Meet Needle 2: An Open 45M-Parameter Tool-Calling Model That Ships as a 14MB Binary and Runs a Full Session in 28MB of RAM

Decode call</button>
<button class=”n2-btn” id=”n2-reset-g”>Reset</button>
</div>
<div class=”n2-term” id=”n2-term”><span class=”q”>&gt; dim the living room to 30</span></div>
<div class=”n2-legend”>
<span><i class=”n2-swatch” style=”background:var(–green)”></i>forced by grammar</span>
<span><i class=”n2-swatch” style=”background:var(–sand)”></i>value chosen by the model</span>
</div>
<div class=”n2-meters”>
<div class=”n2-meter”><span>Vocab rows scored · structural</span><b id=”n2-m1″>0%</b></div>
<div class=”n2-meter”><span>Structural bytes forced</span><b id=”n2-m2″>0</b></div>
</div>
</section>

<!– 03 RETRIEVAL –>
<section class=”n2-panel” role=”tabpanel” data-p=”2″>
<h3 class=”n2-h”>Above five tools, only five ever reach the model</h3>
<p class=”n2-p”>Every schema is embedded once by a built-in contrastive head. Each turn scores the query and
admits the top five; the grammar is then rebuilt over that subset. An unselected tool is unreachable.</p>
<div class=”n2-controls” role=”group” aria-label=”Choose a query”>
<button class=”n2-btn is-on” data-q=”0″>”dim the living room”</button>
<button class=”n2-btn” data-q=”1″>”text Ada I’m late”</button>
<button class=”n2-btn” data-q=”2″>”set it to 21 and play jazz”</button>
</div>
<p class=”n2-lanehead”>Declared catalogue — 12 tools</p>
<div class=”n2-chips” id=”n2-chips”></div>
<p class=”n2-note”>Illustrative simulation of the documented top-5 mechanism, not live model output.</p>
</section>

<!– 04 CONFIDENCE –>
<section class=”n2-panel” role=”tabpanel” data-p=”3″>
<h3 class=”n2-h”>The failure mode is escalation, not wrong execution</h3>
<p class=”n2-p”><code>confidence</code> is the minimum of a calibrated post-hoc head and the decoding
probability of the call tokens. You pick the threshold. Off-topic input returns the empty call
<code>[]</code> before any score applies.</p>
<input class=”n2-slider” id=”n2-slider” type=”range” min=”50″ max=”99″ value=”80″ aria-label=”Confidence threshold”>
<p class=”n2-thresh”>Threshold: <b id=”n2-tv”>0.80</b> — act at or above, re-ask or route to a bigger model below.</p>
<div class=”n2-rows” id=”n2-rows”></div>
<p class=”n2-note”>0.94 is the value shown in the Cactus README response object; the rest are illustrative.</p>
</section>

<!– 05 MEMORY –>
<section class=”n2-panel” role=”tabpanel” data-p=”4″>
<h3 class=”n2-h”>A ceiling, not a curve</h3>
<p class=”n2-p”>Attention runs a 256-token sliding window, and the system turn plus tool declarations are
pinned as KV sinks — so the one thing a tool-calling model must not forget cannot be evicted. Add turns
and watch what happens to each bar.</p>
<div class=”n2-controls”>
<button class=”n2-btn is-on” id=”n2-turn”>+ Add a turn</button>
<button class=”n2-btn” id=”n2-reset-m”>Reset conversation</button>
<span class=”n2-btn” style=”cursor:default” id=”n2-turns”>turns: 1</span>
</div>
<div class=”n2-memgrid”>
<div class=”n2-membar”>
<span class=”n2-bar-l” style=”color:var(–green)”>Needle 2</span>
<div class=”n2-memtrack”>
<div class=”n2-memfill pin” id=”n2-pin” style=”width:34%”></div>
<div class=”n2-memfill win” id=”n2-win” style=”left:34%;width:24%”></div>
</div>
<span class=”n2-bar-v” id=”n2-memv”>28.0 MB</span>
</div>
<div class=”n2-membar”>
<span class=”n2-bar-l”>Unbounded KV</span>
<div class=”n2-memtrack”><div class=”n2-memfill grow” id=”n2-grow” style=”width:22%”></div></div>
<span class=”n2-bar-v” id=”n2-growv”>rising</span>
</div>
</div>
<p class=”n2-note”>Striped block = pinned tools and system facts (KV sinks). Solid block = the 256-token window.</p>
<p class=”n2-overflow” id=”n2-of”></p>
</section>

<div class=”n2-foot”>
<span class=”n2-brand”>Explainer by <b>Marktechpost</b></span>
<span>
<a href=”https://github.com/cactus-compute/needle” target=”_blank” rel=”noopener”>GitHub</a> ·
<a href=”https://huggingface.co/Cactus-Compute/needle2″ target=”_blank” rel=”noopener”>Hugging Face</a> ·
<a href=”https://cactuscompute.com/needle” target=”_blank” rel=”noopener”>Cactus</a> ·
<a href=”https://arxiv.org/abs/2607.18363″ target=”_blank” rel=”noopener”>arXiv:2607.18363</a>
</span>
</div>
</div>

<script>
(function(){
var root=document.getElementById(‘n2-root’);
var reduce=window.matchMedia&&window.matchMedia(‘(prefers-reduced-motion:reduce)’).matches;

/* ———- resize ———- */
function post(){
try{ parent.postMessage({type:’n2-height’, height: root.offsetHeight + 40}, ‘*’); }catch(e){}
}
var ro=window.ResizeObserver?new ResizeObserver(post):null; if(ro) ro.observe(root);
window.addEventListener(‘load’,post); setTimeout(post,120); setTimeout(post,600);

/* ———- tabs ———- */
var tabs=[].slice.call(root.querySelectorAll(‘.n2-tab’));
var panels=[].slice.call(root.querySelectorAll(‘.n2-panel’));
tabs.forEach(function(t){
t.addEventListener(‘click’,function(){
tabs.forEach(function(x){x.setAttribute(‘aria-selected’,’false’)});
t.setAttribute(‘aria-selected’,’true’);
panels.forEach(function(p){p.classList.remove(‘is-on’)});
panels[+t.dataset.t].classList.add(‘is-on’);
if(t.dataset.t===’0′) drawBars(metric);
post();
});
});

/* ———- 01 frontier ———- */
var DATA={
params:{unit:’M params’, note:’Total parameters. Apple FM is approximately 3B. Source: Cactus compute-per-token table.’,
rows:[[‘Needle 2′,45,’45M’],[‘LFM2.5 230M’,230,’230M’],[‘FunctionGemma 270M’,270,’270M’],[‘Apple FM’,3000,’~3B’]]},
flops:{unit:’MFLOPs/token’, note:’2 FLOPs per multiply-accumulate over matmul-active parameters, embeddings tied, attention terms excluded. Needle keeps 8M parameters as gathered engram memory, costing no arithmetic.’,
rows:[[‘Needle 2′,70,’70’],[‘LFM2.5 230M’,460,’460′],[‘FunctionGemma 270M’,540,’540′],[‘Apple FM’,6000,’~6,000′]]},
acc:{unit:’%’, note:’Mobile Actions eval split, 961 rows, ordered strict exact match. Needle 2 at CQ2-bit with retrieval on; LFM2.5 and FunctionGemma at f16 under vLLM; Apple FM on-device.’,
rows:[[‘Needle 2′,63.7,’63.7’],[‘LFM2.5 230M’,69.1,’69.1′],[‘FunctionGemma 270M’,64.0,’64.0′],[‘Apple FM’,57.6,’57.6′]]}
};
var metric=’params’, barsEl=document.getElementById(‘n2-bars’), noteEl=document.getElementById(‘n2-barnote’);
function drawBars(m){
var d=DATA[m], max=Math.max.apply(null,d.rows.map(function(r){return r[1]}));
barsEl.innerHTML=d.rows.map(function(r){
return ‘<div class=”n2-bar’+(r[0]===’Needle 2′?’ is-hero’:”)+'”>’+
‘<span class=”n2-bar-l”>’+r[0]+'</span>’+
‘<div class=”n2-bar-t”><div class=”n2-bar-f”></div></div>’+
‘<span class=”n2-bar-v”>’+r[2]+'</span></div>’;
}).join(”);
noteEl.textContent=d.note;
var fills=barsEl.querySelectorAll(‘.n2-bar-f’);
d.rows.forEach(function(r,i){
var w=Math.max(2,(r[1]/max)*100);
if(reduce){fills[i].style.width=w+’%’;}
else setTimeout(function(){fills[i].style.width=w+’%’;},60*i+40);
});
post();
}
root.querySelectorAll(‘[data-m]’).forEach(function(b){
b.addEventListener(‘click’,function(){
root.querySelectorAll(‘[data-m]’).forEach(function(x){x.classList.remove(‘is-on’)});
b.classList.add(‘is-on’); metric=b.dataset.m; drawBars(metric);
});
});
drawBars(‘params’);

/* ———- 02 grammar ———- */
var CALL=[
[‘{“name”:”‘,1],[‘set_lights’,0],[‘”,”arguments”:{“room”:”‘,1],[‘living room’,0],
[‘”,”on”:’,1],[‘true’,0],[‘,”brightness”:’,1],[’30’,0],[‘}}’,1] ];
var term=document.getElementById(‘n2-term’), m1=document.getElementById(‘n2-m1’), m2=document.getElementById(‘n2-m2’);
var timer=null;
function resetG(){
if(timer){clearInterval(timer);timer=null;}
term.innerHTML='<span class=”q”>&gt; dim the living room to 30</span>’;
m1.textContent=’0%’; m2.textContent=’0′; post();
}
document.getElementById(‘n2-reset-g’).addEventListener(‘click’,resetG);
document.getElementById(‘n2-decode’).addEventListener(‘click’,function(){
if(timer){clearInterval(timer);timer=null;}
var flat=[]; CALL.forEach(function(seg){
for(var i=0;i<seg[0].length;i++) flat.push([seg[0][i],seg[1]]);
});
term.innerHTML='<span class=”q”>&gt; dim the living room to 30</span><br>’;
var out=document.createElement(‘span’); term.appendChild(out);
var caret=document.createElement(‘span’); caret.className=’n2-caret’; caret.textContent=’_’; term.appendChild(caret);
var i=0, forced=0;
function step(){
if(i>=flat.length){ if(timer){clearInterval(timer);timer=null;} caret.remove();
m1.textContent=’2%’; m2.textContent=forced; post(); return; }
var ch=flat[i][0], isF=flat[i][1];
var s=document.createElement(‘span’); s.className=isF?’n2-forced’:’n2-free’;
s.textContent=ch; out.appendChild(s);
if(isF) forced++;
var pct=Math.round(100-(forced/(i+1))*98);
m1.textContent=pct+’%’; m2.textContent=forced;
i++;
}
if(reduce){ while(i<flat.length) step(); }
else timer=setInterval(step,26);
});

/* ———- 03 retrieval ———- */
var TOOLS=[‘set_lights’,’play_music’,’send_message’,’set_thermostat’,’get_weather’,’set_timer’,
‘call_contact’,’navigate_to’,’take_photo’,’lock_door’,’add_calendar_event’,’control_tv’];
var PICKS=[[0,3,11,1,5],[2,6,10,0,4],[3,1,0,5,4]];
var chipsEl=document.getElementById(‘n2-chips’);
chipsEl.innerHTML=TOOLS.map(function(t,i){return ‘<span class=”n2-chip” data-i=”‘+i+'”>’+t+'</span>’}).join(”);
function retrieve(qi){
var sel=PICKS[qi];
[].slice.call(chipsEl.children).forEach(function(c,i){
c.classList.remove(‘is-sel’,’is-out’);
var d=sel.indexOf(i)>-1?0:180;
setTimeout(function(){ c.classList.add(sel.indexOf(i)>-1?’is-sel’:’is-out’); }, reduce?0:d);
});
}
root.querySelectorAll(‘[data-q]’).forEach(function(b){
b.addEventListener(‘click’,function(){
root.querySelectorAll(‘[data-q]’).forEach(function(x){x.classList.remove(‘is-on’)});
b.classList.add(‘is-on’); retrieve(+b.dataset.q);
});
});
retrieve(0);

/* ———- 04 confidence ———- */
var TURNS=[
[‘dim the living room to 30’,0.94,false],
[‘make it warmer in here’,0.81,false],
[‘put the thing on the other thing’,0.58,false],
[‘text Ada that I am running late’,0.88,false],
[‘what is the capital of Peru?’,null,true] ];
var rowsEl=document.getElementById(‘n2-rows’), slider=document.getElementById(‘n2-slider’), tv=document.getElementById(‘n2-tv’);
function renderRows(){
var th=(+slider.value)/100; tv.textContent=th.toFixed(2);
rowsEl.innerHTML=TURNS.map(function(t){
if(t[2]) return ‘<div class=”n2-row ref”><span class=”n2-row-q”>’+t[0]+
‘</span><span class=”n2-row-c”>[ ]</span><span class=”n2-verdict v-ref”>empty call</span></div>’;
var ok=t[1]>=th;
return ‘<div class=”n2-row ‘+(ok?’act’:’esc’)+'”><span class=”n2-row-q”>’+t[0]+
‘</span><span class=”n2-row-c”>’+t[1].toFixed(2)+'</span><span class=”n2-verdict ‘+
(ok?’v-act”>execute’:’v-esc”>escalate’)+'</span></div>’;
}).join(”);
post();
}
slider.addEventListener(‘input’,renderRows); renderRows();

/* ———- 05 memory ———- */
var turns=1, gv=document.getElementById(‘n2-gv’), gf=document.getElementById(‘n2-gf’);
var win=document.getElementById(‘n2-win’), grow=document.getElementById(‘n2-grow’);
var memv=document.getElementById(‘n2-memv’), growv=document.getElementById(‘n2-growv’);
var ofEl=document.getElementById(‘n2-of’), turnsEl=document.getElementById(‘n2-turns’);
function renderMem(){
turnsEl.textContent=’turns: ‘+turns;
var w=Math.min(24, 12+turns*2.5);
win.style.width=w+’%’;
memv.textContent=’28.0 MB’;
gf.style.width=’78%’; gv.innerHTML='<b>28.0 MB</b> / hard ceiling 28 MB’;
var g=Math.min(100, 18+turns*7);
grow.style.width=g+’%’;
growv.textContent=(Math.round((18+turns*7)*0.62*10)/10)+’ MB’;
ofEl.textContent = g>=78 ? ‘Unbounded KV has crossed the 28MB ceiling. Needle 2 has not moved.’ : ”;
post();
}
document.getElementById(‘n2-turn’).addEventListener(‘click’,function(){turns++;renderMem();});
document.getElementById(‘n2-reset-m’).addEventListener(‘click’,function(){turns=1;renderMem();});
renderMem();
})();
</script>
</body>
</html>
“>