokwenene

Zonke iimodeli ezilishumi elinambini kwiSango zihanjiswa ngomxholo ongenayo weetoken ezingama-262,144 kunye nemveliso ephezulu yeetoken ezingama-16,384 kwisicelo ngasinye. Kufana ngabom: ukuphathwa komxholo akufanele ukuze kube lisebe lemodeli nganye kwikhowudi yakho, kwaye ukutshintsha i-inkling ngeyi-MiniMax-M3 kwifayile yocwangciso akufanele kutshintshe ngokuthe cwaka into elinganayo.

Qaphela umsantsa xa kuthelekiswa nesiseko esipapashiweyo: i-Thinking Machines ibhala i-Inkling kwiitoken ezingama-1,048,576, kwaye thina sihambisa isilayi se-262K kuko. Ukuba unomsebenzi wokwenene wesicelo esinye seetoken ezisigidi, asisiyo i-endpoint yaso, kwaye kungcono uyazi phambi kokuba wakhe.

Ngobungakanani kwi-262K: i-PDF yobuchwepheshe yamaphepha angama-400 nemizobo yayo, uvimba wekhowudi ophakathi, amaxesha ambalwa ombhalo weseshoni, okanye iseshoni ende ye-agentic ethwala amashumi eziphumo zezixhobo nembalwa yemifanekiso yesikrini — nganye ilingana kwisicelo esinye. Kwimodeli yombono, khumbula ukuba imifanekiso ityakwaye umxholo; umfanekiso wesikrini oshinyeneyo awusimahla, kwaye iseshoni esasaza izikrini izalisa ifestile ngokukhawuleza ngaphezu kwaleyo yombhalo kuphela. Isakhiwo silungele oku kakuhle: ipatheni ye-5:1 yefestile etyibilikayo kweyehlabathi ithetha ukuba uninzi lweemaleko lunika ingqwalasela ngaphakathi kwendawo yeetoken ezingama-512, iimaleko zehlabathi zibuyisa ubume bomgama omde ngamaxesha athile.

07Amandla okuqiqa: ukuhlawulela ukucinga xa ukudinga kuphela

I-Inkling iveza amandla okuqiqa alawulekayo — elinye lamaqhosha aluncedo ngakumbi kwikhathalogu. Imodeli yasekuqaleni yamkela amanqanaba asuka kwi-none ne-minimal adlula kwi-low, i-medium, i-high, i-xhigh, ukuya kwi-max, ebonakaliswa kuzo zontathu iilwimi ngeparamitha eqhelekileyo ye-reasoning_effort.

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

Leli liqhosha lexabiso elinamazinyo, kuba iitoken zokuqiqa zibhatalwa njengeetoken eziphumayo nge-$6.07 kwitoken enye yezigidi. I-Thinking Machines ixela ukuba kwi-Terminal Bench 2.1 i-Inkling ifikelela kumanqaku alinganayo nawe-Nemotron 3 Ultra kwisithathu malunga seetoken — ukusebenza kakuhle kweetoken yinjongo yoyilo echaziweyo yenkqubo yamandla, ingekuko ingozi.

Ekwenzeni: qhuba i-low okanye i-minimal ngokubhala izihloko, ukwahlula ngeendidi, nokukhupha okuthe ngqo, apho impendulo isemfanekisweni yaye akukho khonkco lokuqiqa lifunekayo. Gcinela i-high nangaphezulu ukuqiqa ngeetshati, uthelekiso lokubonwayo lwamanyathelo amaninzi, nemijikelo ye-agentic apho inyathelo eliphakathi elingachanekanga libiza. Uninzi lwamaqela alusoze luyimisele le paramitha kwaphela, luze lushiye umahluko wexabiso lemveliso ophindwe amaxesha ambalwa etafileni.

08Kuthetha ntoni "ukungacinezelwa" apha — nento kungakuthethiyo

Ngokuthe ngqo: siqhuba le modeli ngaphandle komaleko wokwala nangaphandle kokuhluzwa kwimveliso. Akukho sahluli-ndidi sifunda umyalelo wakho size sigqibe ukuba imodeli ingaphendula na, kwaye akukho nasinye sifunda impendulo phambi kokuba isasazwe. Xa imodeli ingaqinisekanga okanye ingenalo ulwazi, iyatsho — yimodeli leyo, ingesiso isigqibo somgaqo-nkqubo. Into ongasayi kuyifumana kukwala okungenangqondo kumxholo imveliso yabathengi ngokubanzi ewuphatha njengongavumelekanga.

Kumbono le yiyo yonke into exabisekileyo. Izahluli-ndidi zemifanekiso phambi kwee-API zombono azichanekanga, kwaye umlinganiselo weempazamo ufika nzima kakhulu kumsebenzi wobungcali odinga kakhulu esi sakhono: uphando lokhuseleko, iimeko zeklinikhi, intatheli noqhaqho, ukugxekwa kweefilimu nobugcisa, izixhobo zolawulo lomxholo ekufuneka zijonge into ezilawula. I-endpoint eyala ngokwendidi ayikwazi ukusebenzela nanye kwezo ngokuthembekileyo.

Izilumkiso ezimbini ezinyanisekileyo. Okokuqala, i-Thinking Machines ipapasha uvavanyo lokhuseleko ngosasazo lwabo lwereferensi — ama-78.0% kwi-FORTRESS echasayo, ama-95.9% kwi-FORTRESS engenangozi, ama-98.6% kwi-StrongREJECT. Lawo manani achaza isitaki sabo nezikhuseli zabo zisendaweni. Awasichazi esethu, kwaye asizukuwacaphula ngokungathi enza njalo. Sisusile umaleko lawo manani awulinganisa ngokuyinxenye; yiyo leyo imveliso, kwaye luxanduva olo.

Okwesibini, "akukho maleko wokwala" akuthethi "akukho migaqo". I-Shannon Lab LLC isebenza phantsi kwePolisi yoSetyenziso oluNoxanduva nomthetho osebenzayo wase-US. Ukususa isahluli-ndidi kudlulisela isigwebo kuwe, akusicimi. Ukuba wakha kule nto, funda ipolisi phambi kokuba wandise.

09Ukuqalisa

I-URL esisiseko efanayo, isitshixo esinye, iilwimi ezintathu. Ukuba ikhowudi yakho sele ibiza i-OpenAI, i-Anthropic, okanye i-Responses API, ukutshintsha umtya wemodeli kudla ngokuba lolo lonke ufuduko.

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="")

Ireferensi epheleleyo yeeparamitha, imilo ye-/v1/messages neye-/v1/responses, iifomathi zeziganeko zokusasaza, nendawo yokudlala ephilayo kukwi-amaxwebhu e-API. Izikhokelo zeeModeli zezinye iimodeli ezilishumi elinanye zikwi-Shannon Research.

Amanani esiseko ngamanani apapashiweyo e-Thinking Machines Lab endaweni yokugcina ye-Inkling engacutshwanga kwaye achaza uphahla losapho, ingekuko isakhiwo sethu esicutshiweyo. Amaxabiso neeflegi zesakhono ngamaxabiso oluhlu lweSango le-Shannon AI ngokwaSeptemba 2026.

10Imibuzo ebuzwa rhoqo

---

Imithombo: 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 — amanqaku esakhiwo nobalo-mgangatho be-Inkling. Iibalo-mgangatho zesiseko zixelwe yilebhu ngendawo yokugcina engacutshwanga.