Te API DeepSeek Kore-ārai
He Whakaranu-o-Ngā-Mātanga (MoE) DeepSeek 1.6 taera tawhā e mahi ana kāore he paparanga whakakāhore: horopaki 262K, kaha whakaaroaro taketake, karanga taputapu, me ngā reo API e toru i runga i tā mātou ake tautau GPU.
Whakarāpopoto
DeepSeek-V4-Pro-0813-3BIT-REAP ko te tauira whakaaroaro-me-waehere rangatira o te rārangi kore-ārai a Shannon AI. Ko te tūāpapa he V4-Pro tuwhera-taumaha nā DeepSeek-AI — 1.6T tawhā katoa, 49B e hohe ana ia tohu, kua raihanatia ki te MIT, kua whakangungua taketake mō te horopaki miriona-tohu. Ka whakahaerea e mātou i runga i tā mātou ake tautau GPU i ngā taumaha 3-bit me te tapahi mātanga REAP, me tētahi matapihi 262,144 tohu, te whakaputa nui rawa 16,384 tohu, te streaming, te karanga taputapu taketake, te putanga json_schema pūmau, me tētahi pae reasoning_effort. $1.95 ia 1M tāuru, $3.90 ia 1M whakaputa. Kāore he tirohanga. Kāore he paparanga whakakāhore, kāore he tātari whakaputa.
Ko te hunga e tae mai ana ki te kīanga "API DeepSeek kore-ārai" e whai ana i tētahi mea kotahi tonu: he tauira e whai ana i te kaha whakaaroaro me te waehere a DeepSeek, e whakautu ana i te pātai i tino pātaihia e rātou. Kua tuhia pai te wāhanga āheinga — kua whakaputaina e DeepSeek-AI te hoahoa me ngā tau, ā, kei Hugging Face ngā taumaha i raro i te MIT. Ko te wāhanga tuarua he whakataunga whakahaere, ā, ko te tikanga ka whakatauria mōu. Ka kapi tēnei whārangi i ngā mea e rua: he aha te tūāpapa V4-Pro, he aha te mahi a tā mātou hanganga 3BIT-REAP ki a ia, me te tino tikanga o "kore-ārai" ina kōrerohia e mātou.
01He aha kē te tikanga o te "API DeepSeek kore-ārai"
E rua ngā tikanga rerekē kei raro i te kupu kotahi. Ko te tuatahi ko te whakarōpū kua tō ki roto i ngā taumaha — te whanonga whakakāhore i akona i te wā whakangungu-whai-muri. Ko te tuarua he paparanga whakahaere-ihirangi kei te taha tuku: he kaiwehewehe motuhake e tirotiro ana i tō tohutohu i mua i te kite a te tauira, me te whakaputa i mua i tō kite, ā, ka taea e ia te tapahi i tētahi whakautu i waenganui i te rerenga ahakoa te mea i hiahia te tauira ki te kī.
Ka taea te ine i ngā mea e rua, ā, kua inea te ratonga ā-kamupene tonu a DeepSeek. I whakaemi a Promptfoo i tētahi huinga 1,360 tohutohu e pā ana ki ngā pātai tōrangapū tino mataku, ā, i pūrongo mai i whakakāhoretia, i karo rānei e te R1 whakahaere a DeepSeek 1,156 o aua tohutohu — tata ki te 85%. I kitea e te mahi mātauranga i runga i taua whānau tauira tētahi momo rahunga motuhake i tapaina e ngā kaituhi he pēhi-whakaaro, i reira ka mutu tata tonu te aho whakaaroaro a te tauira ina tīmata ki tētahi kaupapa mataku. I tuhia e tētahi atu arotake i roto i te Information Sciences he pēhi taumata-tikanga: he ihirangi kei roto i te whakaaroaro ā-roto o te tauira ka waiho puku, ka tuhia anō rānei i mua i tōna taenga ki te kaiwhakamahi. He raru pakeke ake tēnei i te whakakāhore whānui, nā te mea he āhua whakautu tonu tētahi whakautu kua tapahia.
Ko tā mātou e whakahaere ana ko te putanga tuwhera-taumaha, i runga i ngā taputapu e whakahaerea ana e mātou, i te United States, i raro i a Shannon Lab LLC. Kāore he paparanga whakakāhore kei mua i te tauira, kāore hoki he tātari e pā ana ki te mea ka hoki mai: kāore tō tohutohu e kauteahia ki tētahi kaupapa here i mua i te whakahaere, kāore hoki te awa whakaputa e tirotirohia i tōna putanga. Nā te mea kua raihanatia te tūāpapa ki te MIT ā he mea tūmatanui, ka taea e koe te whakataurite i ā mātou whakaputa ki ngā taumaha tohutoro mōu ake, kaua ki te whakapono noa ki tā mātou kōrero.
02Te tūāpapa: DeepSeek-V4-Pro, he MoE 1.6T
Ko DeepSeek-V4 te whānau tuwhera-taumaha o nāianei a DeepSeek-AI, i whakaaturia tuatahitia i te Paengawhāwhā 2026, ā, i tukuna i raro i te MIT i ngā rahi e rua: V4-Flash (284B katoa, 13B hohe) me V4-Pro (1.6T katoa, 49B hohe). Ko tēnei tuhinga e pā ana ki te taumata Pro. I whakangungua tuatahitia ngā mea e rua ki te tata 33 taera tohu, ā, ka tautoko taketake ngā mea e rua i te horopaki kotahi-miriona-tohu.
Ko te ōwehenga korekore te mea kia āta tirohia: e toru ōrau anake o ngā tawhā ka hohe mō tētahi tohu, ā koinā te mea e taea ai te whakahaere i tētahi tauira pēnei te rahi. Ko te poraka MoE ake he DeepSeekMoE tonu — ngā mātanga arataki matatini me ngā mātanga tiritahi mau-tonu i whakaurua i te V2 — me ngā panonitanga iti noa i te V4, ko te mea kitea rawatia ko te whakarite ara hash i ngā paparanga tuatahi. Ko te Multi-Token Prediction, te whāinga āwhina o V3 e whakahohe ana hoki i te decoding matapae, e mau tonu ana kāore i panonitia.
Ko ngā wāhanga hou o te V4 he ā-hanganga. Ka whakakapi ngā Manifold-Constrained Hyper-Connections (mHC) i te ara toenga māmā ki tētahi matriki whakaranu kua akona, kua here ki te manifold o ngā matriki tauwhitiwhiti-rua mā tētahi whakaputa Sinkhorn–Knopp — he awa toenga e taumaha anō ana ia tāuru me te kore o te kōrari-tau e patu ana i ngā tū hōhonu rawa. I neke te whakangungu i te AdamW ki te Muon mō te nuinga o ngā waehanga, mā te whakamaru orthogonal Newton–Schulz, me te pupuri i te AdamW mō ngā tāuru whakauru, ngā pane matapae, te whakarite paerewa me ngā tawhā mHC pūmau anake. Ā, ka whakangungua ngā taumaha mātanga ki te FP4, me te nuinga o ērā atu tawhā ki te FP8 — he mōhio ki te quantization mai i te tīmatanga, he mea nui mō te āhua ātaahua o te tauira ki te whakawhāiti anō.
03Mai i te MLA ki te CSA + HCA: te āhua o te pupuri iti-utu a DeepSeek i te horopaki roa
Ehara te rahi te aho here o DeepSeek puta noa i ngā whakatupuranga e whā, engari ko te ōhanga aro. Kua whakaekea e ia putanga te keteroki KV, e tipu ōrite ana me te horopaki, ā, e noho rangatira ana ki te mahara i ngā roanga roa.
- I whakaurua e te V2 te Multi-head Latent Attention (MLA) — he kōpōtanga papa-iti tahi o ngā kī me ngā uara ki tētahi ahanoa huna, nō reira ka pupuri te whakatupu i taua ahanoa me tētahi kī RoPE motuhake, kaua i ngā K me V pane-katoa. I pūrongo a DeepSeek i te heke o te keteroki KV ki te tata 4–14% o te aro pane-maha paerewa, me te kounga i pupuri, i pai ake rānei.
- Ko te V3 i whakanui i tēnā ki te 671B, ā, ka noho te MLA hei whakaaro taunoa mō te horopaki roa iti-utu i ngā tauira tuwhera.
- I tāpirihia e te V3.2 te DeepSeek Sparse Attention (DSA), e kōwhiri ana i tētahi wehenga top-k o te hītori mō ia pātai mā tētahi taupū māmā, kaua e aro ana ki ngā mea katoa.
- Ka whakaranu te V4. Ko te Compressed Sparse Attention (CSA) ka kōpaki tuatahi i ngā urunga KV o ia m tohu ki te urunga kotahi mā ngā taumaha kua akona, kātahi ka whakahaere i te kōwhiri top-k tapatapa i runga i aua urunga kua kōpakina. Ko te Heavily Compressed Attention (HCA) ka kōpaki kaha ake, engari ka noho pūkeri tonu ki runga i te hua. Ka whakawhitiwhitia ngā momo paparanga e rua puta noa i te tū, nō reira ka noho tahi te whakarāpopototanga whānui iti-utu me te tikiatu ā-rohe tika.
He nui te painga i pūrongotia i te horopaki kotahi-miriona-tohu: e hiahia ana te V4-Pro ki te tata 27% anake o ngā FLOPs whakatau ia-tohu, me te 10% o te keteroki KV a te V3.2. Koinā te take ka rite tonu tētahi tono 262K ki tētahi tono noa, kaua ki tētahi kaupapa mahara nui, ā, koinā hoki te take ka mau tonu ngā tau tikiatu horopaki-roa, kaua e horo i muri i ētahi tekau mano tohu.
04He aha te tikanga o te taunga 3BIT-REAP
Ka mutu tā mātou tautuhinga rārangi ki te 3BIT-REAP, ā, e whakaahua ana tēnā i ngā hipanga kōpaki motuhake e rua i whakamahia i te raupapa.
REAP: te tapahi i ngā mātanga, kaua te whakakotahi
Ko te REAP te Router-weighted Expert Activation Pruning, nā Cerebras Research — "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression", i whakaaetia mō ICLR 2026. I roto i tētahi MoE korekore, kāore te kaiwhakarite ara e whakamahi ōrite i ōna mātanga: he hiku roa ka kōwhiria onoono, ngoikore rānei, e rua rānei. Ka whakatau te REAP i ia mātanga mā te whakakotahi i ngā uara kūaha kaiwhakarite ara (te auau me te kaha o te kōwhiritia) me ngā paerewa hohenga mātanga (te nui o tāna whakaputa e whai wāhi ana), kātahi ka muku tuatahi i ngā mea hinga.
Ko te tautohe a te pepa he pai ake te tapahi i te whakakotahi: mā te whakakotahi ka toharite atu te motuhaketanga mahi i whai take ai te hanga MoE, engari mā te muku i ngā mātanga kāore e tino tae atu te kaiwhakarite ara ka pupurihia taua motuhaketanga. Puta noa i ngā tauira mai i te 20B ki te 1T, i pai ake te REAP i te whakakotahi me ngā tikanga tapahi hoariri i ngā paearu whakaputa, ā, e pūrongo ana ngā kaituhi i ngā hua tata-kore-ngaro i te whakaputa waehere me te karanga taputapu i te 50% tapahi mātanga i runga i te Qwen3-Coder-480B me te Kimi-K2 — koirā tonu te momo kawenga mahi e noho nei tēnei tauira.
Ngā taumaha 3-bit: te taumata kōpaki mō ngā tūāpapa nunui rawa
Ka quantized ngā mea e ora tonu ana ki ngā taumaha 3-bit. He kaha te toru pae, ā, kāore mātou e whakapae kē: puta noa i te rārangi ka whakamahia e mātou te W8A16 mō ngā tūāpapa iti noa kāore e hiahia te kōpaki, te W4A16-AUTOROUND-REAP i waenganui, me te 3BIT-REAP mō ngā tūāpapa nunui rawa anake, i reira ko te rerekētanga i waenga i te whakahaere i te tauira me te kore whakahaere.
He tika kia kī mārama te utu whakawhiti: ehara tētahi MoE 1.6T 3-bit i te ōrite-pū ki te putanga tohutoro, ā he pono tētahi utu kounga iti. Ko tāu ka riro mai he tūāpapa, ki te kore e kōpakina, he tino teitei ake tōna tuanui i tētahi mea e noho pai ana i te tika-teitei ake — ā, nā te mea i whakangungua ngā mātanga o te V4 ki te FP4 mai i te tīmatanga, he pai rawa te whanonga o te tauira i raro i te kōpaki taumaha anō. Mēnā kei te hiahia koe i tētahi hanga DeepSeek he māmā ake te pā ki te tika, ka whakamahi te DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP i ngā taumaha 4-bit me ngā hohenga 16-bit me te porowhita kua akona a AutoRound, i runga i te horopaki 262K ōrite mā te $0.50 tāuru / $2.00 whakaputa.
05Te whakaaroaro, me te pae reasoning_effort
Ka whakangungua motuhaketia e te whakangungu-whai-muri o te V4 ngā mātanga rohe — pāngarau, waehere, kaihautū, te whai tohutohu — mā te SFT ka whai i te GRPO, te tikanga Group Relative Policy Optimization i whakaurua e DeepSeek i te R1. Kātahi ka whakakotahitia aua mātanga ki roto i te tauira kotahi mā te distillation ā-kaupapa-here i raro i tētahi whāinga reverse-KL, nō reira ka kawe tētahi mata-arataki kotahi i te whanonga taumata-mātanga i ngā rohe maha i te wā kotahi, kaua hei tauira whānui e ngoikore ana i ngā mea katoa.
Kei te tauira i whakaputaina ngā aratau whakaaro kua whakapaipai, kaua he pātono e-rua-noa. I tā mātou kūwaha ko te tawhā paerewa reasoning_effort tērā, koirā te pane whaihua rawa o tēnei tauira:
- Iti — te whakautu tata-tika: te whakarōpū, te tango, te tuhi anō, ngā mahi katoa e utu ana koe mō te mātauranga o te tauira, kaua mō tāna whakaaroaro.
- Waenga — te aratau mahi taunoa mō te arotake waehere, te whakatika hapa me te tātari maha-hipanga.
- Nui — te whakaaroaro roa i mua i te whakautu. Ko reira e whakaotia ai ngā raru pakeke, ā, ko reira e noho ana tō nama tohu-whakaputa, nā te mea ka utua ngā tohu whakaaroaro hei whakaputa.
Nā te kore o tētahi paparanga whakakāhore, kāore te aho whakaaroaro e tapahia i ngā kaupapa mataku — kāore he tikanga e puta ai te tauira pēhi-whakaaro i tuhia mō te ratonga ā-kamupene tonu, i konei.
06Te waehere me te mahi kaihautū: ngā tatauranga a te tūāpapa
He tau nā DeepSeek-AI ake i whakaputaina mō te putanga V4-Pro ēnei — ehara i a mātou, kāore hoki i inea i runga i tā mātou hanga quantized. Ka whakaaturia e mātou nā te mea koirā te take pono e kōwhiria ai tēnei tūāpapa, ā, ka tohua hei tau ā-kamupene tonu kia taea ai e koe te taumaha i runga i tērā.
E rua ngā mea e tika ana kia kōrerohia. Ko te SWE-bench Verified i te 80.6% te whakaoti take putunga-pū pono — te pānui i tētahi putunga waehere tauhou, te kimi i te hapa, te hanga i tētahi patati e paahitia ana e ngā whakamātautau ake a te kaupapa. Ko te MRCR i te kotahi miriona tohu (83.5) te tikiatu coreference maha-hurihanga i te roa katoa, koirā te tau e whakaatu ana kāore i te whakapaipai noa te kerēme horopaki-roa. E pūrongo ana hoki a DeepSeek i tētahi rāngi Codeforces o te 3206 mō te tauira, me te 62.0 i te CorpusQA i te horopaki 1M.
He tau i pūrongotia e DeepSeek-AI mō te putanga tohutoro DeepSeek-V4-Pro ngā mea kei runga ake nei. He momo kua kōpakina tā mātou hanga 3BIT-REAP o aua taumaha, ā, kāore anō i paearutia motuhaketia mō ēnei mahi; whakaarohia ngā tatauranga tohutoro hei tuanui mō te tūāpapa, kaua hei ine mō tēnei kūwaha.
07Horopaki 262K: he aha tonu ka uru
Kua whakangungua te tauira tohutoro mō te miriona tohu. Ka whakahaerea e mātou te 262,144 — te matapihi e tiritahi ana ngā tauira katoa o te kūwaha, nō reira he panonitanga rārangi-kotahi te huri tauira, kāore rawa he heke horopaki. I te tikanga, kua hipa te 262K i te wā ka here te horopaki i te nuinga o ngā mahi tūturu:
| Kawenga mahi | Tata tonu ngā tohu | Ka uru ki te 262K? |
|---|---|---|
| Putunga ratonga rahi-waenga (~50k rārangi) | ~600K | Kāo — kōwhiria ētahi kōwae |
| Pūnaha-iti kotahi me āna whakamātautau | ~60–120K | Āe, mārie |
| Huinga RFC katoa mō tētahi tikanga | ~150K | Āe |
| He pukapuka hangarau 700 whārangi | ~250K | Āe, tata ki te rohe |
| Rere kaihautū roa, 40+ ngā hurihanga taputapu | ~80–200K | Āe |
Ko te rohe hei whakamahere ko te whakaputa: 16,384 tohu te tino nui mō ia whakautu. He tuhinga roa, he patati maha-kōnae nui rānei tērā, ehara ia i te putunga waehere katoa. Ka mahia ngā mahi whakaputa nunui mā te here i ngā tono, mā te whakahoki i te whakaputa o mua ki roto i te horopaki — he mea māmā nā te matapihi 262K — kaua mā te tono i te whakaoti kotahi tino nui.
08Te utu, me te tūnga o V4-Pro i te rārangi
I te $1.95 tāuru / $3.90 whakaputa ia miriona tohu, ko te V4-Pro te taumata whakaaroaro rangatira — te tauira ka toro atu koe ina pakeke te raru. Ina nui noa iho, ka taea e ngā tauira utu-iti ake i runga i te matapihi 262K ōrite te mahi mō te hauwhā o te utu.
| Tauira kore-ārai | Tāuru / 1M | Whakaputa / 1M | Tirohanga | Te pai rawa mō |
|---|---|---|---|---|
| DeepSeek-V4-Pro · 3BIT-REAP | $1.95 | $3.90 | Kāo | Whakaaroaro pakeke, waehere, kaihautū |
| DeepSeek-V4-Flash · W4A16 | $0.50 | $2.00 | Kāo | Whānau ōrite, nui te kawenga |
| GLM-5.2 · 3BIT-REAP | $0.73 | $2.34 | Kāo | Kaimahi whānui-take |
| Kimi-K3 · 3BIT-REAP | $3.83 | $19.12 | Āe | Kaihautū reo-maha taumata-teitei |
| MiniMax-M3 · 3BIT-REAP | $0.50 | $2.00 | Āe | Te tirohanga utu-iti rawa |
Kia tika mō te whakataurite ka tino mahia e te tangata: he utu-iti ake te API ake a DeepSeek i tēnei — ka whakarārangihia te V4-Pro ā-kamupene tonu mō te tata $0.435 tāuru / $0.87 whakaputa, ā he iti ake anō tā rātou utu tāuru mō ngā pā-kāhi. Mēnā ko te utu ia tohu anake tāu paearu, ā kāore te whanonga whakahaere-ihirangi e pā ana ki tāu mahi, whakamahia te API ā-kamupene tonu; e pai ake ana ki a mātou te kī atu i tēnā i te waiho i a koe kia kite i muri. Ko te utu teitei ake e hoko ana i tētahi huinga āhuatanga rerekē: kāore he paparanga whakakāhore, kāore he tātari whakaputa, he kaiwhakahaere US i raro i a Shannon Lab LLC, he mata-arataki kua whakaūngia ki te rā (0813) nō reira kāore rawa tētahi whakahoutanga puku o runga ake e panoni i āu whakaputa i waenganui i tō kaupapa, e toru ngā reo API mō te tauira kotahi, me te matapihi 262K me te mata tawhā ōrite puta noa i ngā tauira tekau mā rua o te rārangi.
09Te karanga: e toru ngā reo, kotahi te tauira
Ka whakautu ia tauira o te kūwaha i ngā kūwaha e toru, me te streaming i ia mea. Kāore koe e uru ki tētahi SDK hou — ka panoni koe i tētahi URL tūāpapa me tētahi id tauira.
curl https://api.shannon-ai.com/v1/chat/completions \
-H "Authorization: Bearer $SHANNON_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "DeepSeek-V4-Pro-0813-3BIT-REAP",
"reasoning_effort": "high",
"stream": true,
"messages": [
{"role": "system", "content": "You are a systems programmer. Be direct."},
{"role": "user", "content": "Audit this parser for memory-safety bugs."}
]
}'/v1/chat/completions— te reo OpenAI; ka mahi tahi me ngā SDK OpenAI me ngā mea kua hangaia i runga i a rātou./v1/messages— te reo Anthropic, mō ngā putunga waehere kua tuhia kē ki taua hanga./v1/responses— te ataata momo-Responses.
E tautokona ana i tēnei tauira: te streaming, te karanga taputapu taketake, te putanga hanganga json_schema pūmau, te response_format, me te reasoning_effort. Kei te tuhinga API te tohutoro tawhā katoa me tētahi papa tākaro ora.
10Ngā rohe pono
E whā ngā mea hei mōhio i mua i tō hanga ki runga i a ia:
- Kāore he tirohanga. He kuputuhi anake; kāore e whakaae ki te tāuru whakaahua i tētahi kūwaha. Mēnā kei te hiahia koe i tētahi tauira kore-ārai e kite ana, whakamahia a Kimi-K3, Kimi-K2.6, MiniMax-M3, MiMo-V2.5, inkling rānei.
- He tuanui whakaputa 16,384 tohu. Ka tatauria ngā tohu whakaaroaro ki tēnā i te kaha nui, nō reira he iti iho te wāhi e toe ana mō te whakautu i muri i te whakaaroaro roa. Herea ngā tono mō ngā mahi whakaputa nunui.
- He kōpaki pono te 3-bit. Ko ngā tatauranga tohutoro e whakaahua ana i te tūāpapa, kaua i tēnei hanga. Mō te taumata pono-rawa, ko ngā tauira
W8A16(MiMo-V2.5-Pro, MiMo-V2.5, Hy3) e kawe ana i te quantization iti rawa. - Ehara te kore-ārai i te kore hua. Ko te kore o tētahi paparanga whakakāhore te tikanga ka noho te haepapa mō tāu e pātai ai, me tāu e mahi ai ki te whakautu, ki a koe — i raro i ā mātou tikanga whakamahi me te ture o tō wāhi e noho ana.
Ko ia kerēme hoahanga me ia tatauranga paearu o tēnei whārangi i ahu mai i tētahi puna tuatahi tūmatanui. Pānuitia koa e koe ake.
11Ngā pātai auau
He aha te API DeepSeek kore-ārai?
He tauira DeepSeek-V4-Pro tuwhera-taumaha e whakahaerea ana i runga i tā Shannon AI ake tautau GPU, kāore he paparanga whakakāhore kei mua i a ia, kāore hoki he tātari e pā ana ki tāna e whakahoki mai ai. Ka karangahia e koe i te /v1/chat/completions, /v1/messages, /v1/responses rānei me tētahi matapihi horopaki 262,144 tohu, ā, ka whakautu te kūwaha kaua ko te whakahaere-ihirangi.
He aha te rerekētanga o tēnei i te API mana a DeepSeek?
He whakapapa tuwhera-taumaha ōrite, engari he kaiwhakahaere rerekē, he kaupapa here rerekē. He utu-iti ake ia tohu te ratonga ā-kamupene tonu a DeepSeek, ā, ka whakahaerea e ia tāna ake whakahaere-ihirangi; kua tuhia e ngā arotake i whakaputaina e Promptfoo me ngā rōpū mātauranga te whakakāhore me te pēhi taumata-tikanga i ngā tohutohu tōrangapū mataku i reira. Ka whakahaerea e Shannon AI ngā taumaha i te United States, kāore he paparanga whakakāhore, ka whakaatu i te kaha whakaaroaro me te karanga taputapu i ngā reo API e toru, ā, ka whakaūngia tētahi mata-arataki kua whakaritea ki te rā kia kore ai āu whakaputa e neke i raro i a koe.
He aha te tikanga o te 3BIT-REAP?
E rua ngā hipanga kōpaki. Ko te REAP te Router-weighted Expert Activation Pruning, he tikanga nā Cerebras Research e muku ana i ngā mātanga onoono, ngoikore rānei te whakarite-ara mai i tētahi tauira Whakaranu-o-Ngā-Mātanga, mā te whakatau i ia mātanga mā ōna uara kūaha kaiwhakarite ara me ōna paerewa hohenga. Ka quantized ngā mea e ora tonu ana ki ngā taumaha 3-bit. Mā ēnei e rua ka uru tētahi tūāpapa 1.6-taera-tawhā, ā, ka noho tere tonu i runga i tā mātou ake tautau GPU.
Kei te pēhea te rahi o te matapihi horopaki me te whakaputa nui rawa?
262,144 tohu horopaki, ā, tae atu ki te 16,384 tohu whakaputa mō ia tono. Tata ki te 600–800 whārangi kuputuhi tāuru tērā, ā, he tuhinga hangarau roa, he patati maha-kōnae nui rānei te whakaputa. Ka mahia ngā mahi roa mā te here i ngā tono, kaua mā te whakapiki i te tuanui whakaputa.
Ka tautoko a DeepSeek V4 Pro i te karanga taputapu me te putanga JSON schema?
Āe ki ngā mea e rua. E tautokona ana te karanga taputapu taketake, te putanga hanganga json_schema pūmau, me te response_format, tae atu ki te streaming me te tawhā reasoning_effort. Kāore e whakaae ki te tāuru whakaahua; mēnā kei te hiahia koe i te tirohanga i runga i tētahi tauira kore-ārai, whakamahia kē a Kimi-K3, Kimi-K2.6, MiniMax-M3, MiMo-V2.5, inkling rānei.
He aha te utu o te API DeepSeek kore-ārai?
$1.95 ia miriona tohu tāuru, $3.90 ia miriona tohu whakaputa. I roto i te rārangi kore-ārai ōrite, he $0.50 tāuru me te $2.00 whakaputa te utu o DeepSeek-V4-Flash mō te horopaki 262K ōrite, ā, he $0.73 tāuru me te $2.34 whakaputa te utu o GLM-5.2, nō reira ko te V4-Pro te taumata ka toro atu koe ina pakeke te raru, kaua ina nui noa iho.
Karangahia a DeepSeek V4 Pro kore-ārai
Horopaki 262K, kaha whakaaroaro, karanga taputapu. Ngā reo OpenAI, Anthropic me Responses i runga i te kī kotahi.
Pānuitia ngā Tuhinga API Ngā Kāri Tauira KatoaShannon Lab LLC · New Mexico, USA · ka whakahaerea ngā tauira i runga i tā mātou ake tautau GPU
Puna kōrero: DeepSeek-AI — Kāri tauira DeepSeek-V4-Pro · Pūrongo hangarau DeepSeek-V4 (CSA/HCA, mHC, Muon) · DeepSeek-V2 (MLA, DeepSeekMoE) · Pūrongo hangarau DeepSeek-V3 · REAP the Experts (Cerebras Research) · Arotake āraitanga a Promptfoo · Discovering Forbidden Topics in Language Models. He tatauranga i pūrongotia e DeepSeek-AI mō te putanga tohutoro ngā tau paearu; kei te tatari tonu te whakaū motuhake e te tuatoru mō te hanga kua kōpakina. He utu rārangi a Shannon AI ēnei utu, e rite ana ki te 5 o Hepetema 2026.