API DeepSeek Flash Kore-ārai
DeepSeek-V4-Flash-0731 · W4A16-AUTOROUND-REAP — te utu-papa, he horopaki 262K, me te kore o tētahi paparanga whakakāhore. Te tauira ka tohua e koe ki te kawenga mahi e rere ana kotahi tekau mano wā ia rā.
Whakarāpopoto
Ko DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP te taumata ranea-nui o te whānau DeepSeek i te Kūwaha Shannon AI. Ka whakahaerea i runga i tā mātou ake tautau GPU i te $0.50 ia miriona tohu tāuru me te $2.00 ia miriona tohu whakaputa — te papa o tā mātou rārangi — me tētahi horopaki 262,144 tohu, te whakaputa nui rawa 16,384 tohu, te streaming, te karanga taputapu, te whakaaroaro me te reasoning_effort, me te putanga hanganga json_schema pūmau. Kāore he paparanga whakakāhore, kāore he tātari whakaputa. Ehara ia i te putanga kua whakangoikoretia o DeepSeek-V4-Pro: he tūāpapa nui rawa a Pro kua whakawhāititia ki te 3-bit, he tūāpapa iti ake a Flash e kawea ana i tētahi quantization AutoRound 4-bit pono-teitei ake. He rua tohu rerekē ērā i runga i te kōpiko ōrite, ā, ko tēhea ka toa, kei tō kawenga mahi anake tērā e whakatau ana.
Ka whakapai te nuinga o ngā tīma e arotake ana i tētahi API LLM i te taurangi hē. Ka paearutia e rātou ngā tohutohu pakeke rawa e rua tekau ka taea e rātou te whakaaro, ka kōwhiri i te toa, kātahi, e ono wiki i muri mai, ka kite he iwa tekau mā rima ōrau o tā rātou hokohoko he mahi poto, hanganga, tukurua e taea ai e tētahi tauira e whā-tekau-ōrau te utu te whakahaere ōrite. Kei te noho a DeepSeek-V4-Flash mō taua iwa tekau mā rima ōrau. Ko tēnei whārangi e pā ana ki te mōhio he aha te wā kei roto koe — ā, he aha te wā kāo.
01He aha tonu ka riro i a koe
| Āhuatanga | DeepSeek-V4-Flash-0731 · W4A16-AUTOROUND-REAP |
|---|---|
| Id tauira | DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP |
| Horopaki / whakaputa nui rawa | 262,144 tohu / 16,384 tohu |
| Utu ia 1M (tāuru / whakaputa) | $0.50 / $2.00 |
| Streaming · karanga taputapu · whakaaroaro | Āe — ngā reo e toru katoa |
| Putanga hanganga | json_schema me te response_format |
| Tāuru whakaahua (tirohanga) | Kāo — kuputuhi anake |
| Paparanga whakakāhore / tātari whakaputa | Kāore |
| E whakahaerea ana i runga i | Tā Shannon ake tautau GPU |
Ka tae atu ngā reo API e toru ki ngā taumaha ōrite: /v1/chat/completions (āhua OpenAI), /v1/messages (āhua Anthropic), me /v1/responses. Ka mahi te streaming i ngā mea e toru. Kei te tuhinga API te tohutoro tawhā katoa.
02Nō whea mai: te whakapapa DeepSeek
Kua whakapau a DeepSeek-AI i ngā whakatupuranga tauira e rima ki tētahi hiahia kaha kotahi: te tiki i te whanonga paetata mai i tētahi tauira he utu-iti ki te whakahaere, kaua ko te utu-iti ki te whakangungu anake. Tata ki ngā whakatau hoahanga katoa o te raupapa e hoki ana ki tērā, ā, kua whakaputaina ngā mea nui. E toru e whakamārama ana i tāu e karanga ana ina pā koe ki tēnei kūwaha.
DeepSeekMoE — ngā mātanga matatini me tētahi tiritahi
Ka whakakapi tētahi paparanga Whakaranu-o-Ngā-Mātanga i tētahi whatunga whāngai-whakamua nui kotahi ki ngā mea iti ake maha me tētahi kaiwhakarite ara e kōwhiri ana i ētahi ia tohu. Ka tapatapahia e te momo a DeepSeek ngā mātanga matatini ake i te whakatakotoranga waru-mātanga māori, nō reira ka hoko ngā FLOPs hohe ōrite i tētahi whakaranu tino tohu-motuhake ake, ā, ka wehe motuhake i tētahi mātanga tiritahi e pahemo ana ia tohu, nō reira kāore te mātauranga noa e ako-anō-tia i roto i ia mātanga takitahi. I tukuna e DeepSeek-V3 tērā hei 256 mātanga arataki me te 1 mātanga tiritahi, e whakahohe ana i te 8 arataki ia tohu — 671B tawhā katoa, 37B hohe — me te whakatairite kawenga kore-utu-āwhina: kaua he tau whiu e whakakino ana i te kounga, engari he pīaha ia-mātanga e neke ana i ngā tatauranga whakarite-ara ki runga, ki raro rānei ia hipanga e ai ki te iti, ki te nui rānei o te kawenga o taua mātanga.
MLA — te aro e uru ana ki te mahara
Ko te haurua tuarua o te raru utu-whakahaere ko te keteroki KV, e tipu ōrite ana me te horopaki, ā, e noho rangatira ana ki te mahara i ngā roanga raupapa roa. Ka kōpaki te Multi-head Latent Attention, i whakaurua i te DeepSeek-V2, i ngā kī me ngā uara tahi ki tētahi ahanoa huna papa-iti, ka rongoā i taua mea, ka hanga anō i te K/V whaihua i te wā tonu. Ko ngā tatauranga V2 i whakaputaina e DeepSeek, hei whakataurite ki tō rātou DeepSeek 67B kiato o mua, e pūrongo ana i te 93.3% te heke o te keteroki KV me te 5.76× te teitei ake o te tere whakaputa nui rawa. Koinā te take he mea māori tonu tētahi matapihi 262K-tohu i tēnei whānau, kaua he tāpiritanga utu-teitei.
Aro korekore — me te tauira mō ngā tohu utu-iti
I tāpirihia e te V3.2-Exp te DeepSeek Sparse Attention: he taupū māmā e whakatau ana i ngā tohu horopaki ki te pātai, ā, ka pupuri tētahi kaikōwhiri i ngā mea tatauranga-teitei anake. Ko te wāhi whakamīharo ko te nama utu, kaua ko te tikanga. I whakaputaina e DeepSeek tētahi whakataurite taha-ki-te-taha me te V3.1-Terminus me te MMLU-Pro ōrite tonu i te 85.0, te AIME 2025 e piki ana (88.4 → 89.3), te SWE-bench Verified e heke iti ana (68.4 → 67.8) — he kounga papatairite tata tonu — ā, i tapahia ō rātou utu API ki te neke atu i te 50% i taua rā tonu. He hoahanga, i pūrongotia pono, e huri tika ana hei utu. Ko Flash taua tautohe kua tāpiri kētia.
Te whakatupuranga V4
Ko te V4, i whakaaturia tuatahitia i te Paengawhāwhā 2026 i raro i te raihana MIT, ka whakakapi anō i te tū aro ki tētahi whakaranu o te Compressed Sparse Attention me te Heavily Compressed Attention, ka whakangungu me te whakapai Muon i te neke atu i te 32T tohu, ā, ka pupuri tonu i te whāinga multi-token-prediction a V3 — e mahi tahi ana hei decoding matapae, ā, i inea e DeepSeek he 85–90% te reiti whakaaetanga tohu-tuarua mō te tata 1.8× te tere decoding i te V3. E pūrongo ana te whakarāpopototanga V4 i a V4-Pro e whakahaere ana i te 27% o ngā FLOPs whakatau tohu-kotahi a V3.2 me te 10% o tāna keteroki KV. Ka whakaputa te whānau i ngā aratau kaha-whakaaroaro e toru, koinā te mea e whakaaturia ana e tā mātou tawhā reasoning_effort.
Puna kōrero: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · DeepSeek-V3 Technical Report (arXiv 2412.19437) · DeepSeek-V3.2-Exp model card · DeepSeek-V4 (arXiv 2606.19348). He tatauranga nā DeepSeek katoa i pūrongotia.
03He aha tonu te tikanga o "Flash"
Koinei te take he hē i te nuinga o ngā whakataurite, nō reira he tika kia kī mārire: ehara a Flash a DeepSeek i tētahi Pro kua quantized, kua tapahia, kua distilled rānei. He tauira Whakaranu-o-Ngā-Mātanga iti ake, i whakangungua motuhaketia, me tāna ake rere whakangungu-tōmua. Ka homai e te pepa whakaatu tuatahi V4 ngā rahi e rua:
He tatauranga nō te pepa whakaatu tuatahi DeepSeek-V4. Ko ngā mata-arataki mana e whakahaerea ana e mātou — V4-Flash-0731 me V4-Pro-0813 — e whakarārangi ana i ngā tapeke iti nui ake, te 304B me te 1.7T, i ō rāua kāri tauira; kāore a DeepSeek e tuhi anō i ngā tawhā hohe mō aua hanga.
E rua ngā tau e mahi ana i konei. Ko ngā tawhā katoa e whakatau ana i te tapuwae mahara, e whakatau ana i te nui o ngā taputapu e nohoia ana e tētahi kape o te tauira. Ko ngā tawhā hohe e whakatau ana i te tatauranga ia tohu, e whakatau ana i te tere whakaputa. He 5.6× te iti ake o Flash mō te tuatahi, he 3.8× mō te tuarua — ā, he mea ātaahua e tika ana kia kitea i taua āputa: ka whakahohe a Flash i te 13 o ngā 284 piriona tawhā, tata ki te 4.6%, engari ka whakahohe a Pro i te 49 o ngā 1,600 piriona, tata ki te 3.1%. Ko Flash te tauira korekore-iti ake. Ehara ia i te Pro noa me ētahi mātanga iti ake kua whakakāha. He hoahoa kiato ake, whakatairite rerekē ake, e whakapau ana i tētahi wāhanga nui ake o ia tonu ia tohu ka hangaia e ia. He kōwhiringa hangarau pono tērā, kaua he whakahekenga.
He aha e kī ana ngā tau a DeepSeek ake
Ka whakaputa ngā kāri tauira mana e rua ngā tatauranga i runga i ngā paearu kaihautū me te hangarautanga pūmanawa ōrite, koinā te mea he whakataurite mā rawa tēnei — te whare rangahau ōrite, te whakatinana aromatawai ōrite, ngā rahi e rua:
| Paearu (i pūrongotia e DeepSeek) | V4-Flash-0731 | V4-Pro-0813 |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | 87.9 |
| Toolathlon-Verified | 70.3 | 74.1 |
| DeepSWE | 54.4 | 62.7 |
| NL2Repo | 54.2 | 61.5 |
| DSBench-FullStack | 68.7 | 71.1 |
Ka toa a Pro i ia rārangi. Koinā te pane pāpaho pono, ā, e pai ake ana ki a mātou kia pānui koe i konei i te kite i te whakahaere. Engari pānuitia ngā āputa, nā te mea koirā tāu e tino hoko ana. I te Terminal Bench 2.1, he 5.2 tohu te āputa i tētahi tauira me te rua-tekau-mā-rima o ngā tawhā; i te whakamahi taputapu, he 3.8 tohu. Ka whānui haere te āputa i te wāhi hōhonu rawa te mahi — 8.3 tohu i te DeepSWE, 7.3 i te NL2Repo, e rua e whakautu ana i te whakaaroaro waehere maha-hipanga mau-roa. Koirā te āhua o te whakawhitinga i te ripanga kotahi: ka kati a Flash i te nuinga o te tawhiti mō ngā mahi herea, ā, ka tuku i te whenua mō ngā mahi tuwhera-mutunga.
Ka kaha ake te kī a te kāri Flash a DeepSeek ake: ka toa a Flash-0731 i te whakaaturanga tuatahi V4-Pro i ngā paearu kua whakarārangihia, ahakoa te iti rawa ake o ōna tawhā hohe. He āhuatanga māori i tēnei mātai te toa o tētahi tauira iti hou ake i tētahi tauira nui tawhito ake, ā, koinā te take kāore rawa "taumata nui ake" e tika hei whakautu taunoa māu.
He kupu mō te horopaki
Ka whakaputa a DeepSeek i tētahi horopaki 1M-tohu mō te whānau V4. Ka whakahaerea e te Kūwaha Shannon AI tēnei tauira i te 262,144 tohu — te matapihi ōrite puta noa i ngā tauira tekau mā rua o tā mātou rārangi, ā, ko te matapihi ōrite ka riro i a koe i runga i a Pro. E pai ake ana ki a mātou kia rite te whanonga o ia tauira o te rārangi mō te horopaki i te whakamahara i a koe i tētahi tuanui rerekē mō ia id tauira.
04Te pānui i te taunga: W4A16-AUTOROUND-REAP
Ka kī mārama ō mātou id tauira i tō rātou tikanga kōpaki, nā te mea koinā te mea e tino pā ana ki te kounga whakaputa, ā, tata kāore he tangata e whakapuaki. Ka wetewetehia tēnei ki ngā kerēme e toru.
W4A16 — taumaha 4-bit, hohenga 16-bit
Ka rongoātia ngā taumaha ki te 4 pae ia mea; ka rere ngā hohenga i te tika 16-bit. He quantization-taumaha-anake tēnei, ā, e aro ana ki te tino paepae-whāiti o te whakaputa tohu, arā ko te whānuitanga rerenga mahara, kaua ko te tatauranga. Mā te quantizing i ngā hohenga hoki, ka riro he iti ake te tere, engari ka tūpono ki tētahi pari tika nui ake; ko te waiho i a rātou ki te 16-bit te kōwhiringa taritari, pono-teitei. He whakapoto ā-ao-whakahaere te tuhinga — ka tuhia te mea ōrite e te tuhinga quantization hei W4 me tētahi rahi rōpū.
AUTOROUND — te porowhita kua akona, kaua te porowhita tata rawa
Ka porowhitatia e te quantization māori ia taumaha ki te uara tata rawa e taea ana. Kua whakaatuhia kāore tērā i te mea pai rawa: mā te porowhita i tētahi taumaha ki te ara "hē" ka whakakore i te hapa i tāpiritia e ōna hoa-tata, ā, i ngā piriona taumaha, ka noho rangatira aua whakawhitiwhitinga. Ka whakaaro te AutoRound, nā Intel, i te whakatau porowhita ake hei mea hei ako — he whakaruru ia-taumaha ka taea te whakangungu, kua herea ki te ±0.5, tino nui hei huri i tētahi whakatau porowhita kotahi, kāore rawa he nui ake, me ngā awhe kuti ka taea te ako, kua whakapaingia mā te signed gradient descent ki tētahi whāinga hanga-anō-whakaputa ā-poraka i runga i tētahi huinga whakaū iti. E rua rau hipanga, kātahi ka mutu.
Kei ngā whakataurite i whakaputaina e Intel, i toa te AutoRound i te GPTQ i te 30 o ngā 32 āhuatanga i whakamātauhia, i te AWQ i te 27 o te 32, i te OmniQuant i te 29 o te 32 i ngā tautuhinga taumaha-anake 4-bit, me ngā piki i tua atu i te porowhita-ki-te-tata-rawa māori o te 6.9 ki te 33.2 tohu tika i te 2 pae. Kāore he utu o tēnei i te wā whakatau: he kōwhiringa tau pai ake, i mahia kotahi tau, motuhake i te ipurangi.
REAP — te muku i ngā mātanga kāore he tangata e arataki ana ki reira
Ko te REAP (Router-weighted Expert Activation Pruning), nā Cerebras Systems me te University of Calgary, he kōpaki hurihanga-kotahi mō ngā tauira MoE. Ka whakatau ia mātanga mā te kaha me te auau o te kōwhiritanga a te kaiwhakarite ara ki a ia, kua whakarea ki te rahi o te whakaputa a taua mātanga, kātahi ka muku tuatahi i ngā mātanga tatauranga-iti rawa — kāore he whakangungu anō i muri.
Ko te kitenga matua a te pepa te mea whaihua: i ngā ōwehenga kōpaki teitei, he pai ake te tapahi mātanga i te whakakotahi. Ka puta i te whakakotahi i ngā mātanga motuhake-mahi tētahi mea i tapaina e ngā kaituhi he horonga wāhanga-mokowā-mahi — ka ngaro i te kaiwhakarite ara te mana matatini, whakawhirinaki-tāuru i whai take ai te hanga MoE, ā, ka mahue ki te ārahi i waenga i ngā toharite pūmau. Ko ā rātou tau puta noa i ngā tauira mai i te 20B ki te 1T tawhā: i te 25% kōpaki mātanga i ngā mahi waehere, ka ngaro te REAP i te 2.8% toharite tika i te wā ka ngaro te whakakotahi i te neke atu i te 5%; i te 50%, ka ngaro te REAP i te 8.0% i te wā ka ngaro te whakakotahi i te neke atu i te 20%. I runga i te Qwen3-Coder-480B i te 50% tapahi, he 1.4% te hekenga, hei whakataurite ki te 39.3% mō te tapahi ā-auau māori. Kua whakaputaina e Cerebras ngā mata-arataki REAP mō te raupapa DeepSeek me ētahi atu.
He aha ai he 4-bit i konei, he 3-bit i runga i a Pro
Me uru, me noho tere tonu ngā taumata e rua i runga i te momo taputapu ōrite. E kawe ana a Pro i te 1.7 taera tawhā, nō reira e hiahia ana ki te maimoatanga 3-bit kaha ake kia uru ai rawa. E kawe ana a Flash i te tata 300 piriona, nō reira kāore — ā, ko te pae e penapenatia ana ka whakapaua ki te pono, kaua ki te uru.
He nui ake te take o taua pae tāpiri i tāna e whakaatu ana te tatauranga, nā te mea ehara i te ōrite te kino o te quantization. I ngā ine WikiText-2 a te pepa AWQ, ka neke a Llama-2-7B i te 5.47 perplexity i te FP16 ki te 5.60 i te 4-bit — he whiu 2.4% — engari ki te 6.24 i te 3-bit, he whiu 14%. He tata ono whakarea te heke mō te 25% kōpaki tāpiri, ā, ka mau tonu taua tauira i te 70B. He mārama ake anō te pepa GPTQ: i te 3 pae, ka horo te porowhita-ki-te-tata-rawa māori i ngā tauira taumata-175B, ki ngā mano o te perplexity, i te mea ka noho whaihua tonu i te 4 pae. I raro i te whā pae, ka mutu te noho hei mea pai noa iho te motuhēni quantization pai, ka noho hei mea kotahi anake e pupuri ana i te tauira kia kotahi.
Nō reira, ko te pānui pono o ēnei taunga e rua: e noho tata ake ana a Flash ki tōna ake mea kāore i kōpakina, i a Pro ki tōna ake. Ko Pro tonu te tauira kaha ake — kua whakatauhia tērā e te ripanga paearu i runga ake nei. I tuku noa ia i te nui ake kia tae mai ai ki konei, ā, ka utua e koe mō taua mana.
Puna kōrero: AutoRound — "Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs" (arXiv 2309.05516), Intel, Findings of EMNLP 2024. REAP — "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression" (arXiv 2510.13999), Cerebras Systems. He tatauranga perplexity nō te AWQ (arXiv 2306.00978) me te GPTQ (arXiv 2210.17323).
05Flash rāua ko Pro i te kūwaha
I kapi te wāhanga 03 i ngā tūāpapa. Anei te āhua o ngā taumata e rua hei hua i tā mātou kūwaha:
| DeepSeek-V4-Flash-0731 | DeepSeek-V4-Pro-0813 | |
|---|---|---|
| Quantization | W4A16-AUTOROUND-REAP | 3BIT-REAP |
| Te tika taumaha | 4-bit, porowhita kua akona | 3-bit |
| Tūāpapa (katoa / hohe) | ~304B / 13B | ~1.7T / 49B |
| Tāuru ia 1M tohu | $0.50 | $1.95 |
| Whakaputa ia 1M tohu | $2.00 | $3.90 |
| Horopaki / whakaputa nui rawa | 262,144 / 16,384 | 262,144 / 16,384 |
| Taputapu · putanga hanganga · streaming | Āe | Āe |
| Tirohanga | Kāo | Kāo |
| Paparanga whakakāhore | Kāore | Kāore |
Kia mātau ki te mea kāore e panoni: he ōrite tonu te horopaki, te tuanui whakaputa, te karanga taputapu, te putanga hanganga, me te kore o te paparanga whakakāhore. Ko ngā mea anake ka neke ko te tūāpapa, te quantization, me te utu. Koirā te take he panonitanga aho-kotahi noa te huri i waenga i a rāua i tō waehere, he whakatau ōhanga noa i tō hoahanga.
He aha te utu o tētahi karanga rahi-nui-rawa kotahi
Whakakīa te horopaki 262,144-tohu katoa, ka tango i te whakaputa 16,384-tohu katoa:
I te utu rārangi. Flash: $0.131 tāuru + $0.033 whakaputa. Pro: $0.511 + $0.064. Ōwehenga 3.51×. Ka tiritahi hoki a Flash i te papa $0.50 / $2.00 me MiniMax-M3, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro me Hy3 — tirohia te rārangi katoa mō te rerekētanga o aua tauira mō te tirohanga me te putanga hanganga.
06Te pāngarau ranea, kua mahia
He pono anake te "3.9× te iti ake" mō te tāuru. Mō te whakaputa, he 1.95× te iti ake a Flash. Ko tō tino penapena, nō reira, kei tō ōwehenga tāuru-ki-te-whakaputa e whakawhirinaki ana, ā, he whānui tō horapa kia panoni ai ngā whakatau hoahoa. E whā ngā kawenga mahi tūturu, ia mea kotahi rau mano tono:
| Kawenga mahi | Tāuru / whakaputa ia karanga | Flash | Pro | Penapena |
|---|---|---|---|---|
| Whakarōpū tīkiti | 1,200 / 120 | $84 | $281 | 3.3× |
| Te whakautu pātai RAG | 8,000 / 600 | $520 | $1,794 | 3.4× |
| Whakarāpopoto tuhinga-roa | 60,000 / 1,500 | $3,300 | $12,285 | 3.7× |
| Hanga raraunga waihanga | 500 / 2,000 | $425 | $878 | 2.1× |
Ko te tauira: he nui ake tō pana horopaki, he iti ake tō whakaputa kuputuhi, he nui ake te penapena a Flash i a koe. Ko ngā ara tikiatu, te tirotiro tuatahi tuhinga, te tātari rārangi-rekoata me te tuhinga tuhinga, me te tango horopaki-roa te wāhi tino pai — he taumaha-tāuru rātou mā tētahi taumata whakarea, ā, ko te tāuru te wāhi e noho ana te utu-heke 3.9×. Mēnā kē ko tō kawenga mahi he tohutohu poto e whakaputa ana i ngā mano tohu kōrero, ka whāiti te āputa ki te tata 2×, ā, ka kaha ake te tautohe āheinga mō Pro.
Kotahi ake tatauranga e tika ana kia mau ki a koe: i te $0.50 ia miriona tohu tāuru, he 13.1 heneti te utu o te whakakī i te matapihi 262,144-tohu katoa. Ka mutu te noho hei pātai kōti te tohutohu horopaki-roa i taua utu, ā, i ētahi wā ko te tikanga ka taea e koe te peka i te hanga paparanga tikiatu, ka waiho i te matapihi horopaki kia mahi — he penapena hangarau i tua atu i te penapena tohu.
07He aha te tikanga o "kore-ārai" i konei, tino tika
He kupu e whakamahia māmātia ana, nō reira anei tā mātou tautuhinga, kua kīia hei āhuatanga o te kūwaha: kāore he kaiwehewehe whakakāhore kei mua i te tauira, kāore hoki he tātari ihirangi e pā ana ki tāna whakaputa. Kāore tētahi mea e hopu ana i tētahi tohutohu, e whakahoki mai ana i tētahi whakakāhoretanga kua rite kētia; kāore tētahi mea e tirotiro ana i te whakaotinga, e whakangaro ana i a ia. Mō te mahi whakahaere ranea-nui, he iti ake te whakamīharo, he nui ake te whaihua o ngā hua i tā te kupu e kī ana:
- Kāore he tika-hē i te hokohoko māia. Ka whakakāhore auau ngā kūwaha kua āraia i ngā whakarāpopoto rongoā, ngā tuhinga haumarutanga, te whakawā ā-ture, te pakimaero, ngā rauemi whakaiti-kino me te mahi whakahaere-ihirangi ake — ko ngā momo tonu e tukatuka ana tētahi mahi pupuri i te kuputuhi mō tētahi kaupapa mataku, kaua e hanga ana i tētahi mea kino. I te kotahi rau mano tono, he kotahi mano rārangi pakaru tētahi reiti whakakāhoretanga-hē kotahi-ōrau.
- Ngā kirimana whakatau-pūmau. Tonoa he whakautu
json_schema, ka riro i a koe he JSON e rite ana ki te hanganga, kaua he parāweta whakapāha i te wāhi me noho ai he ahanoa. Kāore tō wetewete e hiahia i tētahi peka whakakāhore, kāore hoki tō arorangi whakamātau-anō e hiahia wehe i tētahi hapa kawe mai i tētahi hapa kaupapa here. - Nāu te kaupapa here. Koirā te whakawhitinga. Mēnā e hiahia ana tō hua i ētahi arai haumaru mō āu kaiwhakamahi whakamutunga, hangaia e koe i tō ake paparanga, i reira e mōhio ana koe ki āu kaiwhakamahi me tō mana whakahaere, kaua e riro i a koe ngā matapae a tētahi kaiwhakarato mō ērā e rua. Ka whakahaerea tonu tō pūkete e tā mātou Kaupapa Here Whakamahi Haepapa.
Ehara te kore-ārai i te ōrite ki te kore-whakarōpū, ki te kore-āwhina rānei. Ka whai tonu te tauira i ngā tohutohu, ka whakaute tonu i tētahi tohutohu pūnaha. Mēnā e hiahia ana koe kia whakakāhore ia i ētahi momo mō āu kaiwhakamahi, kīa i reira — nāu ake tērā tohutohu ināianei hei tuhi, kaua mā mātou e whakatau.
08Ngā āheinga me ngā rohe pono
E rua ngā āheinga e kawe ana i te nuinga o te taumaha mō te mahi ranea. Ko te putanga json_schema pūmau e huri ana i tētahi whakaotinga hei rārangi pātengi raraunga me te kore paparanga wetewete, me te kore peka whakakāhore. Ko te karanga taputapu i ngā reo e toru e whakarite ana i a Flash hei pūkaha whai-take mō ngā hipanga-iti utu-iti a tētahi kaihautū e whakahaerea ana tāna kaimahere taumata-runga i runga i tētahi tauira kaha ake. Tāpiritia te reasoning_effort, e taea ai e koe te whakapau tohu tāpiri ki te wāhanga anake o te hokohoko e hiahia ana, me tētahi horopaki 262,144-tohu e ōrite ana ki tā Pro — kāore koe e tuku i te horopaki mā te kōwhiri i te taumata utu-iti, he mea onoono tērā.
He ōrite te tino o ngā rohe:
- Kāore he tirohanga. He kuputuhi anake a Flash, ā, ka whakakāhoretia ngā wāhanga whakaahua i tētahi karere. Mō ngā whakaahua i runga i tētahi tauira kore-ārai i te papa $0.50 / $2.00 ōrite, ko MiniMax-M3 me MiMo-V2.5 ngā kōwhiringa whai-tirohanga; ko Kimi-K2.6 me Kimi-K3 ngā tauira tirohanga kaha ake i runga ake i taua taumata.
- Ehara i te pito teitei o te arawhata. Ka whakatakoto ngā tau a DeepSeek ake i a Pro ki mua i ia paearu kaihautū kua whakaputaina, me te āputa whānui rawa i te whakaaroaro waehere maha-hipanga mau-roa. I te wāhi he utu-nui ake tētahi whakautu hē kotahi i te kotahi mano utu-iti, whakapaua te $1.45 tāpiri ia miriona ki a DeepSeek-V4-Pro.
- 16,384 tohu whakaputa mō ia tono. Me wāwāhi ngā whakaputa roa puta noa i ngā karanga. He tuanui ōrite tēnei i ia tauira e whakahaerea ana e mātou, ā, ka here i mua noa i te matapihi horopaki.
09Te karanga i a ia
Tohua tētahi kiritaki rite-ki-OpenAI ki te kūwaha, ka huri i te aho tauira. Kāore tētahi atu mea i tō tū e hiahia neke.
curl https://api.shannon-ai.com/v1/chat/completions \
-H "Authorization: Bearer $SHANNON_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP",
"messages": [
{"role": "system", "content": "Label the ticket. Reply only with the schema."},
{"role": "user", "content": "The invoice PDF renders blank on Safari 18."}
],
"response_format": {
"type": "json_schema",
"json_schema": {
"name": "ticket_label",
"schema": {
"type": "object",
"properties": {
"category": {"type": "string"},
"severity": {"type": "integer"},
"needs_human": {"type": "boolean"}
},
"required": ["category", "severity", "needs_human"],
"additionalProperties": false
}
}
}
}'Ka mahi te id tauira ōrite, kāore i panonitia, i te /v1/messages me te /v1/responses, koirā te huarahi tere rawa ki te whakamātau A/B i tētahi tauira utu-iti ake i raro i tētahi whakahiatonga kei a koe kē, me te kore tuhi anō i tō kiritaki. Kua tuhia ngā tawhā, ngā āhua kaupapa streaming, ngā hanganga karanga-taputapu me ngā rohe reiti i te tohutoro API; kua whakarārangihia ia tauira o te rārangi i Shannon Research.
10Ngā pātai auau
He aha te API DeepSeek Flash kore-ārai?
Ko te kūwaha o te Kūwaha Shannon AI mō DeepSeek-V4-Flash-0731-W4A16-AUTOROUND-REAP: he hanga AutoRound 4-bit o te tūāpapa V4-Flash 13B-hohe a DeepSeek, e whakahaerea ana i runga i tā mātou ake tautau GPU kāore he paparanga whakakāhore, kāore hoki he tātari whakaputa, i te $0.50 ia miriona tohu tāuru me te $2.00 ia miriona tohu whakaputa, me tētahi horopaki 262,144 tohu.
He aha te utu o DeepSeek V4 Flash ia miriona tohu?
$0.50 tāuru me te $2.00 whakaputa, te papa utu i te kūwaha Shannon. He $1.95 me te $3.90 a DeepSeek-V4-Pro-0813, nō reira he 3.9× te iti ake o Flash i te tāuru, he 1.95× te iti ake i te whakaputa. He tata $0.13 te utu o te whakakī i te horopaki 262,144-tohu katoa i runga i a Flash, hei whakataurite ki te $0.51 i runga i a Pro.
Me whakamahi ahau i a DeepSeek V4 Flash, i a DeepSeek V4 Pro rānei?
Ko Flash ina te maha o ngā karanga te rohe: te whakarōpū, te tango, te whakarāpopoto, te raraunga waihanga, te whakautu RAG, ngā hipanga-iti kaihautū. Ko Pro ina te pakeke o tētahi karanga kotahi te rohe. Ehara a Flash i te Pro kua whakahekea; he tūāpapa iti ake i tētahi quantization 4-bit pono-teitei ake, i te mea he tūāpapa 1.7T-tawhā i te 3-bit a Pro.
Ka tautoko a DeepSeek-V4-Flash-0731 i te tirohanga me te putanga JSON schema?
Ka tautoko ia i te putanga hanganga json_schema pūmau me te response_format, tae atu ki te karanga taputapu, te streaming me te reasoning_effort. Kāore ia e whakaae ki te tāuru whakaahua. Mō ngā whakaahua i runga i tētahi tauira kore-ārai i te papa utu ōrite, ko MiniMax-M3 me MiMo-V2.5 ngā kōwhiringa whai-tirohanga.
He aha te tikanga o te W4A16-AUTOROUND-REAP?
Ko te W4A16 he taumaha 4-bit me ngā hohenga 16-bit. Ko te AutoRound te tikanga porowhita kua akona a Intel, e whakapai ana i ngā whakatau porowhita mā te signed gradient descent, kaua mā te porowhita ki te uara tata rawa. Ko te REAP te tapahi mātanga hurihanga-kotahi a Cerebras Systems mō ngā tauira Whakaranu-o-Ngā-Mātanga, e muku ana i ngā mātanga arataki-iti rawa me te kore whakangungu anō.
He kore-ārai pono te tauira, kāore he paparanga whakakāhore?
Āe. Kāore he kaiwehewehe whakakāhore e noho ana i mua i te tauira, kāore hoki he tātari e pā ana ki tāna whakaputa, nō reira ka riro i a koe te tohatoha ake o te tauira. Ko te tikanga hoki, nāu ake te tū haumaru o tō hua, ā, ka hangaia e koe ngā arai haumaru i tō ake paparanga; ka whakahaerea tonu te pūkete e tā mātou Kaupapa Here Whakamahi Haepapa.
Whakahaerea tō ranea i runga i a Flash
$0.50 tāuru, $2.00 whakaputa, horopaki 262K, kāore he whakakāhoretanga. Hurihia kotahi aho tauira.
Pānuitia ngā Tuhinga API Tirohia Ngā Tauira Tekau mā Rua KatoaShannon Lab LLC · New Mexico, USA · e whakahaerea ana e tā mātou Kaupapa Here Whakamahi Haepapa
Puna kōrero: DeepSeek-V2 (arXiv 2405.04434) · DeepSeekMoE (arXiv 2401.06066) · DeepSeek-V3 Technical Report (arXiv 2412.19437) · DeepSeek-V4 (arXiv 2606.19348) · Ngā kāri tauira DeepSeek-V4-Flash-0731 me DeepSeek-V4-Pro-0813 · AutoRound (arXiv 2309.05516, Intel) · REAP (arXiv 2510.13999, Cerebras Systems) · AWQ (arXiv 2306.00978) · GPTQ (arXiv 2210.17323). He tatauranga tūāpapa me te paearu i whakaputaina e ō rātou kaituhi ake katoa; he uara rārangi nā Shannon AI ake ngā utu me te horopaki e whakahaerea ana, e rite ana ki te Hepetema 2026.