NVIDIA Releases Alpamayo 2 Edge: A 2B Reasoning VLA for DRIVE Thor

0


NVIDIA has released Alpamayo 2 Edge, a compact reasoning vision-language-action (VLA) model for autonomous driving. Alpamayo 2 Edge is a 2B-parameter model that plans trajectories and explains them inside the vehicle. It targets NVIDIA DRIVE AGX Thor, not a datacenter GPU. The goal: reasoning-based planning within onboard compute and latency budgets. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.

TL;DR

  • Size: 2B-class. Card breakdown: 2.45B backbone + 0.2B action expert (HF metadata says 3B). Context length: not disclosed (measured prompt: 3,236 tokens).
  • Runs on: NVIDIA DRIVE AGX Thor with the FP8 + TensorRT Edge-LLM recipe. BF16 PyTorch needs 16 GB+ GPU memory, validated on 1x H100 80GB.
  • Performance: Strong closed-loop driver for its size; mid-pack on reasoning QA.
  • Best: 15.15 km per at-fault incident in AlpaSim, about 60x Qwen-Drive-1.0 4B (0.25 km).
  • Worst: nuReasoning numeric score of 48.67; LingoQA 75.2 trails Alpamayo 2 Super (79.2).
  • Bottom line: Best: reasoning-based planning at 160 to 190 ms on in-car hardware. Worst: only H100 validated, and real-time Thor speed needs the FP8 recipe.

What is Alpamayo 2 Edge?

Alpamayo 2 Edge is an open NVIDIA VLA model that turns surround-camera video into a 6.4-second plan. It is the in-car member of the Alpamayo family, next to the 34B cloud model Alpamayo 2 Super.

Per the model card, the 2.45B backbone pairs a 1.96B language model with a 0.41B SigLIP2 encoder. A 0.2B diffusion action expert sits on top. The backbone comes from NVIDIA Cosmos 3 Edge and is Nemotron-based. NVIDIA markets the model as 2B, while Hugging Face metadata shows 3B.

What goes in and what comes out?

For trajectory prediction, the model reads 6 cameras and 22 frames per observation. The 2 front cameras give 5 frames each, the other 4 give 3 each. It also takes a navigation instruction and 20 historical ego poses.

Output is 64 waypoints from 0.1 to 6.4 seconds. The action expert predicts acceleration and curvature, then a unicycle model integrates them. The text output is a Chain-of-Causation (CoC) reasoning trace. The model also answers free-form questions about a driving scene.

How does adaptive reasoning work?

During routine driving, the model emits special <NOMINAL> tokens instead of a full reasoning trace. It writes CoC reasoning only when a scene needs it. Fewer tokens means less compute per planning step.

How was Alpamayo 2 Edge trained?

NVIDIA team lists these ingredients:

  • Offline distillation from CoC autolabels generated by Alpamayo 2 Super.
  • About 115,000 hours of multi-camera driving video and about 3.7M CoC traces.
  • Closed-loop RL post-training in AlpaGym, learning from its own actions in simulation.
  • Multi-resolution inputs for cheaper encoding of surround views.

How fast does it run on DRIVE AGX Thor?

Hugging Face weights ship in BF16 and run as a PyTorch model. Real-time use on Thor needs the Alpamayo Developer Hub recipe. It stages the checkpoint and the DFlash2 drafter, then calibrates FP8 scales. Next it exports ONNX graphs and builds TensorRT engines with TensorRT Edge-LLM plugins.

NVIDIA reports 160 ms to 190 ms pixel-to-trajectory latency, depending on how much the model reasons. The reference setup is DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9. In NVIDIA’s demo video, PyTorch BF16 plans at 0.76 Hz. The optimized FP8 Thor pipeline plans at 13.02 Hz in AlpaSim. Partner Embedl says it reached 10 Hz on AGX Thor in under 48 hours.

On a workstation GPU, the card recommends at least 16 GB of memory. NVIDIA measured 12.06 GiB peak process memory on 1 H100 with 6 trajectory candidates. Other GPU architectures are not yet validated.

How does Alpamayo 2 Edge score on benchmarks?

NVIDIA compares Edge with Alibaba’s 4B Qwen-Drive-1.0. In AlpaSim closed-loop driving, Edge covers 15.15 km between at-fault incidents versus 0.25 km. That is about 60x. Open-loop minADE is 0.53 m versus 1.68 m. Both are NVIDIA’s own runs.

The model card adds:

  • AlpaSim: 15.15 ± 0.04 km per at-fault incident over 875 NuRec scenarios, 3 rollouts.
  • Open-loop minADE_6 at 5 s: 0.52 m on 1,392 PhysicalAI-AV samples.
  • LingoQA (Lingo-Judge): 75.2.
  • SURDS: 86.50, KITScenes scene: 89.46, nuReasoning numeric: 48.67.

<iframe id="mtp-a2e-frame" title="Alpamayo 2 Edge interactive explainer" style="width:100%;max-width:880px;height:600px;border:0;display:block;margin:0 auto;overflow:hidden" scrolling="no" loading="lazy" srcdoc="<!DOCTYPE html><html><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><style>body{margin:0;background:transparent}</style></head><body><style>
#mtp-a2e{max-width:880px;margin:0 auto;background:#0c0c0c!important;color:#e8e8e8!important;border:1px solid #262626!important;border-radius:14px;font-family:Inter,system-ui,-apple-system,Segoe UI,Roboto,sans-serif;overflow:hidden;box-sizing:border-box}
#mtp-a2e *{box-sizing:border-box}
#mtp-a2e hr,#mtp-a2e p:empty,#mtp-a2e del,#mtp-a2e s{display:none!important}
#mtp-a2e .hd{padding:20px 24px 14px;border-bottom:1px solid #222!important;background:linear-gradient(135deg,#111 0%,#0c0c0c 60%,#132000 100%)!important}
#mtp-a2e .tag{display:inline-block;font-size:11px;letter-spacing:.08em;text-transform:uppercase;color:#0c0c0c!important;background:#76B900!important;padding:3px 8px;border-radius:4px;font-weight:700}
#mtp-a2e h3{margin:10px 0 4px;font-size:22px;line-height:1.25;color:#fff!important;font-weight:800}
#mtp-a2e .sub{margin:0;color:#a8a8a8!important;font-size:14px}
#mtp-a2e .tabs{display:flex;gap:4px;padding:10px 12px;background:#111!important;border-bottom:1px solid #222!important;flex-wrap:wrap}
#mtp-a2e .tb{background:transparent!important;color:#bdbdbd!important;border:1px solid #2a2a2a!important;border-radius:8px;padding:8px 12px;font-size:13px;font-weight:600;cursor:pointer;transition:all .2s}
#mtp-a2e .tb:hover{border-color:#76B900!important;color:#fff!important}
#mtp-a2e .tb.on{background:#76B900!important;color:#0c0c0c!important;border-color:#76B900!important}
#mtp-a2e .pn{display:none;padding:20px 24px 22px;animation:a2eIn .35s ease}
#mtp-a2e .pn.on{display:block}
@keyframes a2eIn{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}}
#mtp-a2e .grid{display:grid;grid-template-columns:repeat(4,1fr);gap:10px}
#mtp-a2e .st{background:#151515!important;border:1px solid #262626!important;border-radius:10px;padding:14px;cursor:default;transition:border-color .2s,transform .2s}
#mtp-a2e .st:hover{border-color:#76B900!important;transform:translateY(-2px)}
#mtp-a2e .st b{display:block;font-size:24px;color:#76B900!important;font-weight:800}
#mtp-a2e .st span{font-size:12px;color:#a8a8a8!important;line-height:1.35;display:block;margin-top:4px}
#mtp-a2e .note{font-size:12px;color:#8c8c8c!important;margin:12px 0 0;line-height:1.5}
#mtp-a2e .txt{font-size:14px;line-height:1.6;color:#d6d6d6!important;margin:0 0 12px}
#mtp-a2e .flow{display:flex;align-items:stretch;gap:8px;margin:6px 0 14px}
#mtp-a2e .nd{flex:1;background:#151515!important;border:1px solid #2a2a2a!important;border-radius:10px;padding:12px 10px;text-align:center;font-size:12.5px;color:#cfcfcf!important;transition:all .3s;position:relative}
#mtp-a2e .nd strong{display:block;color:#fff!important;font-size:13px;margin-bottom:4px}
#mtp-a2e .nd.lit{border-color:#76B900!important;background:#142000!important;box-shadow:0 0 0 1px #76B900,0 0 18px rgba(118,185,0,.35)}
#mtp-a2e .ar{align-self:center;color:#76B900!important;font-weight:800}
#mtp-a2e .ctl{display:flex;gap:8px;align-items:center;flex-wrap:wrap}
#mtp-a2e .btn{background:#1b1b1b!important;color:#fff!important;border:1px solid #76B900!important;border-radius:8px;padding:8px 14px;font-size:13px;font-weight:600;cursor:pointer}
#mtp-a2e .btn.act{background:#76B900!important;color:#0c0c0c!important}
#mtp-a2e .btn:hover{background:#76B900!important;color:#0c0c0c!important}
#mtp-a2e .desc{background:#111!important;border-left:3px solid #76B900!important;padding:12px 14px;border-radius:6px;font-size:13.5px;line-height:1.55;color:#ddd!important;min-height:64px;margin-top:12px}
#mtp-a2e .road{background:#111!important;border:1px solid #262626!important;border-radius:10px;padding:14px;margin:10px 0}
#mtp-a2e .stream{font-family:ui-monospace,SFMono-Regular,Menlo,monospace!important;font-size:12.5px;line-height:1.6;min-height:84px;color:#cfcfcf!important;background:#0a0a0a!important;border:1px solid #222!important;border-radius:8px;padding:10px 12px;white-space:pre-wrap}
#mtp-a2e .nom{color:#0c0c0c!important;background:#76B900!important;padding:1px 6px;border-radius:4px;font-weight:700}
#mtp-a2e .rs{color:#ffd166!important}
#mtp-a2e .meter{height:8px;background:#1d1d1d!important;border-radius:6px;overflow:hidden;margin-top:8px}
#mtp-a2e .meter i{display:block;height:100%;background:#76B900!important;width:0;transition:width .8s ease}
#mtp-a2e .bars{display:flex;flex-direction:column;gap:12px;margin-top:6px}
#mtp-a2e .br .lb{display:flex;justify-content:space-between;font-size:12.5px;color:#cfcfcf!important;margin-bottom:5px}
#mtp-a2e .br .lb em{font-style:normal;color:#fff!important;font-weight:700}
#mtp-a2e .trk{height:22px;background:#181818!important;border-radius:6px;overflow:hidden}
#mtp-a2e .fil{height:100%;width:0;border-radius:6px;transition:width 1.1s cubic-bezier(.2,.8,.2,1)}
#mtp-a2e .g{background:#76B900!important}
#mtp-a2e .gr{background:#5a5a5a!important}
#mtp-a2e .steps{list-style:none;padding:0;margin:0;counter-reset:s}
#mtp-a2e .steps li{counter-increment:s;padding:10px 12px 10px 44px;position:relative;border:1px solid #232323!important;border-radius:8px;margin-bottom:6px;font-size:13.5px;color:#d0d0d0!important;background:#131313!important;cursor:pointer;transition:all .2s}
#mtp-a2e .steps li:before{content:counter(s);position:absolute;left:12px;top:9px;width:22px;height:22px;border-radius:50%;background:#262626!important;color:#76B900!important;font-weight:800;font-size:12px;display:flex;align-items:center;justify-content:center}
#mtp-a2e .steps li.on{border-color:#76B900!important;background:#142000!important}
#mtp-a2e .steps li .x{display:none;color:#a8a8a8!important;font-size:12.5px;margin-top:5px}
#mtp-a2e .steps li.on .x{display:block}
#mtp-a2e .seg{display:flex;gap:6px;margin-bottom:12px;flex-wrap:wrap}
#mtp-a2e code{background:#1a1a1a!important;color:#76B900!important;padding:1px 5px;border-radius:4px;font-family:ui-monospace,Menlo,monospace!important;font-size:12px;border:0!important}
#mtp-a2e .ft{display:flex;justify-content:space-between;gap:10px;flex-wrap:wrap;padding:10px 24px;border-top:1px solid #222!important;background:#0f0f0f!important;font-size:11.5px;color:#8a8a8a!important}
#mtp-a2e .ft b{color:#76B900!important}
@media (max-width:640px){
#mtp-a2e .grid{grid-template-columns:repeat(2,1fr)}
#mtp-a2e .flow{flex-direction:column}
#mtp-a2e .ar{transform:rotate(90deg)}
#mtp-a2e h3{font-size:19px}
#mtp-a2e .pn{padding:16px}
#mtp-a2e .tb{padding:7px 9px;font-size:12px}
}
</style>
<div id="mtp-a2e">
<div class="hd"><span class="tag">NVIDIA · Open model</span><h3>Alpamayo 2 Edge, explained</h3><p class="sub">A 2B-class reasoning VLA that plans and explains driving decisions on DRIVE AGX Thor.</p></div>
<div class="tabs">
<button class="tb on" data-t="ov">Overview</button>
<button class="tb" data-t="pl">How it works</button>
<button class="tb" data-t="ad">Adaptive reasoning</button>
<button class="tb" data-t="th">Thor deployment</button>
<button class="tb" data-t="bm">Benchmarks</button>
</div>

<div class="pn on" id="a2e-ov">
<div class="grid">
<div class="st"><b>2.65B</b><span>2.45B backbone + 0.2B action expert (marketed as 2B)</span></div>
<div class="st"><b>160-190 ms</b><span>Pixel-to-trajectory on Thor, FP8 + TensorRT</span></div>
<div class="st"><b>15.15 km</b><span>Per at-fault incident, AlpaSim closed loop</span></div>
<div class="st"><b>6.4 s</b><span>64 waypoints at 0.1 s steps</span></div>
<div class="st"><b>6 cams</b><span>22 frames per observation</span></div>
<div class="st"><b>16 GB+</b><span>Recommended GPU memory, BF16 PyTorch</span></div>
<div class="st"><b>75.2</b><span>LingoQA Lingo-Judge score</span></div>
<div class="st"><b>OpenMDW</b><span>Weights OpenMDW-1.1, code Apache 2.0</span></div>
</div>
<p class="note">Sources: Hugging Face model card and NVIDIA Technical Blog, Nov 10, 2026. Benchmarks are NVIDIA-reported.</p>
</div>

<div class="pn" id="a2e-pl">
<p class="txt">Step through one planning cycle. Click Next or Auto-play.</p>
<div class="flow">
<div class="nd" data-i="0"><strong>Inputs</strong>6 cameras, 22 frames, nav text, 20 ego poses</div><div class="ar">›</div>
<div class="nd" data-i="1"><strong>Encoders</strong>Multi-res SigLIP2 vision + motion + text</div><div class="ar">›</div>
<div class="nd" data-i="2"><strong>Reasoner</strong>Cosmos 3 Edge Nemotron-based backbone</div><div class="ar">›</div>
<div class="nd" data-i="3"><strong>Action expert</strong>0.2B diffusion, 10 steps</div><div class="ar">›</div>
<div class="nd" data-i="4"><strong>Outputs</strong>64 waypoints + CoC trace</div>
</div>
<div class="ctl"><button class="btn" id="a2e-prev">‹ Back</button><button class="btn" id="a2e-next">Next ›</button><button class="btn" id="a2e-play"> Auto-play</button></div>
<div class="desc" id="a2e-desc">Press Next to start.</div>
</div>

<div class="pn" id="a2e-ad">
<p class="txt">The model writes a Chain-of-Causation trace only when a scene needs it. In routine driving it emits <code>&lt;NOMINAL&gt;</code> tokens instead.</p>
<div class="seg"><button class="btn" id="a2e-rt">Routine highway</button><button class="btn" id="a2e-cx">Complex scene</button></div>
<div class="stream" id="a2e-stream">Pick a scenario.</div>
<div class="road"><div style="display:flex;justify-content:space-between;font-size:12.5px;color:#cfcfcf"><span>Reasoning tokens generated (illustrative)</span><span id="a2e-tk">0</span></div><div class="meter"><i id="a2e-mt"></i></div></div>
<p class="note">Example traces and token counts are illustrative. NVIDIA states the latency range of 160 to 190 ms depends on the amount of reasoning.</p>
</div>

<div class="pn" id="a2e-th">
<p class="txt">Planning rate shown in NVIDIA’s demo video (AlpaSim, DRIVE AGX Thor).</p>
<div class="bars" id="a2e-hz">
<div class="br"><div class="lb"><span>PyTorch BF16 on Thor</span><em>0.76 Hz</em></div><div class="trk"><div class="fil gr" data-w="5.8"></div></div></div>
<div class="br"><div class="lb"><span>Optimized FP8 Thor pipeline</span><em>13.02 Hz</em></div><div class="trk"><div class="fil g" data-w="100"></div></div></div>
</div>
<p class="txt" style="margin-top:16px">Deployment recipe (Alpamayo Developer Hub). Click a step.</p>
<ol class="steps" id="a2e-steps">
<li class="on">Stage checkpoint + DFlash2 drafter<div class="x">DFlash2 speculative decoding speeds up online reasoning.</div></li>
<li>Calibrate FP8 scales<div class="x">Weights ship in BF16 on Hugging Face.</div></li>
<li>Export ONNX graphs<div class="x">Graphs feed the TensorRT build.</div></li>
<li>Build TensorRT engines<div class="x">Uses NVIDIA TensorRT Edge-LLM plugins.</div></li>
<li>Run the Python harness<div class="x">Outputs are checked against a BF16 PyTorch reference.</div></li>
</ol>
<p class="note">Reference: DRIVE AGX Thor-X, DRIVE OS 7.2.7.0, CUDA 13.4, TensorRT 11.3.1.9.</p>
</div>

<div class="pn" id="a2e-bm">
<div class="seg"><button class="btn" data-b="cl">Closed loop</button><button class="btn" data-b="ol">Open loop</button><button class="btn" data-b="lq">LingoQA</button></div>
<div class="bars" id="a2e-bars"></div>
<p class="note" id="a2e-bn"></p>
</div>

<div class="ft"><span>Data: NVIDIA model card, NVIDIA Technical Blog, Qwen-Drive-1.0 model card. Self-reported.</span><b>© Marktechpost</b></div>
</div>
<script>
(function(){
var R=document.getElementById(‘mtp-a2e’);if(!R)return;
function q(s){return R.querySelectorAll(s)}
function ping(){try{var h=R.offsetHeight+40;parent.postMessage({mtpA2eHeight:h},’*’)}catch(e){}}
function fill(box){var f=box.querySelectorAll(‘.fil’);for(var i=0;i<f.length;i++){f[i].style.width=”0″}setTimeout(function(){for(var i=0;i<f.length;i++){f[i].style.width=f[i].getAttribute(‘data-w’)+’%’}},60)}
var tb=q(‘.tb’);for(var i=0;i<tb.length;i++){tb[i].onclick=function(){for(var j=0;j<tb.length;j++)tb[j].classList.remove(‘on’);this.classList.add(‘on’);var p=q(‘.pn’);for(var k=0;k<p.length;k++)p[k].classList.remove(‘on’);var id=’a2e-‘+this.getAttribute(‘data-t’);document.getElementById(id).classList.add(‘on’);if(id==’a2e-th’)fill(document.getElementById(‘a2e-hz’));if(id==’a2e-bm’)bm(‘cl’);setTimeout(ping,80)}}
var D=[‘Inputs: front wide and front tele give 5 frames each, 4 side and rear cameras give 3 each. Plus a navigation instruction and 20 ego poses at 0.1 s steps.’,’Encoders: multi-resolution inputs make online encoding of surround views cheaper. Vision runs through a 0.41B SigLIP2 encoder.’,’Reasoner: a 1.96B Nemotron-based language model from Cosmos 3 Edge decides whether to reason, then writes a Chain-of-Causation trace if needed.’,’Action expert: a 0.2B diffusion head predicts acceleration and curvature, integrated by a unicycle model. Default is 10 flow-matching steps.’,’Outputs: 64 waypoints from 0.1 to 6.4 s with XYZ and rotation, plus reasoning text. It can also answer free-form scene questions.’];
var nd=q(‘.nd’),cur=-1,tm=null;
function show(i){cur=i;for(var j=0;j<nd.length;j++)nd[j].classList.toggle(‘lit’,j<=i);document.getElementById(‘a2e-desc’).textContent=D[i];ping()}
document.getElementById(‘a2e-next’).onclick=function(){stop();show((cur+1)%5)};
document.getElementById(‘a2e-prev’).onclick=function(){stop();show(cur<=0?4:cur-1)};
function stop(){if(tm){clearInterval(tm);tm=null;document.getElementById(‘a2e-play’).textContent=”▶ Auto-play”}}
document.getElementById(‘a2e-play’).onclick=function(){if(tm){stop();return}this.textContent=”❚❚ Pause”;show((cur+1)%5);tm=setInterval(function(){show((cur+1)%5)},1800)};
var st=document.getElementById(‘a2e-stream’),tk=document.getElementById(‘a2e-tk’),mt=document.getElementById(‘a2e-mt’),ti=null;
function run(parts,n){if(ti)clearInterval(ti);st.innerHTML=”;var k=0;mt.style.width=”0″;tk.textContent=”0″;ti=setInterval(function(){if(k>=parts.length){clearInterval(ti);return}st.innerHTML+=parts[k];k++;var c=Math.round(n*k/parts.length);tk.textContent=c;mt.style.width=(c/256*100)+’%’;ping()},220)}
document.getElementById(‘a2e-rt’).onclick=function(){run([‘<span class="nom">&lt;NOMINAL&gt;</span>n’,’trajectory: 64 waypoints, keep lane’],4)};
document.getElementById(‘a2e-cx’).onclick=function(){run([‘<span class="rs">CoC:</span> ‘,’Cyclist ahead is drifting into the lane. ‘,’Oncoming car blocks a wide pass. ‘,’Slow down and keep a safe gap ‘,’until the oncoming lane clears.n’,’trajectory: 64 waypoints, decelerate’],180)};
var B={cl:{n:’AlpaSim, km between at-fault incidents (higher is better). NVIDIA-run comparison, Figure 2 of the blog. About 60x.’,r:[[‘Alpamayo 2 Edge (2B)’,’15.15 km’,100,’g’],[‘Qwen-Drive-1.0 (4B)’,’0.25 km’,1.65,’gr’]]},ol:{n:’Open-loop minADE in metres (lower is better). NVIDIA-run comparison from the blog chart. The card reports 0.52 m minADE_6 at 5 s.’,r:[[‘Alpamayo 2 Edge (2B)’,’0.53 m’,31.5,’g’],[‘Qwen-Drive-1.0 (4B)’,’1.68 m’,100,’gr’]]},lq:{n:’LingoQA Lingo-Judge score (higher is better). Each value from its own model card or NVIDIA blog. Qwen-Drive value is SFT under official LingoJudge, self-reported.’,r:[[‘Qwen-Drive-1.0 SFT (4B)’,’79.4′,79.4,’gr’],[‘Alpamayo 2 Super (34B)’,’79.2′,79.2,’gr’],[‘Alpamayo 2 Edge (2B)’,’75.2′,75.2,’g’],[‘Alpamayo 1.5 (10B)’,’74.2′,74.2,’gr’]]}};
function bm(k){var b=B[k],h=””;for(var i=0;i<b.r.length;i++){var r=b.r[i];h+='<div class="br"><div class="lb"><span>’+r[0]+'</span><em>’+r[1]+'</em></div><div class="trk"><div class="fil ‘+r[3]+’" data-w="’+r[2]+’"></div></div></div>’}var bx=document.getElementById(‘a2e-bars’);bx.innerHTML=h;document.getElementById(‘a2e-bn’).textContent=b.n;fill(bx);var bb=R.querySelectorAll(‘[data-b]’);for(var j=0;j<bb.length;j++){bb[j].classList.toggle(‘act’,bb[j].getAttribute(‘data-b’)==k)}setTimeout(ping,80)}
var bb=R.querySelectorAll(‘[data-b]’);for(var j=0;j<bb.length;j++){bb[j].onclick=function(){bm(this.getAttribute(‘data-b’))}}
window.addEventListener(‘load’,ping);window.addEventListener(‘resize’,ping);setTimeout(ping,300);
})();
</script>
</body></html>”>
(function(){var f=document.getElementById(“mtp-a2e-frame”);window.addEventListener(“message”,function(e){if(e.data&&e.data.mtpA2eHeight&&f&&e.source===f.contentWindow){f.style.height=e.data.mtpA2eHeight+”px”}})})();

How does Alpamayo 2 Edge compare with other driving VLAs?

FeatureAlpamayo 2 EdgeAlpamayo 2 SuperAlpamayo 1.5Qwen-Drive-1.0-4BOrgNVIDIANVIDIANVIDIAAlibaba QwenTotal params2B marketed; 2.45B backbone + 0.2B action expert (card)34B: 32B backbone + 2B action expert (blog)10B (NGC)4B shared VLM on Qwen3.5-4B (card)Target hardwareIn-vehicle, DRIVE AGX Thor (blog)Cloud, teacher model (NVIDIA)Cloud, 1 GPU 24 GB+ (NGC)Not disclosedMin GPU memory (BF16)16 GB recommended (card)Not disclosed24 GB+ (NGC)Not disclosedTrajectory horizon6.4 s, 64 waypoints (card)Not disclosedNot disclosed5 s, 50 points at 10 Hz (card)Adaptive reasoningYes, <NOMINAL> tokens (blog)Not disclosedNot disclosedNot disclosedLingoQA (Lingo-Judge)75.2 (card)79.2 (blog)74.2 (NVIDIA)79.4 SFT, self-reported (card)Closed loop (AlpaSim)15.15 km per at-fault incident (card)AlpaSim Score 1.50 ± 0.13 (blog)AlpaSim Score 1.37 ± 0.10 (blog)0.25 km per incident in NVIDIA’s run (blog); own card: at-fault score 0.37 (RL)Weights licenseOpenMDW-1.1OpenMDW-1.1OpenMDW-1.1Apache 2.0ReleasedNov 10, 2026Weights Aug 4, 2026Mar 19, 2026Paper Aug 31, 2026 (arXiv)

AlpaSim metrics differ across sources (km per incident vs AlpaSim Score vs at-fault score), so compare only within a row’s same metric.

What else shipped with it?

NVIDIA also released AlpaJudge, a human-calibrated judge for reasoning traces. It scores traces, filters training data and supplies RL rewards. As an RL reward, it lifted reasoning quality 10.2% (judge) and 11.4% (human).

Availability and license

Weights use the OpenMDW-1.1 license, and code uses Apache 2.0. Inference code is on GitHub. The model supports fine-tuning.

Key Takeaways

  • 2B-class reasoning VLA built to run inside the car on DRIVE AGX Thor.
  • Adaptive reasoning: <NOMINAL> tokens skip CoC traces in routine driving.
  • 160 ms to 190 ms pixel-to-trajectory latency with the FP8 TensorRT recipe.
  • 15.15 km per at-fault incident in AlpaSim, about 60x Qwen-Drive-1.0 4B.
  • Weak spot: LingoQA 75.2, behind Alpamayo 2 Super and Qwen-Drive-1.0.

Check out the model card, the technical blog and the GitHub repo. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.

Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us

The post NVIDIA Releases Alpamayo 2 Edge: A 2B Reasoning VLA for DRIVE Thor appeared first on MarkTechPost.



Source link

You might also like
Leave A Reply

Your email address will not be published.