MiMo API Yosasidwa pa Kulondola kwa 8-Bit Kwenikweni

Mindandanda yambiri imakugulitsani chiwerengero cha zigawo. Ndicho nambala yolakwika yoti mugulire, chifukwa mtundu uliwonse wa zolemera zotseguka waukulu umene mungathe kupeza kwenikweni wapanikizidwa kuti ukwane pena. Funso loona si "izi zinali zazikulu bwanji zisanapanikizidwe?" koma "ndi zochuluka bwanji zomwe zikadalipo?" Nkhaniyi ikukhudza funso limeneli, ndi za mtundu mu mndandanda wathu kumene yankho ndi pafupifupi zonse zilipobe.

01Chomwe MiMo API yosasidwa imakupatsani kwenikweni

MiMo API yosasidwa ndi id imodzi ya mtundu, MiMo-V2.5-Pro-W8A16, yofikika kudzera mu zilankhulo zitatu zomwe zimafikira zolemera zomwezo pa gulu lathu la ma GPU: /v1/chat/completions, /v1/messages ndi /v1/responses. Zonse zitatu zimachita streaming ndipo zimathandiza kuyitanitsa ntchito ndi kuganiza.

Chosasidwa chili ndi tanthauzo lopapatiza pano: palibe wosanjikiza wa kukana patsogolo pa mtunduwo, palibe chosefera pa zotulutsa zake. Sitiimitsa, sitigawa m'magulu, kapena kulemba mwatsopano. Mumapeza khalidwe lenileni la maziko — chimene kwa mtundu wa zolemera zotseguka wa mlingo wa kafukufuku chikutanthauza kuti umagwira ntchito ndi kafukufuku wa chitetezo, nkhani zopeka za akuluakulu kapena zachiwawa, kuganiza kwa zachipatala ndi zamalamulo, ndi mapempho amene ma endpoint amalonda otetezedwa amakana mwachizolowezi. Ndi chida cha akatswiri pansi pa Ndondomeko yathu ya Kagwiritsidwe Kaudindo, osati kusapezeka kwa malamulo.

02Mbadwo wa MiMo: kudalira kwa Xiaomi pa pretraining yoyamba ndi kuganiza

MiMo si mtundu wochokera ku Llama wokhala ndi dzina latsopano; ndi mzere wodziyimira payokha wochokera ku gulu la MiMo la Xiaomi, ndipo lingaliro lake loyambira ndilofunika musanayese mtunduwo. Umayamba ndi MiMo-7B (April 30, 2025, wa layisensi ya MIT), umene lipoti lake la ukadaulo likunena kuti labotale zambiri zimasamalira kuganiza ngati vuto la pambuyo pa maphunziro loti liikidwe ndi RL, ndipo izi ndi zosemphana. Xiaomi inakonzanso pretraining mozungulira izo m'malo mwake: kusakanikirana kwa magawo atatu pa matiriliyoni 25 a ma token, gawo lachiwiri liyendetsa pafupifupi 70% masamu ndi code, gawo lachitatu likuphatikiza zizindikiro za kuganiza zopeka pomwe likukulitsa context kuchoka 8K kupita 32K — kuphatikiza cholinga cha Multi-Token Prediction chimene chimagwiranso ntchito ngati mutu wa speculative-decoding.

Zotsatira zake zinali mtundu wa maziko wa 7B umene unkakhala ngati chinthu chachikulu kwambiri. MiMo-7B-Base inapeza 32.9 pa AIME 2024 pomwe Qwen2.5-7B inapeza 10.1, Gemma-2-9B 0.0 ndi Llama-3.1-8B 0.3. Maphunziro apambuyo pake kenako anayenda RL pa mavuto 130,000 otsimikizika a masamu ndi code okhala ndi mphotho ya code yoyendetsedwa ndi kuvuta kwa mayeso, kupewa msampha wa mphotho zochepa umene umasokoneza RL ya code. Zotsatira zake zinafika patsogolo pa o1-mini:

AIME 2025, ziwerengero zofalitsidwa ndi Xiaomi kuchokera ku lipoti la ukadaulo la MiMo (arXiv:2505.07608). MiMo-7B-RL inanenanso 95.8 pa MATH-500 ndi 57.8 pa LiveCodeBench v5.

Banjali linakulirakulira m'malo mopatuka: MiMo-VL-7B (June 2025) inaika wosindikiza wa Qwen2.5-ViT pa msana womwewo, ndipo December 2025 inabweretsa MiMo-Audio-7B pamodzi ndi chinsinsi choyamba cha MoE, MiMo-V2-Flash (309B zonse / 15B zogwira ntchito, ma token 27T, kulondola kosakanikirana kwa FP8). MiMo-V2-Pro inatsatira mu March 2026 pa pafupifupi 1T / 42B zogwira ntchito ndi context ya 1M, pambuyo pa milungu pa OpenRouter monga "Hunter Alpha". MiMo-V2.5 ndi MiMo-V2.5-Pro zinatumizidwa pamodzi pa April 22, 2026, zonse za MIT.

03Mkati mwa MiMo-V2.5-Pro: momwe maziko amawonekera

Malinga ndi kaadi ya mtundu yofalitsidwa ndi Xiaomi, MiMo-V2.5-Pro ndi mtundu wa Mixture-of-Experts wosakanikirana wokhala ndi zigawo zonse tiriliyoni 1.02 ndi mabiliyoni 42 zogwira ntchito pa token pa zosanjikiza 70 — imodzi yolimba, 69 za MoE — ndi akatswiri 384 ogawidwa ntchito, 8 pa token. Kukula kobisika ndi 6144, ndi kukula kwa MoE kwapakati kopapatiza kwa 2048.

Zisankho ziwiri zimapanga mmene zimamvekera pa kugwiritsa ntchito. Unyolo wa chisamalira ndi wosakanikirana: zosanjikiza 10 za chisamalira chonse ndi zosanjikiza 60 za sliding-window zokhala ndi zenera la 128, kuphatikiza chikoka chophunziridwa cha attention-sink. Xiaomi ikunena kuti izi zimachepetsa kusunga kwa KV-cache pafupi ndi kuwirikiza 7 poyerekeza ndi chisamalira chonse cha global pomwe zikusunga mtundu wa context yaitali — kusiyana pakati pa zenera limene ndi lalitali mwa dzina ndi limene mungathe kudzaza. Imaperekanso zosanjikiza zitatu za Multi-Token Prediction, kupatsa decoder mutu wa speculative.

Maphunziro anali matiriliyoni 27 a ma token pa kutalika kwachibadwa kwa mzere wa 32K mu kulondola kosakanikirana kwa FP8 (E4M3), kenako kukonza bwino koyang'aniridwa, RL yaikulu ya agentic ndi distillation ya multi-teacher on-policy. Sungani mfundo ya FP8, ndicho chotsutsana cholimba kwambiri chokha mu nkhaniyi chotumikira maziko awa pa 8-bit.

Xiaomi ikunena za mtunduwo pa 57.2% pa SWE-bench Pro, 72.9 pa τ3-Bench pa kugwiritsa ntchito zida ndi 48.0% pa Humanity's Last Exam — manambala a wogulitsa okha pa mtundu wosapanikizidwa, otchulidwa motero.

Timatumikira zenera la ma token 262,144, osati 1M yathunthu ya maziko — context imene tingatsimikizire pa kufulumira, m'malo monkhalira nambala yaikulu imene imawonongeka pansi pa katundu.

04Chomwe W8A16 amatanthauza, molondola

Mtundu uliwonse mu mndandanda uwu umanyamula chowonjezera cha quantization chimene chimafotokoza kusintha kwenikweni, koyerekezeka m'malo mwa mlingo wa malonda. Athu ndi 3BIT-REAP, W4A16-AUTOROUND-REAP ndi W8A16. MiMo-V2.5-Pro ndi wachitatu.

Mawuwa amagawika m'magawo awiri. W8 ndi kulondola kwa zolemera: parameter iliyonse imasungidwa ngati chiwerengero cha integer cha 8-bit m'malo mwa float ya 16-bit, ndi chiwerengero chogawana cha scaling pa gulu laling'ono lililonse choncho mlingo weniweni wa gulu ulionse ukupulumuka. A16 ndi kulondola kwa zochitika: mtengo umene umayenda pakati pa zosanjikiza pa nthawi ya kuganiza umakhalabe pa mabiti 16. Zolemera zimapanikizidwa; masamu safunika.

Theka lachiwiri limeneli limagwira ntchito yofunika mokhalira. Zochitika ndi zovuta kwambiri kupanikiza kuposa zolemera: zili ndi njira zodabwitsa kwambiri — magawo ochepa amene kukula kwawo kumapambana chilichonse china ndipo amanyamula gawo losagwirizana la khalidwe la mtunduwo. Kankhani izo ku kulondola kotsika ndipo mumadula ndendende chizindikiro chimene munkafunikira. Njira za zolemera zokha zimapewa vutoli posiya njira ya zochitika yosasinthidwa, ndicho chifukwa ntchito yeniyeni ya quantization yasonkhana pa izo.

Choncho funso lenileni limadzicheperetsa ku chinthu chimodzi: ndi ma bit ochepa bwanji amene kulemera kungathe kupirira?

05W8A16 motsutsana ndi W4A16 motsutsana ndi 3-bit: kumene kutayika kumapita kwenikweni

Kafukufuku waukulu kwambiri wofalitsidwa pa izi ndi wa Neural Magic ndi Red Hat "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization — kupitirira miyeso 500,000 pa banja la Llama-3.1 pa 8B, 70B ndi 405B. Zotsatira zawo zazikulu, mu manambala awo:

Mawonekedwe	Chomwe chapanikizidwa	Kubwezeretsedwa kofalitsidwa poyerekeza ndi BF16
W8A8-FP (8-bit)	Zolemera + zochitika, float	~99.75% avereji — "pafupifupi popanda kutayika"
W8A8-INT (8-bit)	Zolemera + zochitika, integer	99–100% pa 8B/70B; 98.2–99.3% pa 405B
W4A16-INT (4-bit)	Zolemera zokha	98–99.4% (8B); 98.6–99.8% (70B)

Ziwerengero zofalitsidwa ndi Neural Magic / Red Hat (arXiv:2411.02355) pa mitundu ya Llama-3.1 — mawonekedwe a kusinthana, osati chiwerengero cha mtundu uwu.

Zowerengedwa moona, izi zikuwoneka ngati chotsutsana chosatchinjiriza kusamala: 4-bit imabweza 98–99%, choncho n'chifukwa chiyani mungasamale? Chifukwa izo ndi ma benchmark a maphunziro a zisankho zambiri, pafupi ndi chida chosalondola kwambiri chopezeka pa kuzindikira kuwonongeka kwa quantization. Funso la benchmark lili ndi yankho limodzi lolondola lofikika mu token imodzi kapena ziwiri, ndipo cholakwika cha quantization ndi chaching'ono, phokoso lofanana ndi zero-mean pa logit iliyonse; pa funso laifupi lokhala ndi mpata wotakata, phokosolo limatengedwa.

Tsopano lembani ma token 4,000 a chikalata cha ukadaulo, sungani mawu pa nkhani ya ma token 40,000, kapena tulutsani JSON imene iyenera kutsimikizika nthawi yoyamba. Phokoso lomwelo limatengedwa maulendo masauzande motsatizana, chitsanzo chilichonse chikukhudza chotsatira. Njira ya kulephera si yankho lolakwika — ndi kupatuka: kukakamira kofooka kwa lamulo kuchokera pa kutembenuka koyamba, munthu amene amasiya kumveka ngati iye yekha pafupi ndi token 3,000, chibokosi chotseka chimene sichifika konse.

Kumene ma bit amapita kwenikweni

Cholakwika cha quantization sichifalikira mofanana pa maluso onse a mtundu; chimasonkhana mu mchira. Khalidwe la mlingo wapamwamba — Chingerezi cholondola pa galamala, mawonekedwe a ntchito ya Python, zowona chikwi chapamwamba za dziko lapansi — zimalembedwa mobwerezabwereza pakati pa zolemera zambiri ndipo zimapulumuka pafupifupi chilichonse. Malo olondola kwambiri, aang'ono a kagawidwe ka zolemera amasunga zinthu zosowa: mawonekedwe enieni a API yosadziwika, kayendedwe ka wolemba wina, kusiyana pakati pa ma CVE awiri ozungulira. Amenewo ndi mitengo imene grid yolimba imazungulira pamodzi.

Sitepe yochoka pa mabiti 4 kupita ku 3 ndicho kumene izi zimasiya kukhala msonkho ndikukhala chigwa. Malamulo a kukula kwa k-bit inference a Dettmers ndi Zettlemoyer anapeza kuti 4-bit ndicho malo amene amakweza kulondola kwa zero-shot pa bit ya kukumbukira kwa mtundu, ndi kuwonongeka kukuchuluka pansi pa apo. Mabiti anayi amapereka milingo 16 pa gulu; atatu amapereka 8. Mukugawa pakati kulondola kwa grid yomwe inali yolimba kale, ndipo cholakwika cha kuzungulira pa kulemera kulikonse chimafika pafupifupi kaŵiri.

Masamu amene amasankha izo

Kuyerekezera koyenera kupanga ndi masamu, osati benchmark. Chomwe chimapulumuka kupanikiza ndi pafupifupi zigawo × ma bit pa zigawo — kuchuluka konyamula chidziwitso kwa fayilo ya kulemera. Tengani maziko awiri oganiziridwa a mtundu wofanana pa zigawo iliyonse:

• Maziko A: zigawo 400B pa mabiti 3 → pafupifupi 150 GB ya kulemera kosungidwa.
• Maziko B: zigawo 160B pa mabiti 8 → pafupifupi 160 GB ya kulemera kosungidwa.

Pa chiwerengero chenicheni cha zigawo A imapambana kaŵiri ndi hafu ku imodzi — nambala imene kaadi ya mtundu imayamba nayo. Pa chidziwitso chosungidwa ndi zofanana, ndipo B ili ndi ubwino wa kapangidwe kuwonjezera: zolemera zake zili pa grid yolondola mokwanira kotero kuti khalidwe lake la mchira lilibe kuwonongeka, pomwe mchira wa A wapanikizidwa mkati mwa gulu lonse. A imadziwa zinthu zambiri; B imadziwa zomwe ikudziwa molondola kwambiri. Pa kupanga kwautali, mawonekedwe olondola, kuwongolera kwa mtundu ndi kutsatira kolondola kwa context yaitali, tengani B.

Masamu ochititsa chidwi, osati benchmark: zikuwonetsa chifukwa chomwe "maziko akulu kwambiri" ndi "mtundu wotumikiridwa bwino kwambiri" ndi zonena zosiyana.

06Chifukwa matiriliyoni 27 a ma token a maphunziro amapangitsa quantization yolimba kukhala yoipa kwambiri

Chotsatira chachiwiri, chosavuta kumveka, chikukhudza mwachindunji maziko awa. Pepala la Kumar et al. la ICLR 2025 Scaling Laws for Precision likupeza kuti deta yambiri imene mtundu unaphunzitsidwa nayo, ndi kuwonongeka kwambiri kumene quantization ya pambuyo pa maphunziro imachita pa iwo. Kulondola kotsika kumachepetsa chiwerengero cha zigawo chogwira ntchito cha mtundu, ndipo kuwonongeka kumeneku kumakula ndi deta ya pretraining yotengedwa — mpaka pamlingo umene deta yowonjezera ya pretraining ingakhale yowononga chenicheni ngati mukufuna kuchita quantization yolimba, chifukwa mukuika kapangidwe kolondola kamene grid idzachifafaniza.

Mtundu wophunzitsidwa pa ma token 5T uli ndi kapangidwe ka kulemera kolimba, kobwerezabwereza ndi malo owonjezera. Mtundu wophunzitsidwa pa ma token 27T — ndalama za MiMo-V2.5-Pro — inagwiritsa ntchito deta yowonjezerayo kulemba kusiyana kolondola kwambiri mu mitengo yolondola ya zolemera, ndicho ndendende chimene grid ya 3-bit singathe kuyimira.

Kenako pali mfundo ya FP8 kuchokera ku gawo 03, chotsutsana chowoneka bwino kwambiri mu nkhaniyi. MiMo-V2.5-Pro inaphunzitsidwa mu kulondola kosakanikirana kwa FP8. Zolemera zake zinapangidwa kuyambira poyamba ndi optimizer yogwira ntchito pa pafupifupi ma bit asanu ndi atatu a kulondola. Kuitumikira pa zolemera za 8-bit choncho si chinthu chowonongeka choganiziridwa pambuyo pake choikidwa pa mtundu wa 16-bit — ndi pafupi ndi kubweza zolemera ku dziko la manambala limene zinaphunziridwa mmenemo. Pali kapangidwe kolondola kochepa kwambiri pansi pa bit yachisanu ndi chitatu koti katayike. Chomwecho sichinganenedwe cha zolemera zomwezo zokakamizidwa kufika ku zitatu.

07REAP: chifukwa chomwe kudula akatswiri ndi mtundu wosiyana wa kutayika kuposa kuzungulira

Awiri mwa milingo itatu amanyamula chowonjezera cha REAP — ntchito yosiyana kwathunthu ndi quantization. REAP (Router-weighted Expert Activation Pruning, Cerebras Research, arXiv:2510.13999, ICLR 2026) ndi kupanikiza kamodzi kokha ka mitundu ya MoE yosakanikirana: imayerekezera katswiri aliyense pa mfundo za router gate ndi norm ya kugwira ntchito kwa katswiri, kenako imachotsa opeza mfundo zochepa kwambiri. Cerebras ikunena za zotsatira pafupifupi zopanda kutayika pa kupanga code ndi kuyitanitsa ntchito ngakhale pa 50% ya akatswiri odulidwa.

Timagwiritsa ntchito REAP chifukwa cha chomwe chimapangira: pa maziko akulu kwambiri ndicho chimene chimawapangitsa kutha kutumikiridwa. Koma onani kusiyana kwa mtundu. Quantization imasokoneza kulemera kulikonse pang'ono; kudula kumachotsa akatswiri kwathunthu. Ngati ntchito yanu imayenda kwambiri ku kuthekera kumene kunali mkati mwa katswiri wosankhidwa pang'ono — chinenero chaching'ono cha kupanga mapulogalamu, mawu apadera — sikuwonongeka. Kwatha, ndipo palibe kufunsa kumene kungakubwezereni. Avereji zimakhalabe zabwino kwambiri; nkhani yanu ya mchira mwina si choncho.

MiMo-V2.5-Pro · W8A16 sinyamula chowonjezera cha REAP. Akatswiri onse 384 alipo, ndipo router imasankha kuchokera ku gulu lonse limene maziko anaphunzitsidwa nalo.

Mlingo	Ma bit a zolemera	Akatswiri	Chomwe mukusinthana
W8A16	8	Onse alipo	Palibe chokhudza kapangidwe. Mlingo wa kulondola.
W4A16-AUTOROUND-REAP	4	Odulidwa	Tsatanetsatane wolondola wa zolemera kuphatikiza akatswiri a magalimoto ochepa, posinthana ndi maziko akulu kwambiri pa gulu lomwelo la mtengo.
3BIT-REAP	3	Odulidwa	Kulondola kwakukulu kwa zolemera kuphatikiza akatswiri a magalimoto ochepa — mtengo woyendetsa maziko akulu kwambiri pa kufulumira.

Mlingo wapakati uyenera kumveka bwino. AutoRound siyo kuzungulira kosaganizira: njira ya Intel (arXiv:2309.05516) imagwiritsa ntchito signed gradient descent pa masitepe pafupifupi 200 kukonza zisankho za kuzungulira ndi malire a kudula pa gulu la kulemera lililonse, ndipo yanenedwa kuti ndi pafupifupi yopanda kutayika pa mabiti 4 mu makonzedwe ambiri, kupambana GPTQ, AWQ, HQQ ndi OmniQuant pa makonzedwe ambiri oyesedwa. Mtundu wa W4A16-AUTOROUND ndi mtundu weniweni wabwino wa 4-bit. Ndi mtundu wa 4-bit chidakhalabe.

08Ndani ayenera kusankha mlingo wa 8-bit — ndi ndani ayi

MiMo-V2.5-Pro · W8A16 ili pa mtengo wotsika kwambiri wa mndandanda wonse — $0.50 kulowa / $2.00 kutuluka pa miliyoni ya ma token, gulu lomwelo la mitundu ya mlingo woyambira wa 4-bit ndipo pafupifupi gawo limodzi mwa anayi la mtengo wa kulowa wa mtundu wokwera mtengo kwambiri umene timatumikira. Simulipira mtengo wowonjezera pa kulondola; mukusankha mzere wina wa ndalama zomwezo.

Sankhani pamene

• Kutalika kwa zotulutsa ndicho ntchito. Kulemba kwa ukadaulo kwautali, nkhani zopeka za kutalika kwa mutu, kupitiriza konse kwa zolemba — token iliyonse yopangidwa ndi imene phokoso la kupanikiza limasonkhana.
• Mawu ndi kugwirizana ndikofunika. Ntchito ya munthu, mawu a mtundu, mtundu wosungidwa pa gawo lalitali. Mawonekedwe amakhala mu mchira, ndipo mchira ndicho quantization yolimba imadya poyamba.
• Zotuluka zokonzedwa ziyenera kutsimikizika nthawi yoyamba. Mtunduwo umathandiza json_schema ndi response_format; kutsatira schema ndi mkhalidwe wa kulondola, ndipo kuyesanso kulikonse kopewedwa ndi ndalama ndi kuchedwa kosungidwa.
• Mukudzaza context. Kutsatira malangizo pa token 200,000 ndikovuta kuposa pa token 2,000, ndipo kumawonongeka ndi kulondola kwa zolemera — mofananamo unyolo wautali wa kuganiza, kumene zolakwika zimasonkhana chimodzimodzi.
• Ntchito ndi yosasidwa mwachibadwa. Kafukufuku wa chitetezo, red-teaming, nkhani zopeka zosasefedwa, kusanthula kwapamwamba kwa zachipatala kapena zamalamulo — kumene yankho lofooketsedwa ndi loipa kuposa palibe yankho konse.

Sankhani chinanso pamene

• Mukufunika masomphenya. Mtunduwu ndi mawu okha; mnzake MiMo-V2.5 · W8A16 ali ndi masomphenya pa mtengo womwewo ndi mlingo womwewo.
• Mukufunika chidziwitso chotakata kwambiri cha dziko lapansi pa mafunso aafupi. Umenewo ndi mkhalidwe kumene zigawo zowonjezera za maziko akulu kwambiri a 3-bit zimapambana ndipo kuwonongeka kwake kwa quantization sikuoneka kwambiri.
• Ntchito yanu ndi kuika m'magulu kwaifupi kwa kuchuluka kwakukulu. Kuwonongeka kwa quantization ndi kosaoneka kumeneko — gulani chilichonse chofulumira kwambiri.

09Kuyitana MiMo API yosasidwa

Ngati code yanu ikuyang'ana SDK ya OpenAI, kusintha base URL ndi id ya mtundu ndicho kusamuka konse:

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
  }'

/v1/messages ndi /v1/responses zimalandira id yomweyo ndi kuthekera komweko. Mndandanda wa ma parameter, ma schema a kuyitanitsa chida, zitsanzo za zotuluka zokonzedwa ndi malo oyeserera amoyo zili mu zolemba za API za Shannon AI.

Kuthekera	MiMo-V2.5-Pro · W8A16
Context	ma token 262,144
Kutuluka kwambiri pa pempho	ma token 16,384
Streaming	Inde — zilankhulo zonse zitatu
Kuyitanitsa ntchito	Inde
Kuganiza / reasoning_effort	Inde
Zotuluka zokonzedwa za json_schema	Inde
response_format	Inde
Masomphenya / zolowetsa za zithunzi	Ayi — gwiritsani ntchito MiMo-V2.5
Wosanjikiza wa kukana / kusefa kwa zotulutsa	Palibe
Mtengo pa ma token 1M	$0.50 kulowa / $2.00 kutuluka

10Mafunso ofunsidwa kawirikawiri

---

Magwero: Lipoti la ukadaulo la MiMo, arXiv:2505.07608 · Kaadi ya mtundu ya 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 · Malamulo a kukula kwa k-bit inference, arXiv:2212.09720 · Scaling Laws for Precision, Kumar et al., ICLR 2025. Ziwerengero zonse za benchmark ndi zofalitsidwa ndi olemba awo. Shannon AI ndi chinthu cha Shannon Lab LLC, New Mexico, USA.