API MiMo Kore-ārai i te Pono 8-Pī
Te tūāpapa MiMo whakaaroaro-tuatahi a Xiaomi, e whakahaerea ana me te kore paparanga whakakāhore i te W8A16 — tā mātou taumata quantization pono-rawa — mō te utu papa o te rārangi.
Whakarāpopoto
Ko MiMo-V2.5-Pro · W8A16 te MiMo-V2.5-Pro taumaha-tuwhera a Xiaomi e whakahaerea ana i runga i tā Shannon ake tautau GPU i ngā taumaha 8-pī / hohenga 16-pī — te taumata kōpaki-iti rawa o tā mātou rārangi tekau mā rua. He horopaki 262,144-tohu, 16,384 tohu whakaputa ia tono, te karanga taputapu, te whakaaroaro, te putanga hanganga JSON-schema, te streaming i ngā reo API e toru, me te kore paparanga whakakāhore, kore hoki tātari whakaputa, i te $0.50 tāuru / $2.00 whakaputa ia miriona tohu — te papa o te rārangi. He kuputuhi-anake; whakamahia a MiMo-V2.5 · W8A16 mō ngā whakaahua.
Ko te nuinga o ngā rārangi e hoko ana i a koe ki te tatauranga tawhā. He nama hē tēnā hei hokohoko, nā te mea kua kōpakina ia tauira taumaha-tuwhera nui ka taea e koe te utu kia uru ki tētahi wāhi. Ehara te pātai pono i te "e hia te rahi o tēnei i mua i tana parea?" engari he "e hia tonu o tēnei kei konei tonu?" Mō tēnei pātai tēnei tuhinga, ā, mō te tauira o tā mātou rārangi e mea ana te whakautu he tata katoa tonu.
01He aha tino ka riro mai i a koe i te API MiMo kore-ārai
Ko te API MiMo kore-ārai he id tauira kotahi, MiMo-V2.5-Pro-W8A16, ka taea te tae atu mā ngā reo e toru e pā ana ki ngā taumaha ōrite i runga i tā mātou ake tautau GPU: /v1/chat/completions, /v1/messages me /v1/responses. Ka streaming ngā mea e toru, ā, ka tautoko i te karanga taputapu me te whakaaroaro.
He tikanga kuiti tō kore-ārai i konei: kore paparanga whakakāhore i mua i te tauira, kore tātari e pā ana ki tana whakaputa. Kāore mātou e ārai, e whakarōpū, e tuhi anō rānei. Ka riro mai i a koe te whanonga ake o te tūāpapa — mō tētahi tauira taumaha-tuwhera taumata-rangahau, ko te tikanga ka whai wāhi ia ki te rangahau haumarutanga, te tuhinga-parau pakeke, tūkino rānei, te whakaaroaro rongoā me te ture, me ngā tono e whakakāhoretia aunoatia ana e ngā pito mutunga arumoni kua whakahaerehia. He taputapu ngaio i raro i tā mātou Kaupapa Here Whakamahi Haepapa, ehara i te korenga o ngā ture.
02Te whakapapa MiMo: te tāpae a Xiaomi mō te whakangungu-tōmua whakaaroaro-tuatahi
Ehara a MiMo i te uri o Llama me tētahi ingoa hōu; he whakapapa motuhake nō te tīma MiMo a Xiaomi, ā, he nui tōna whakaaro whakaputa i mua i tō arotake i te tauira. Ka tīmata ki a MiMo-7B (30 o Āperira 2025, raihana MIT), e tohe ana tāna pūrongo hangarau he nuinga o ngā whare rangahau e whakaaro ana i te whakaaroaro he raru whakangungu-whai-muri hei whakapiri mā te RL, ā, he hē tēnei whakaaro. I whakaritea kē e Xiaomi te whakangungu-tōmua ki tēnei: he whakaranu e toru wāhanga puta noa i te 25 taera tohu, ko te wāhanga tuarua e tata 70% pāngarau me te waehere, ko te wāhanga tuatoru e whakauru ana i ngā ara whakaaroaro waihanga i te wā e toro atu ana te horopaki mai i te 8K ki te 32K — me tētahi whāinga Multi-Token Prediction e noho ana anō hei pane wetewete-whakapae.
Ko te hua he tauira taketake 7B e whanonga ana pēnei i tētahi mea nui rawa ake. I eke a MiMo-7B-Base ki te 32.9 i te AIME 2024, ā, ko Qwen2.5-7B i te 10.1, ko Gemma-2-9B i te 0.0, ko Llama-3.1-8B i te 0.3. Kātahi ka rere te whakangungu-whai-muri i te RL puta noa i te 130,000 rapanga pāngarau me waehere ka taea te whakaū, me tētahi utu waehere e ārahina ana e te uauatanga whakamātautau, e karo ana i te mahanga utu-onge e whakangāwari ana i te RL waehere. I tae te hua ki mua i a o1-mini:
AIME 2025, ngā tau i pūrongotia e Xiaomi mai i te pūrongo hangarau MiMo (arXiv:2505.07608). I pūrongo hoki a MiMo-7B-RL i te 95.8 i te MATH-500 me te 57.8 i te LiveCodeBench v5.
I whānui ake te whānau kaua i kōtiti: ka whakanohoia e MiMo-VL-7B (Hune 2025) tētahi kaiwhakawaehere Qwen2.5-ViT ki te tuara ōrite, ā, i te Tīhema 2025 ka puta a MiMo-Audio-7B me te urunga MoE tuatahi, MiMo-V2-Flash (309B katoa / 15B hohe, 27T tohu, tika-ranu FP8). I whai mai a MiMo-V2-Pro i te Poutū-te-rangi 2026 i te tata 1T / 42B hohe me te horopaki 1M, whai muri i ngā wiki i runga i a OpenRouter hei "Hunter Alpha". I tukuna ngātahitia a MiMo-V2.5 me MiMo-V2.5-Pro i te 22 o Āperira 2026, e rua nei he MIT.
03I roto i a MiMo-V2.5-Pro: te āhua o te tūāpapa
E ai ki te kāri tauira i whakaputaina e Xiaomi, he tauira Whakaranu-o-Ngā-Mātanga korekore a MiMo-V2.5-Pro me te 1.02 taera tawhā katoa me te 42 piriona hohe ia tohu puta noa i te 70 paparanga — kotahi kiato, 69 MoE — me te 384 mātanga kua whakaritea-ara, 8 ia tohu. Ko te rahi huna he 6144, me te rahi waenga MoE kuiti o te 2048.
E rua ngā kōwhiringa e whakatakoto ana i tōna āhua i te mahi. He whakaranu te paparua aro: 10 paparanga aro-katoa me te 60 paparanga aro matapihi-neke i tētahi matapihi 128, me tētahi whakapēhitanga kōpū-aro ka akona. E pūrongo ana a Xiaomi ka tapahia e tēnei te rokoranga KV-cache tata ki te 7x i te aro-whānui-katoa, me te pupuri tonu i te kounga horopaki-roa — te rerekētanga i waenga i tētahi matapihi he roa ingoa noa me tētahi ka taea e koe te utu kia whakakīa. Ka mau hoki e ia e toru paparanga Multi-Token Prediction, e hoatu ana ki te wetereo he pane whakapae.
I rere te whakangungu ki te 27 taera tohu i te roa raupapa ā-taketake 32K i te tika-ranu FP8 (E4M3), kātahi ka whai i te whakatikatika-motuhake, te RL kaihautū rahi-nui, me te distillation maha-kaiako aro-kaupapa. Kia mau ki taua taipitopito FP8; koinā te tautohenga kaha rawa kotahi mō te whakahaere i tēnei tūāpapa i te 8 pī.
E pūrongo ana a Xiaomi i te tauira ki te 57.2% i te SWE-bench Pro, te 72.9 i te τ3-Bench mō te whakamahi taputapu, me te 48.0% i te Humanity's Last Exam — he tau nā te kaihoko anō mō te tauira kāore i kōpakina, kua whakahuahia pērā.
Ka whakahaerea e mātou he matapihi 262,144-tohu, kaua te 1M katoa o te tūāpapa — he horopaki ka taea e mātou te whakaū i te tere, kaua he nama matua e heke ana te kounga i raro i te taumahatanga.
04He aha tino te tikanga o W8A16
Ka mau ia tauira o tēnei rārangi i tētahi taunga quantization e whakaahua ana i tētahi whakawhitinga tūturu, ka taea te ine, kaua he taumata hokohoko noa. Ko ā mātou ko 3BIT-REAP, W4A16-AUTOROUND-REAP me W8A16. Ko MiMo-V2.5-Pro te tuatoru.
Ka wehea te tuhinga ki te rua. Ko W8 te tika taumaha: ka rokohia ia tawhā hei tau tōpū 8-pī kaua hei rite-tere 16-pī, me tētahi take whakarea tiritahi mō ia rōpū iti kia ora tonu ai te awhe hurihuri ā-rohe o ia rōpū. Ko A16 te tika hohenga: ka noho tonu ki te 16 pī ngā uara e rere ana i waenga i ngā paparanga i te whakatau. Ka kōpakina ngā taumaha; kāore te tatauranga.
He mahi puku, nui hoki tā te wāhanga tuarua. He uaua ake te quantize i ngā hohenga i ngā taumaha: kei roto i a rātou ngā ara motuhake tino nui — he iti noa ngā āhuatanga e nui rawa ana ō rātou rahinga i ērā atu katoa, ā, e kawe ana i tētahi wāhanga kore-ōrite o te whanonga a te tauira. Ki te pana koe i ērā ki te tika iti, ka kutikuti tika koe i te tohu i hiahiatia e koe. Ka karo ngā tikanga taumaha-anake i te raru mā te waiho i te ara hohenga; koinā te take kua huihui ngā mahi quantization tino nui ki runga i a rātou.
Nō reira, ka kuiti te pātai whaitake ki te mea kotahi: e hia noa ngā pī ka manawanui ai tētahi taumaha?
05W8A16 me W4A16 me te 3-pī: kei hea tino te ngaronga e haere ana
Ko te rangahau nui rawa i whakaputaina mō tēnei ko tā Neural Magic me Red Hat, "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization — neke atu i te 500,000 arotake puta noa i te whānau Llama-3.1 i te 8B, 70B me te 405B. Ko tā rātou hua matua, i roto i ā rātou tau:
| Hōputu | He aha kua kōpakina | Whakahokinga i pūrongotia ki te BF16 |
|---|---|---|
| W8A8-FP (8-pī) | Taumaha + hohenga, rite-tere | ~99.75% toharite — "tata kore-ngaro" |
| W8A8-INT (8-pī) | Taumaha + hohenga, tau tōpū | 99–100% i te 8B/70B; 98.2–99.3% i te 405B |
| W4A16-INT (4-pī) | Taumaha anake | 98–99.4% (8B); 98.6–99.8% (70B) |
He tau i pūrongotia e Neural Magic / Red Hat (arXiv:2411.02355) mō ngā tauira Llama-3.1 — te āhua o te whakawhitinga, kaua te tatauranga o tēnei tauira.
Ki te pānui pono, he tautohenga tēnā e kī ana kaua e whai whakaaro: ka whakahoki te 4-pī i te 98–99%, nō reira he aha te take kia aro? Nā te mea he paearu kōwhiri-maha mātauranga ērā, tata ki te taputapu kino rawa e wātea ana hei kite i te kino quantization. He whakautu tika kotahi tō tētahi pātai paearu, ka eke i te tohu kotahi, rua rānei, ā, he iti te hapa quantization, he pāhorangi tata-toharite-kore i runga i ia logit; i tētahi pātai poto me te taumārō whānui, ka horomia taua pāhorangi.
Nā, tuhia he 4,000 tohu o tētahi tuhinga hangarau, pupuria he reo puta noa i tētahi kōrero 40,000-tohu, whakaputahia rānei he JSON me whakaū i te whakamātautau tuatahi. Ka tāuiratia taua pāhorangi tonu i ngā mano wā i te raupapa, ā, ka whakaahua ia tāuiratanga i tērā e whai ake. Ehara te momo hapa i te whakautu hē — he kōtiti: he pupuri kua ngāwari haere i te tohutohu mai i te hurihanga tuatahi, he āhua tangata kua mutu tana rite ki a ia anō i te tata tohu 3,000, he kōpani-tuhi kāore rawa e tae mai.
Kei hea tino ngā pī e haere ana
Kāore te hapa quantization e horapa ōrite ana puta noa i ngā āheinga o tētahi tauira; ka kohia ki te waha. Ko ngā whanonga nui-tapeke — te wetereo tika Ingarihi, te āhua o tētahi mahi Python, ngā mano meka matua mō te ao — kua whakawaeahia tauritenga puta noa i ngā taumaha maha, ā, ka ora tonu ahakoa he aha. Ko ngā kokonga tino-whakatikaina, rahinga-iti o te tohatoha taumaha e mau ana i ngā mea onge: te waitohu tika o tētahi API mōhio-iti, te rangi ā-tuhi a tētahi kaituhi motuhake, te rerekētanga i waenga i ngā CVE tata-tonu e rua. Ko ērā ngā uara ka whakaporowhitatia ngātahitia e tētahi ripanga anga-rahi.
Ko te hikoi mai i te 4 pī ki te 3 te wāhi ka mutu te noho hei tāke, ka noho hei pari. I kitea e ngā ture whakarahi whakatau-pī a Dettmers rāua ko Zettlemoyer ko te 4-pī te wāhi e whakanui rawa ai te tika kore-tauira ia pī o te mahara tauira, ā, ka noho matua te hekenga i raro iho i tēnā. E 16 ngā taumata ia rōpū mō te 4-pī; e 8 mō te 3. Kei te tapahi koe i te rautaki o tētahi ripanga kua anga-rahi kē, ā, ka tata ruarua te hapa porowhita ia taumaha.
Te tatauranga e whakatau ana
Ko te whakataurite whai take he tatauranga, ehara i te paearu. Ko te mea e ora tonu ana i muri i te kōpaki he tata ki te tawhā × pī ia tawhā — te rahinga kawe-mōhiohio o te kōnae taumaha. Whakaarohia ngā tūāpapa whakapae e rua he kounga ōrite ia tawhā:
- Tūāpapa A: 400B tawhā i te 3 pī → tata ki te 150 GB taumaha ka pupurihia.
- Tūāpapa B: 160B tawhā i te 8 pī → tata ki te 160 GB taumaha ka pupurihia.
Mā te tatauranga tawhā mata, ka toa a A e rua me te hawhe ki te kotahi — te nama e ārahina ana e te kāri tauira. Mā te mōhiohio kua pupurihia, he ōrite rāua, ā, he painga hanganga anō tō B: e noho ana ōna taumaha i runga i tētahi ripanga menemene rawa e mau tonu ai tōna whanonga waha, engari kua quantized te waha o A ki roto i te papatipu. E maha ake ngā mea e mōhio ana a A; ka mōhio a B ki āna e mōhio ai i te mārama ake. Mō te hanga roa, te hōputu tika, te whakahaere āhua, me te ū pono ki te horopaki-roa, tangohia a B.
He tatauranga whakaahua, ehara i te paearu: e whakaatu ana he aha i rerekē ai te kī "tūāpapa nui ake" me te kī "tauira i whakahaeretia pai ake".
06He aha ai te 27 taera tohu whakangungu e whakakino ake ai te quantization kaha
He hua tuarua, iti ake te māramatanga wawe, e pā tika ana ki tēnei tūāpapa. I kitea e te pepa a Kumar mā (ICLR 2025), Scaling Laws for Precision, kia nui ake ngā raraunga i whakangungua ai tētahi tauira, kia nui ake te kino ka mahia e te quantization-whai-muri-whakangungu ki a ia. Ka whakaiti te tika iti i te tatauranga tawhā whai-hua a tētahi tauira, ā, ka tipu taua hekenga me ngā raraunga whakangungu-tōmua kua mitihia — tae rawa atu ki te wāhi ka kino kē te raraunga whakangungu-tōmua taapiri mēnā ka whakaaro koe ki te quantize kaha, nā te mea kei te whakatakoto koe i tētahi hanganga menemene ka mukua e te ripanga.
He hanganga taumaha anga-rahi, tauritenga tō tētahi tauira i whakangungua ki te 5T tohu, me te wātea toenga. Ko tētahi tauira i whakangungua ki te 27T tohu — te pūtea a MiMo-V2.5-Pro — i whakapaua taua raraunga tāpiri ki te tuhi i ngā wehewehenga menemene ake ki ngā uara taumaha tino tika ake, koirā tino te mea kāore e taea e tētahi ripanga 3-pī te whakaatu.
Kātahi ko te meka FP8 mai i te wāhanga 03, te tautohenga mā rawa o tēnei tuhinga. I whakangungua-tōmua a MiMo-V2.5-Pro i te tika-ranu FP8. I whakaahuatia ōna taumaha mai i te tīmatanga e tētahi kaiwhakatika e mahi ana i te tata waru pī o te tika. Nō reira, ko te whakahaere i a ia i ngā taumaha 8-pī ehara i te whakaaro-muri ngaro kua akiakihia ki runga i tētahi tauira 16-pī — he tata ki te whakahoki i ngā taumaha ki te rangi tatauranga i akona ai rātou. He iti noa te hanganga menemene kei raro iho i te pī tuawaru hei ngaro. Kāore e taea te kī pērā mō ngā taumaha kua akiakihia ki te toru.
07REAP: he aha ai he momo ngaronga rerekē te tapahi mātanga i te porowhita
E rua o ngā taumata e toru e mau ana i te taunga REAP — he mahi rerekē tino motuhake i te quantization. Ko REAP (Router-weighted Expert Activation Pruning, Cerebras Research, arXiv:2510.13999, ICLR 2026) he kōpaki hikoi-kotahi mō ngā tauira MoE korekore: ka whakatau ia mātanga i runga i ngā uara kēti a te kaiwhakarite ara me te ine whakaputa a te mātanga, kātahi ka muku i ngā mea tatauranga-iti rawa. E pūrongo ana a Cerebras i ngā hua tata-kore-ngaro mō te hanga waehere me te karanga taputapu ahakoa i te 50% o ngā mātanga kua tapahia.
Ka whakamahia e mātou te REAP mō te take i hangaia ai: i ngā tūāpapa nui rawa, koinā te mea e taea ai te whakahaere i a rātou. Engari kia mōhio ki te rerekētanga momo. Ka whakapoke iti te quantization i ia taumaha; ka tango katoa te tapahi i ngā mātanga. Mēnā ka whakaritea-ara kaha tō kawenga mahi ki tētahi āheinga i noho ai i tētahi mātanga onge-kōwhiritia — he reo hōtaka iti, he kupu ngaio motuhake — kāore i te heke te kounga. Kua ngaro, ā, kāore he tono ka whakahoki mai i a ia. Ka pai tonu ngā toharite; kāore pea tō kēhi waha.
Kāore he taunga REAP kei runga i a MiMo-V2.5-Pro · W8A16. Kei konā tonu ngā mātanga 384 katoa, ā, ka kōwhiri te kaiwhakarite ara mai i te huinga katoa i whakangungua ai te tūāpapa.
| Taumata | Pī taumaha | Mātanga | He aha koe ka tuku ai |
|---|---|---|---|
| W8A16 | 8 | Kei konā katoa | Kāore he mea hanganga. Te taumata pono. |
| W4A16-AUTOROUND-REAP | 4 | Kua tapahia | Ngā taipitopito taumaha menemene me ngā mātanga hokohoko-iti, hei utu mō tētahi tūāpapa nui ake i te awhe utu ōrite. |
| 3BIT-REAP | 3 | Kua tapahia | He rautaki taumaha nui me ngā mātanga hokohoko-iti — te utu mō te whakahaere i ngā tūāpapa nui rawa i te tere. |
He mea whai take kia whakarongo tikaina te taumata waenga. Ehara a AutoRound i te porowhita ngāwari: ka whakamahia e te tikanga a Intel (arXiv:2309.05516) te heke-paiaka-tohu puta noa i te tata 200 hikoi hei whakatikatika i ngā whakatau porowhita me te awhe kutikuti ia rōpū taumaha, ā, e pūrongotia ana he tata-kore-ngaro i te 4 pī i te nuinga o ngā tautuhinga, e toa ana i a GPTQ, AWQ, HQQ me OmniQuant puta noa i te nuinga o ngā whirihoranga i whakamātauria. He tauira 4-pī tino pai tētahi tauira W4A16-AUTOROUND. He tauira 4-pī tonu ia.
08Ko wai me kōwhiri i te taumata 8-pī — ko wai kaua
E noho ana a MiMo-V2.5-Pro · W8A16 ki te utu papa o te rārangi katoa — $0.50 tāuru / $2.00 whakaputa ia miriona tohu, te awhe ōrite ki ngā tauira 4-pī taumata-tāuru, ā, tata ki te hauwhā o te utu tāuru o te tauira utu-nui rawa e whakahaeretia ana e mātou. Kāore koe e utu i tētahi utu nui ake mō te pono; kei te kōwhiri koe i tētahi mata rerekē o te pūtea ōrite.
Kōwhiria ina
- Ko te roa o te whakaputa te mahi. Te tuhinga hangarau whitinga-roa, te tuhinga-parau roa-upoko, ngā rerenga tuhinga katoa — he wāhi ia tohu i whakaputaina e whakarea ai te pāhorangi kōpaki.
- He nui te reo me te ū-tonutanga. Te mahi āhua-tangata, te oro pakihi, he reo ka mau puta noa i tētahi wātea roa. Kei te waha te noho o te āhua, ā, ko te waha te mea tuatahi e kaingia ana e te quantization anga-rahi.
- Me whakaū te putanga hanganga i te whakamātautau tuatahi. Ka tautoko te tauira i te
json_schemame teresponse_format; he āhuatanga tika te ū-ki-te-schema, ā, ko ia tarai anō ka karohia he moni, he wā roa hoki kua ora. - Kei te whakakī koe i te horopaki. He uaua ake te ū tohutohu i te tohu 200,000 i te tohu 2,000, ā, ka heke me te tika taumaha — pērā anō he ara whakaaroaro roa, i reira ka whakarea ngā hapa i te ara ōrite.
- He kore-ārai te mahi mai i tōna ake āhua. Te rangahau haumarutanga, te tīma-whero, te tuhinga-parau kore-tātari, te tātaritanga rongoā, ture rānei tino whakaraerae — i reira he kino ake tētahi whakautu kua tāmi i te kore whakautu.
Kōwhiria tētahi atu ina
- Me whai koe i te tirohanga. He kuputuhi-anake tēnei tauira; he whai-tirohanga tōna tuakana/teina, a MiMo-V2.5 · W8A16, i te utu me te taumata ōrite.
- Me whai koe i te mātauranga-ao whānui rawa mō ngā pātai poto. Koinā te rangi e toa ai ngā tawhā taapiri o tētahi tūāpapa 3-pī nui ake, ā, kāore e tino kitea tōna kino quantization.
- He whakarōpū poto rahi-nui tō kawenga mahi. Kāore e tino kitea te kino quantization i reira — hokona te mea tere rawa.
09Te karanga i te API MiMo kore-ārai
Mēnā kei te aro tō waehere ki te SDK OpenAI, ko te huri i te URL tūāpapa me te id tauira te nuku katoa:
curl https://api.shannon-ai.com/v1/chat/completions \
-H "Authorization: Bearer $SHANNON_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "MiMo-V2.5-Pro-W8A16",
"messages": [
{"role": "user", "content": "Draft the threat model section."}
],
"reasoning_effort": "high",
"max_tokens": 16384,
"stream": true
}'Ka whakaae a /v1/messages me /v1/responses ki te id me ngā āheinga ōrite. Kei te tuhinga API a Shannon AI te tohutoro tawhā, ngā schema karanga-taputapu, ngā tauira putanga-hanganga, me tētahi papa whakamātau ora.
| Āheinga | MiMo-V2.5-Pro · W8A16 |
|---|---|
| Matapihi horopaki | 262,144 tohu |
| Whakaputa nui rawa ia tono | 16,384 tohu |
| Streaming | Āe — ngā reo e toru |
| Karanga taputapu | Āe |
Whakaaroaro / reasoning_effort | Āe |
Putanga hanganga json_schema | Āe |
response_format | Āe |
| Tirohanga / tāuru whakaahua | Kāo — whakamahia a MiMo-V2.5 |
| Paparanga whakakāhore / tātari whakaputa | Kāore |
| Utu ia 1M tohu | $0.50 tāuru / $2.00 whakaputa |
Ko ia tau ā-waho o konei nā tētahi atu, ehara i a mātou, i whakaputa, ā, kua tohuhia i te tuhinga. Nō ngā rangahau whānau Llama ngā hua whakahokinga quantization: e whakaatu ana rātou i te āhua o te whakawhitinga, kaua i ngā tatauranga a tēnei tauira.
10Ngā pātai auau
He aha te API MiMo kore-ārai?
Ko te pito mutunga a Shannon AI mō MiMo-V2.5-Pro, te tauira Whakaranu-o-Ngā-Mātanga taumaha-tuwhera a Xiaomi, e whakahaerea ana i runga i tā mātou ake tautau GPU me te kore paparanga whakakāhore, kore hoki tātari whakaputa. Ka taea te tae atu mā /v1/chat/completions, /v1/messages me /v1/responses, me te horopaki 262,144-tohu me te 16,384 tohu whakaputa ia tono.
He aha te tikanga o W8A16, ā, he aha i nui ai?
Ko te tikanga o W8A16 kua rokohanga ngā taumaha ki te 8 pī ia mea, engari ka tatauhia ngā hohenga ki te 16 pī. Koinei te taumata quantization pono-rawa o tā mātou rārangi. E kite ana ngā rangahau i whakaputaina ka tata whakahoki katoatia e ngā hōputu taumaha-anake 8-pī te tika o te tauira taketake, engari ka utu te 4-pī i tētahi wāhi iti o te pono mō te mahara, ā, ka utu te 3-pī i tētahi wāhanga nui.
Ka taea e tētahi tauira iti ake i te 8-pī te toa i tētahi tauira nui ake i te 3-pī?
He maha ngā wā, āe. Ko te mea e ora tonu ana i muri i te kōpaki he tata ki te tatauranga tawhā whakarea ki ngā pī ia tawhā, ā, ka kohitia te hapa quantization ki te whanonga onge, ki te waha. Ka pupuri tētahi tauira i te 8-pī i tōna kōherenga whitinga-roa, tōna hōputu tika, me tōna mātauranga waha; ka pupuri tētahi tauira nui ake kua parea ki te 3-pī i ōna tatauranga paearu whānui engari ka ngaro ērā āhuatanga.
Ka tautoko a MiMo-V2.5-Pro i te tirohanga, i te tāuru whakaahua rānei?
Kāo. He kuputuhi-anake a MiMo-V2.5-Pro i runga i a Shannon AI. Ka tautokona te karanga taputapu, te whakaaroaro me te reasoning_effort, te putanga hanganga JSON-schema me te response_format. Mēnā e hiahiatia ana te tāuru whakaahua, whakamahia te tauira tuakana/teina a MiMo-V2.5 W8A16, he whai-tirohanga i te utu ōrite.
E hia te utu o te API MiMo kore-ārai?
$0.50 ia miriona tohu tāuru me te $2.00 ia miriona tohu whakaputa. Koinei te utu papa o te rārangi Shannon AI, e waiho ai a MiMo-V2.5-Pro hei ara utu-iti rawa ki te hoko i te pono 8-pī i tēnei kūwaha.
Whakamātauria te API MiMo kore-ārai
Pono waru-pī, horopaki 262K, kore paparanga whakakāhore — i te utu papa o te rārangi.
Pānuitia ngā Tuhinga API Ngā Kāri Tauira KatoaE whakahaerea ana i raro i te Kaupapa Here Whakamahi Haepapa a Shannon AI
Ngā puna kōrero: Pūrongo hangarau MiMo, arXiv:2505.07608 · Kāri tauira MiMo-V2.5-Pro (Xiaomi) · "Give Me BF16 or Give Me Death"?, arXiv:2411.02355 · REAP the Experts, arXiv:2510.13999 · AutoRound, arXiv:2309.05516 · k-bit inference scaling laws, arXiv:2212.09720 · Scaling Laws for Precision, Kumar mā, ICLR 2025. Ko ngā tatauranga paearu katoa e rite ana ki tā ngā kaituhi i pūrongo ai. He hua a Shannon AI nā Shannon Lab LLC, New Mexico, USA.