Cloudflare has released Clef and Clef-flash, the first models trained by its Workers AI team. They are decision models, not chatbots. Each reads an input state and a schema of typed questions. It returns a probability for every allowed answer, with no free-form text. Both are open-weight under Apache 2.0 and compatible with TypeSafe AI’s Jev API.
Is it deployable? Yes, Both models run today on Workers AI, and the weights are on Hugging Face for self-hosting.
What a Decision Model Does
An LLM generates tokens one at a time, and its output still needs parsing. A decision model only answers a fixed set of questions about an input. Clef supports 3 question types:
noul: yes/no, returns the probability of yes.choice: picks 1 named option, with per-option probabilities and a confidence value.score: rates against an ordered rubric, returning a probability-weighted score.
On Workers AI, 1 request carries up to 64 questions and up to 4 images.
TypeSafe AI launched Jev, its first ‘System One’ model, on September 15, 2026. Open alternatives like Kev-9B and Laya followed. Clef uses the same System One API. Switching from Jev means changing the endpoint and model name.
How Clef Works
Clef is post-trained from Qwen3.8-27B, and Clef-flash from Qwen3.5-9B. Both keep the backbone’s vision encoder.
Inference has 2 stages. The backbone first runs a single prefill-only pass over the state and questions. A small transformer, the joint schema head, then reads the final hidden states. It routes evidence to each question, lets fields cross-attend, and scores all options jointly. A per-question softmax turns logits into probabilities.
Training froze both backbones and jointly optimized the routing head with rank-256 low-rank adapters. The loss pairs label-smoothed cross-entropy with a Brier loss for calibration. A secondary objective, Reinforcement Learning for Calibrated Decisions (RLCD), gives partial credit to adjacent ordinal choices.
Interactive Explainer
‘});$(‘lanes’).innerHTML=h;setTimeout(rs,30)}
lanes();
R.querySelectorAll(‘.seg button’).forEach(function(b){b.onclick=function(){R.querySelectorAll(‘.seg button’).forEach(function(x){x.classList.remove(‘on’)});b.classList.add(‘on’);mode=b.dataset.m;lanes()}});
$(‘race’).onclick=function(){
lanes();var mx=Math.max.apply(null,L.map(function(l){return mode===’med’?l[1]:l[2]}));
setTimeout(function(){L.forEach(function(l,i){var v=mode===’med’?l[1]:l[2],d=v*4,f=$(‘lf’+i),t0=performance.now();
f.style.transition=’width ‘+d+’ms linear’;f.style.width=(v/mx*100)+’%’;
(function a(t){var p=Math.min(1,(t-t0)/d);$(‘lv’+i).textContent=(v*p).toFixed(1)+’ ms’;if(p’+b[0].split(‘ (‘)[0]+”});$(‘chips’).innerHTML=ch;
$(‘chips’).querySelectorAll(‘.chip’).forEach(function(c){c.onclick=function(){$(‘chips’).querySelectorAll(‘.chip’).forEach(function(x){x.classList.remove(‘on’)});c.classList.add(‘on’);bench(+c.dataset.i)}});
function bench(i){
var b=B[i],n=[‘Clef’,’Clef-flash’,’Jev’],v=[b[1],b[2],b[3]],mx=Math.max.apply(null,v),h=””;
n.forEach(function(nm,j){h+=’
‘});
$(‘bars’).innerHTML=h;
setTimeout(function(){n.forEach(function(nm,j){$(‘bb’+j).style.width=v[j]+’%’})},40);
var w=n[v.indexOf(mx)];
$(‘bcap’).innerHTML=’‘+b[0]+’: ‘+w+’ leads this row among these 3. ‘+(i===9?’DiffusionGemma Jev (85.35) tops PhishNChips overall.’:”)+(w===’Jev’?’ Jev also leads GPQA Diamond, MMLU-Pro and BBH on the full card.’:”);
setTimeout(rs,60);
}
window.addEventListener(‘load’,rs);window.addEventListener(‘resize’,rs);setTimeout(rs,300);
})();
Credit: Source link
