Dellu ci xibaar yi
ProduitAI Understanding

OpenBMB genne na xeetu 2B bu ubbeeku ak dooley làkku japonais

GIGAZINE xamlena ni sosete juntuwaay yu bees yu waa Chine OpenBMB gennena MiniCPM5-2B, benn model bu ubbeeku bu am 2 milyaar ci ay poñ yu ñu méngale ak Gemma 4 12B bu gina magg ci Index bu Xarañteg Analysis Artifisiyel.

4 min readRead the linked source
Source-provided image accompanying OpenBMB releases open 2B model with reported Japanese-language strength
RoyuwaaySource biñ enregistre
Siiwalkat
gigazine.net
Lëkkalekaayu cosaan
gigazine.nethttps://gigazine.net/gsc_news/en/20260908-openbmb-minicpm5-2b/
Xeetu balluwaay
Source buñ lëkkale — joxe wuñu status source bu njëkk bi.
KontekstXam lii ci 60 seconde

Tambalil fii

Term yu am solo

Memoire (Memoire agent)
Kontekst buñ denc bi ab ndawu IA di jëfandikoo ci jéego yi wala sesioŋ yi ngir gëna mëna wéy.
Modèle ouvert
Modèle buñ génne ak poids publik wala kode ngir saytu, méngale ak jëfandikoowaat.
Kantifikaasioŋ
Soppi diisaayu model ci formaa yu gëna ndaw yu melni 8-bit wala 4-bit.
Nattal sa boppModèlu IA leeral quiz

Lu xew

GIGAZINE dafa wax ni OpenBMB genne na MiniCPM5-2B ci 7 septembre 2026. Modèle bi ubbeeku amna 2,516,756,480 paramet, guddaay bu gëna mag di 131,072 jeton, te dañu koy séddale ci Hugging Face ak Model2C. GIGAZINE itam dafa wax ni amna demo navigateur. Sunu sukkandikoo ci GIGAZINE, test yu OpenBMB dañu wane ni MiniCPM5-2B moo gëna am doole ci Qwen3.5-4B ci jàngat yu bari. Xeetu gennup bi neena Cammbar bu ñu defar bi joxna ko 14 poñ ci Index Intelligence v4.3, mu def ko ci poñ yi ñu siiwal ci Gemma 4 12B. GIGAZINE dafay wax itam ni tontub làkku japonais dafa jaar yoon buñu ko méngale ak yeneen xeetu tontu yu ndaw.

GIGAZINE dafa wax ni OpenBMB, benn kompiñi IA bu waa Chine, genne na MiniCPM5-2B ci 7 septembre 2026. Xeet wi dañu ko wax ni mooy wuutu wala yokkute bu sukkandiko ci MiniCPM5-1B bi gëna ndaw te amna 2,516,756,480 paramet. Guddaayi fimu gëna mag biñ wax mooy 131,072 jeton.

Dafay wax ni OpenBMB dafa taxawal MiniCPM5-2B muy xeetu xët bu ubbeeku ngir jëfandikoo ci pegg yi. Gigazine neena model bi mën nga ko amee te doo fay ci Hugging Face ak ModelScope, ak lisence Apache 2.0. Dafay wax itam ni amna demo Space bu sukkandiko ci navigateur Hugging Face.

Ci wàllu liggéey, GIGAZINE neena OpenBMB dafa wane ni MiniCPM5-2B dafa raw Qwen3.5-4B ci test yu bari. Rapport bi neena itam ni Analysis Artifisiel joxna model bi 14 ci Index Intelligence v4.3, ap poñ bu GIGAZINE wax ni noonu la Gemma 4 12B joxe. Lii ay nattug rapoor la, duñu ay resultaa yuñ joxe ci jàngat bii.

GIGAZINE dafa joxe benn misaalu làkku japonais buy xelal nit ñi ci jëfandikoo ab kaaraange ekraŋ telefon bu xarañ, ba noppi wax ni tontu bi dafa defar bu baax. Ci misaal boobu, dafay wane ni làkku japonais amul benn njariñ ci liggéey yépp.

Ay leeral ci cosaan: gigazine.net ↗

Lu tax mu am solo

Benn xeetu ubbeeku bu mëna am lu tollu ci 2 milyaar ciy paarametre mënna tax ay aplikaasioŋu IA yu bërëb mbaa yu am ay jumtukaay yu gina am njariñ, rawatina fu doxal benn xeetu lu gëna magg seer lool mbaa yeex. Performance biñ wax ci làkku japonais amna solo itam ndax model yu ndaw yi dañuy faral di génne làkku japonais bu gëna néew doole wala bu wuute ci wàllu nàfar. Waaye, li ñuy wax ci liggéey bi dafay des ci source-reported: jàngat bii mënul defar boppam benchmark yi, xool anam yi ñuy méngale, wala wane ni model bi dafay def lu yaatu ni model 12-milyaar-parametre.

Modèlu làkk yu ndaw yi mën nañu am njariñ sudee latency, mémoire, njëgu hardware, privacy, wala liggéey bi nekkul ci net bi dañu am solo. Benn model bu am lisence Apache 2.0 mën na gëna yombal developpeur yi ngir saytu, méngale, ak boole, lépp di aju ci lisence bi ak këyitu model bi boppam.

Tegtale biñ wax ak Gemma 4 12B lu am solo la ndax dafay wax ci poñ yiñ joxe ci benchmark bi, du ci lim parametre yi kese. Buñu ko jàppee ni firnde ci kàttan gu mat sëkk: rapoor biñ joxe leeralul njaxasu benchmark bi, laajtu test yi, aparey bi, kantite bi, wala tóo njuumte yi.

Jàppale japonais mën na yaatal valeur pratique bu model compact ngir aplikaasioŋu lokal, waaye balluwaay bi dafay joxe benn misaal rek, du benn benchmark làkk wala jàngat jëfandikukat bu moom boppam.

Interactive Mechanism

Mekanism buy weccoo xalaat: naka lay doxee

Saytu xarala yu bees yi ci ginaaw yokkute bii ci anam wu weccoo xalaat.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Saytu konsept buy weccoo xalaat+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Li nga wara seetaan ci topp

Seetal jàngat yi moom seen bopp ci MiniCPM5-2B ci jamonoy japonais, xalaat, kodage, dëggu, kaaraange, ak liggéey yu yàgg. Developpeur yi dañu wara xoolaat kàrtu model bi ak li ñuy laaj ci samp gi balaa ñu koy jëfandikoo ci liggéey bi, lu ci melni runtime yiñ jàppale, soxlay hardware bi, version yiñ kantite, ak bépp tënk ci jëfandikoo gi. Modèle bi dañu ko xamle ni mën nañu ko yebbi ci Hugging Face ak ModelScope ci suufu Apache 2.0, ak demo navigateur bu jàppandi. Boroom bi ñu joxe mënul firndeel ni demo bi wala yebbite yi am nañu ci gox bu nekk, ni model bi dafa baax ci telefon yu xarañ yi, wala ni bépp serwiis buñ dalal amna gaaraati jëfandikoo bu ñuy fay wala bu amul fayda.

Test moomel sa bopp war na leeral ndax poñ yiñ joxe ci Intelligence Index dañu tekki ci liggéey bu am njariñ ci biti liggéey yiñ jàngat, rawatina ci xalaat japonais, topp tegtal, kodage, ak tontu yu dëggu.

Dañu wara xool anam yi ñuy duggee ci xëti Hugging Face ak ModelScope yiñ lëkkale. Source bi dafay wax ni dañuy séddale te doo fay, waaye du bind njëgu serwiis biy dalal, disponibilite ci gox bi, li ñuy laaj ngir yebbi, wala ndimbalu liggéey bi.

Guddaayi fiñu wax ci contexte bi mooy 131,072 token, waaye ci dëgg-dëgg contexte bu gudd bi ak memory bi ñuy laaj nekkul ci rapoor biñ joxe.

Jëfandikukat yiy xalaat dugal dañu wara xool bu baax kàrtu model bi, kaadar yiñ jàppale, li ñuy laaj ci hardware bi, këyitu kaaraange yi, ak bépp tënk buñu waxul ci xët wi.

Gid ak quiz yu ci méngoo

Model IA leeral nañu koTransformatërTaggat ci IANatt li nga xam — natt quiz IA bu amul faydaSeetal benn baat IA ci sunu glossaireToppal toppukaayu génne xeetu IA
Gis nga lii am njariñ?