https://jigsawstack.ai/_next/static/chunks/pages/blog/what-even-is…-language-model-now--ai-e69d00354ea372f4.js

1(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[5842,5211,7329],{60632:function(e,n,t){(window.__NEXT_P=window.__NEXT_P||[]).push(["/blog/what-even-is-a-small-language-model-now--ai",function(){return t(73447)}])},91873:function(e,n,t){"use strict";var i=t(11121),s=t(14295),l=t(29966),a=t(48541),r=t(25608),o=t(67992),d=t(63202),c=t(7275),h=t(671),u=t(90751),m=t(49855);let g="".concat(m.v2,"/jigsaw_icon.png");n.Z=(0,d.memo)(e=>{var n,t;let{slug:d,children:p}=e,x=h.find(e=>e.slug===d),y=(null==x?void 0:x.date_modified)||(null==x?void 0:x.date),j="".concat(m.v2,"/blog/").concat(d),f=(null==x?void 0:x.og_image)||(0,u.Q)(m.v2,"Blog",x);return f.startsWith("http")||(f="".concat(m.v2).concat(f)),(0,i.jsxs)(i.Fragment,{children:[(0,i.jsx)(c.Z,{title:null==x?void 0:x.title,description:null==x?void 0:x.description,canonical:j,openGraph:{type:"article",article:{publishedTime:null==x?void 0:x.date,modifiedTime:y,authors:null==x?void 0:null===(n=x.author)||void 0===n?void 0:n.map(e=>e.name),tags:null==x?void 0:x.tags},images:[{url:f,alt:(null==x?void 0:x.title)||""}]},additionalMetaTags:(null==x?void 0:null===(t=x.tags)||void 0===t?void 0:t.length)?[{name:"keywords",content:x.tags.join(", ")}]:void 0}),x&&(0,i.jsx)(o.dX,{type:"BlogPosting",url:j,title:x.title,images:[f],datePublished:x.date,dateModified:y,authorName:(x.author||[]).map(e=>e.name),description:x.description,publisherName:"JigsawStack",publisherL

1ogo:g}),x&&(0,i.jsx)(o.gR,{itemListElements:[{position:1,name:"Home",item:m.v2},{position:2,name:"Blog",item:"".concat(m.v2,"/blog")},{position:3,name:x.title,item:j}]}),(0,i.jsx)(r.Z,{w:["100%","3xl"],children:(0,i.jsxs)(s.k,{mt:["1rem","2rem"],mb:"5rem",flexDir:"column",w:"100%",children:[(null==x?void 0:x.og_image)&&(0,i.jsx)(l.x,{overflow:"hidden",borderRadius:"lg",mb:"1rem",w:"100%",h:["3xs","sm"],bgColor:"bg.600",children:(0,i.jsx)(a.Z,{src:x.og_image,alt:x.title,w:"100%",h:"100%",objectFit:"cover"})}),p]})})]})})},91322:function(e,n,t){"use strict";var i=t(11121),s=t(29966),l=t(14295),a=t(958),r=t(42320),o=t(69035),d=t(63202),c=t(86430),h=t(80620);n.Z=(0,d.memo)(e=>{let{code:n,language:t="auto",codeWhiteSpace:u="pre",codeStyle:m,showCopy:g=!0,...p}=e,[x,y]=(0,d.useState)(""),[j,f]=(0,d.useState)(!0),w=(0,d.useCallback)(async()=>{f(!0);try{let e=await (0,h.CZ)(n,{lang:"curl"==t?"bash":t,theme:"slack-dark"});y(e)}finally{f(!1)}},[t,n]);return(0,d.useLayoutEffect)(()=>{w()},[w,n,t]),(0,i.jsxs)(a.Z,{width:"100%",pos:"relative",...p,children:[(0,i.jsx)(s.x,{width:"100%",pl:"1rem",pr:g?"3rem":"1rem",py:"1rem",textAlign:"left",fontSize:"sm",overflowY:"scroll",overflowX:"scroll",sx:{"& .shiki":{backgroundColor:"transparent !important"},"& .shiki > code":{whiteSpace:u}},dangerouslySetInnerHTML:{__html:j||!x?"...":x},backgroundColor:"transparent",...m}),g&&(0,i.jsx)(l.k,{borderRadius:"md",mr:"0.5%",right:0,p:"0.5rem",_hover:{bgColor:"bg.400"},cursor:"pointer",position:"absolute",onClick:e=>{e.preventDefault(),navigator.clipboard.writeText(n),(0,c.Am)("Copied to clipboard")},children:(0,i.jsx)(r.Z,{as:o.Z})})]})})},90751:function(e,n,t){"use strict";t.d(n,{Q:function(){return l}});var i=t(93371),s=t.n(i);let l=(e,n,t)=>{var i,l;let a=new URLSearchParams({title:(null==t?void 0:t.title)||"",label:n}),r=null==t?void 0:null===(l=t.author)||void 0===l?void 0:null===(i=l[0])||void 0===i?void 0:i.name;return r&&a.set("author",r),(null==t?void 0:t.date)&&a.set("date",s()(t.date).format("DD MMM YYYY")),"".concat(e,"/api/og?").concat(a.toString())}},73447:function(e,n,t){"use strict";t.r(n),t.d(n,{default:function(){return o}});var i=t(11121),s=t(10562),l=t(91873);let a=function(e){let{children:n}=e;return(0,i.jsx)(l.Z,{slug:"what-even-is-a-small-language-model-now--ai",children:n})};function r(e){let n={a:"a",em:"em",h1:"h1",h2:"h2",h3:"h3",li:"li",p:"p",strong:"strong",ul:"ul",...(0,s.a)(),...e.components};return(0,i.jsxs)(i.Fragment,{children:[(0,i.jsx)(n.h1,{children:"What Even Is a Small Language Model Now?"}),"\n",(0,i.jsxs)(n.p,{children:['If you asked someone in 2018 what a "small model" was, they\'d probably say something with a few million parameters that ran on a Raspberry Pi or your phone. Fast-forward to today, and we\'re calling 30B parameter models "small"—because they ',(0,i.jsx)(n.em,{children:"only"})," need one GPU to run."]}),"\n",(0,i.jsx)(n.p,{children:'So yeah, the definition of "small" has changed.'}),"\n",(0,i.jsx)(n.h2,{children:"Small Used to Mean... Actually Small"}),"\n",(0,i.jsx)(n.p,{children:'Back in the early days of machine learning, a "small model" might\'ve been a decision tree or a basic neural net that could run on a laptop CPU. Think scikit-learn, not LLMs.'}),"\n",(0,i.jsx)(n.p,{children:"Then came transformers and large language models (LLMs). As these got bigger and better, anything not requiring a cluster of A100s suddenly started to feel... small by comparison."}),"\n",(0,i.jsxs)(n.p,{children:["Today, small is more about ",(0,i.jsx)(n.strong,{children:"how deployable"})," the model is, not just its size on paper."]}),"\n",(0,i.jsx)(n.h2,{children:"Types of Small Models (By 2025 Standards)"}),"\n",(0,i.jsx)(n.p,{children:"We now have two main flavors of small language models:"}),"\n",(0,i.jsx)(n.h3,{children:"1. Edge-Optimized Models"}),"\n",(0,i.jsx)(n.p,{children:"These are the kind of models you can run on mobile devices or edge hardware. They're optimized for speed, low memory, and offline use."}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Examples"}),": Phi-3-mini (3.8B), Gemma 2B, TinyLlama (1.1B)"]}

1),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Use cases"}),": voice assistants, translation on phones, offline summarization, chatbots embedded in apps"]}),"\n"]}),"\n",(0,i.jsx)(n.h3,{children:"2. GPU-Friendly Models"}),"\n",(0,i.jsxs)(n.p,{children:["These still require a GPU, but just ",(0,i.jsx)(n.strong,{children:"one GPU"}),'—not a whole rack. In this category, even 30B or 70B models can qualify as "small".']}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Examples"}),": Meta Llama 3 70B (quantized), MPT-30B"]}),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Use cases"}),": internal RAG pipelines, chatbot endpoints, summarizers, code assistants"]}),"\n"]}),"\n",(0,i.jsx)(n.p,{children:"The fact that you can now run a 70B model on a single 4090 and get decent throughput? That would've been science fiction a few years ago."}),"\n",(0,i.jsx)(n.h2,{children:"Specialization: The Real Power Move"}),"\n",(0,i.jsxs)(n.p,{children:["One big strength of small models is that they don't need to do everything. Unlike GPT-4 or Claude that try to be general-purpose brains, small models are often ",(0,i.jsx)(n.strong,{children:"narrow and optimized"}),"."]}),"\n",(0,i.jsx)(n.p,{children:"That gives them a few key advantages:"}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"They stay lean"})," — no need to carry weights for tasks they’ll never do."]}),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"They’re more accurate in-domain"})," — a small legal model will outperform a general-purpose LLM on legal docs."]}),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"They’re easier to fine-tune"})," — less data, faster iteration."]}),"\n"]}),"\n",(0,i.jsx)(n.p,{children:"Small models shine when you know what you want. Think: summarizing medical records, identifying security vulnerabilities, parsing invoices—stuff that doesn't need general reasoning across the internet."}),"\n",(0,i.jsx)(n.h2,{children:"30B+ Models: Still Small?"}),"\n",(0,i.jsx)(n.p,{children:'Sounds weird, but yes. The bar for what’s considered "small" keeps shifting.'}),"\n",(0,i.jsx)(n.p,{children:"With the right quantization and engineering, even a 70B model can run comfortably on a high-end consumer GPU:"}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Llama 3.1 70B"})," can be shrunk from 140GB (FP16) to 21GB (2-bit), running on a single 24GB VRAM card."]}),"\n",(0,i.jsx)(n.li,{children:"Throughput? ~60 tokens/sec — totally usable for many production workloads."}),"\n"]}),"\n",(0,i.jsx)(n.p,{children:'So now we talk about models being "small" if they’re:'}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsx)(n.li,{children:"Deployable without distributed inference"}),"\n",(0,i.jsx)(n.li,{children:"Runnable on one GPU (especially consumer-grade)"}),"\n",(0,i.jsx)(n.li,{children:"Tunable without a lab full of TPUs"}),"\n"]}),"\n",(0,i.jsxs)(n.p,{children:["It’s less about size, more about ",(0,i.jsx)(n.em,{children:"practicality"}),"."]}),"\n",(0,i.jsx)(n.h2,{children:"Everyday Small Models: The Unsung Heroes"}),"\n",(0,i.jsx)(n.p,{children:"Not all small models are new. Some of the most widely used models today have been around for years, quietly powering everyday tools we rely on."}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Google Translate"}),": Since 2006, it's been translating billions of words daily. In 2016, Google switched to a neural machine translation system, GNMT, which uses an encoder-decoder architecture with long short-term memory (LSTM) layers and attention mechanisms. This system, with over 160 million parameters, significantly improved translation fluency and accuracy."]}),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"AWS Textract"}),": This service extracts text and data from scanned documents. It's been a staple in automating document processing workflows, handling everything from invoices to medical records."]}),"\n"]}),"\n",(0,i.jsx)(n.p,{children:"These models may not be cutting-edge by today's standards, but they've been instrumental in shaping the AI landscape and continue to serve millions daily."}),"\n",(0,i.jsx)(n.h2,{children:"Why This Matters"}),"\n",(0,i.jsx)(n.p,{children:"Small models are becoming a huge deal:"}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Startups"})," can deploy LLMs without spending six figures on infra."]}),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Developers"})," can run local models for privacy-focused apps."]}),"\n",(0,i.jsxs)(n.li,{children:[(0,i.jsx)(n.strong,{children:"Enterprises"})," can fine-tune task-specific LLMs without massive overhead."]}),"\n"]}),"\n",(0,i.jsx)(n.p,{children:'And when a "small model" can hold its own against GPT-3.5 in benchmarks? The game has officially changed.'}),"\n",(0,i.jsx)(n.h2,{children:"TL;DR"}),"\n",(0,i.jsxs)(n.ul,{children:["\n",(0,i.jsx)(n.li,{children:'Small models used to mean tiny. Now they mean "runs without drama."'}),"\n",(0,i.jsx)(n.li,{children:"You’ve got edge models, GPU-ready models, and everything in between."}),"\n",(0,i.jsx)(n.li,{children:"Specialization is where small models shine."}),"\n",(0,i.jsx)(n.li,{children:"30B and 70B models can be small—if they’re optimized well."}),"\n",(0,i.jsx)(n.li,{children:"Practicality > parameter count."}),"\n"]}),"\n",(0,i.jsx)(n.p,{children:"In a world chasing ever-bigger models, small ones are quietly doing more with less—and that's exactly what makes them powerful."}),"\n",(0,i.jsx)(n.h2,{children:"\uD83D\uDC65 Join the JigsawStack Community"}),"\n",(0,i.jsxs)(n.p,{children:["Have questions or want to show off what you’ve built? Join the JigsawStack developer community on ",(0,i.jsx)(n.a,{href:"https://discord.com/invite/dj8fMBpnqd",children:"Discord"})," and ",(0,i.jsx)(n.a,{href:"https://x.com/jigsawstack",children:"X/Twitter"}),". Let’s build something amazing together!"]})]})}function o(){let e=arguments.length>0&&void 0!==arguments[0]?arguments[0]:{};return(0,i.jsx)(a,{...e,children:(0,i.jsx)(r,{...e})})}},65211:function(e,n,t){"use strict";t.d(n,{W:function(){return s}});var i=t(63202);function s(e){return i.Children.toArray(e).filter(e=>(0,i.isValidElement)(e))}},98126:function(e,n,t){"use strict";t.d(n,{DE:function(){return y},GS:function(){return g},HC:function(){return x},QI:function(){return p},aV:function(){return m}});var i=t(11121),s=t(8758),l=t(52125),a=t(65211),r=t(19771),o=t(61844),d=t(80120),c=t(30474);let[h,u]=(0,l.k)({name:"ListStylesContext",errorMessage:"useListStyles returned is 'undefined'. Seems you forgot to wrap the components in \"<List />\" "}),m=(0,o.G)(function(e,n){let t=(0,d.j)("List",e),{children:l,styleType:r="none",stylePosition:o,spacing:u,...m}=(0,s.L)(e),g=(0,a.W)(l);return(0,i.jsx)(h,{value:t,children:(0,i.jsx)(c.m.ul,{ref:n,listStyleType:r,listStylePosition:o,role:"list",__css:{...t.container,...u?{"& > *:not(style) ~ *:not(style)":{mt:u}}:{}},...m,children:g})})});m.displayName="List";let g=(0,o.G)((e,n)=>{let{as:t,...s}=e;return(0,i.jsx)(m,{ref:n,as:"ol",styleType:"decimal",marginStart:"1em",...s})});g.displayName="OrderedList";let p=(0,o.G)(function(e,n){let{as:t,...s}=e;return(0,i.jsx)(m,{ref:n,as:"ul",styleType:"initial",marginStart:"1em",...s})});p.displayName="UnorderedList";let x=(0,o.G)(function(e,n){let t=u();return(0,i.jsx)(c.m.li,{ref:n,...e,__css:t.item})});x.displayName="ListItem";let y=(0,o.G)(function(e,n){let t=u();return(0,i.jsx)(r.J,{ref:n,role:"presentation",...e,__css:t.icon})});y.displayName="ListIcon"}},function(e){e.O(0,[7992,1610,5245,1038,562,8538,2888,9774,179],function(){return e(e.s=60632)}),_N_E=e.O()}]);