Dellu ci xibaar yi
ProduitAI Understanding

PrismML genne na Bonsai 2, xeetu 27B buñ komprime ngir ay aparey lokal

PrismML genne na Bonsai 2, muy xeetu Qwen3.8 27B bu Alibaba bu am 5.9GB, muy dox ci telefon yu xarañ yi ak PC yi, fekk mingi wéy di am 98% ci màndarga yi mu amoon.

4 min readRead the linked source
Source-provided image accompanying PrismML releases Bonsai 2, a compressed 27B model for local devices
RoyuwaaySource biñ enregistre
Siiwalkat
techjuice.pk
Lëkkalekaayu cosaan
techjuice.pkhttps://www.techjuice.pk/prismml-bonsai-2-27b-llm-compression-smartphone-pc-ai-device/
Xeetu balluwaay
Source buñ lëkkale — joxe wuñu status source bu njëkk bi.
KontekstXam lii ci 60 seconde

Tambalil fii

Term yu am solo

Modelu làkk bu mag (LLM)
Benn xeetu làkk buñ tàggat ci corpus mbind yu bari ngir sos ak jàngat mbind.
IA ci aparey
Inference IA dañu koy defee ci barab bi ci hardware jëfandikukat bi, du ci sarwiisu cloud bu sori.
Référence
Test buñ yamale wala ensemble done yuñ jëfandikoo ngir natt ak méngale liggéeyu model bi.
Nattal sa boppModèlu IA leeral quiz

Lu xew

PrismML genne na Bonsai 2 27B, muy xeetu làkk bu yaatu buñ defar ngir mu mëna dox ci ordinatëri bopp ak telefon yu xarañ te amul infrastructure cloud. Modèle bi dafa wàññi Qwen3.8 27B bu Alibaba ba 5.9 gigabytes ci jëfandikoo kompresioŋ ternaire, ba noppi tëye 98% ci liggéeyu référence bi njëkk. Diisaay yi am nañu ci ndigalu Apache 2.0 te dañuy jàppale GPU NVIDIA jaaraleko ci CUDA ak aparey Apple jaaraleko ci MLX.

PrismML, benn entreprise bu gëstukat yu Caltech taxawal, ki ko jiite mooy Babak Hassibi, moo genne Bonsai 2 27B ci Alxamis. Modèle bi benn xeetu Qwen3.8 27B bu Alibaba la, wàññeeku ba 5.9 gigabytes jaaraleko ci kompresioŋ ternaire, biy soppi parametre 16-bit ci ñatti valeur: benn bu baax, benn bu baaxul, wala nul. Pexem dafay dindi ay parametre yu bari, loolu mooy tax model bi mëna tëye 98% ci performance bi njëkk, fekk mingi méngoo ak aparey bi ci dëkk bi.

Li ñuy génne dafay màndargaal yokkute bu am solo ci modelu Bonsai bi njëkk ci weeru màrs, bi denc 95% ci performance bi. Sunu sukkandikoo ci PrismML, xeetu njëkk bi yebbi nañu ko lu ëpp 11 milioŋ yoon, ak yeneen model yu ndaw yu dajale 2.6 milioŋ ciy yebbi, loolu dafay wane ni developpeur yi dañu jël IA bu lalu ci aparey. Modèle bu bees bi dafay dox ci GPU NVIDIA jaaraleko ci CUDA ak ci aparey Apple (Mac, iPhone, iPad) jaaraleko ci MLX, di jëfandikoo kernel yu ndaw yuñ jagleel ngir mëna def liggéey bu baax ci gox bi.

Diisaayu model yi am nañu ci saasi ci ndigalu Apache 2.0 License, loolu dafay tax ñu mëna boole ak soppali lu bari. PrismML dajale na 22,25 milioŋ ciy dolaar ci liggéey bi, te Khosla Ventures, Cerberus Capital, ak Caltech ñoo ko jàppale, Ion Stoica, benn ci ñi sos Databricks, moo nekkoon digalkat bi. Xeetu liggéeyukaay bi dafay xoole ci jëfandikoo redundancy ci model yu gëna mag ngir am parity performance bu kawe ak dayo fichier bu ndaw, pexe Hassibi digle ni dina gëna am njariñ su dayo model yi di yokk.

Ay leeral ci cosaan: techjuice.pk ↗

Lu tax mu am solo

Bii génne dafay wane ni IA bi mëna xalaat mën na doxal ci barab bi ci hardware konsomatër bi, di saafara jafe-jafey nëbbëtu ak dindi latency cloud ak njëgu abonemaa bi. Su PrismML amee parite performance bu kawe ak dayo fichier bu gëna ndaw, dafay awaase ci coppite bi joge ci IA bu sukkandiko ci cloud dem ci xalaat buñ séddale, bu sukkandiko ci aparey, def kàttani IA yu xarañ yi ñu mëna jëfandikoo te duñu soxla infrastructure yu mag ci santu done yi.

Mën doxal IA bu mëna xalaat ci barab bi ci aparey konsomatër yi dafay saafara jafe-jafe yu am solo ci wàllu kumpa ndax dafay fexe ba done jëfandikukat yi duñu musa génn ci aparey bi. Loolu dina dindi soxla transmisioŋ cloud, wàññi latency ak dindi njëgu abonemaa biy wéy di méngoo ak sarwis IA yi bawoo ci cloud. Ngir developpeur yi ak jëfandikukat yi, lii dafay tekki coppite bu am solo ci xalaat buñ séddale, bu sukkandiko ci aparey, muy lu am njariñ te nekk lu kenn xamul.

Tolluwaayu tëye bu rëy bi tollu ci 98% ci performance ci biir benn fichier bu 5.9GB dafay wane ni pexem kompresioŋ mën nañu baña yàq xarañteg model yu mag yi. Loolu amna njeexital yu am solo ci entreprise yi soxla IA ci aparey yi, lu ci melni pajum wérgi-yaram, xaalis, ak assistant personel, fu soberaanite done ak tontu ci jamono dëgg ñoo gëna am solo. Nature open-source bi ci génne gi dafay gaawlu jëfandikoo gi ndax dafay may askan wi mu tabax ci kaw ak gëna mëna jëfandikoo model bi ngir ay jafe-jafe hardware.

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

Developpeur yi war nañu xool bu baax ndax kernel yu ndaw yi am nañu ci ay architecture hardware yu bari, ak mébetu liggéeyukaay bi di kompresse ay model yu tollu ci téemeeri miliyaar ciy parametre ci weer yii ñu jémm. Rax ci dolli, boole gi mëna am ak ay aparey Apple, te waxtu génne bi dafay wane, dafay laaj ñu bàyyi xel ci kàttanu IA yi ci wàllu aparey yi ci kanam.

PrismML dafa wax ni mébetam bi ci topp mooy kompresse model yu am téemeeri milyaari parametre ci weer yii ci topp. Su amee ndam ci wàll wii dina gëna firndeel scalability bu ternaire compression ngir model frontier-class, mën na tax IA yu gëna xarañ yi jàppandi ci hardware bi ci dëkk bi.

Waxtu génne gi jurna ay xalaat ci mëna boole ay aparey Apple. Bi CEO Babak Hassibi bañee wax ci rapoor yi ci waxtaan yi ak Apple, jàppale bi Apple def ci kaadar MLX dafay wane ni dafa wara jëflante ak ecosystem Apple. Koom-koom yiy ñëw ci wàllu lëkkaloo ofisel wala gëna xéewale aparey yi mën nañu am njeexital yu am solo ci ni ñuy jëfandikoo IA ci platform iOS ak macOS.

Gid ak quiz yu ci méngoo

Model IA leeral nañu koTaggat ci IAAgent 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?