Black Forest Labs (BFL) has released FLUX 3, a multimodal foundation model that learns from images, videos and audio inside a single architecture. It is also the first FLUX model to ship video, audio and action prediction from one set of weights.
The Black Forest Labs (BFL) research team argues that no single modality gives a complete description of the world. Images capture spatial structure at one instant. Video restores time and exposes physical dynamics. Audio reveals causal relationships between mechanical events and sound. Each is treated as a lossy projection of the same underlying reality.
Training on all of them at once means the modalities constrain each other. The sound has to match the impact. The motion has to obey the mass. The research team calls FLUX 3 its first model built entirely on that principle.
The method underneath: Self-Flow
FLUX 3 builds on Self-Flow, BFL’s method for aligning multimodal generation and understanding in one architecture. Self-Flow combines the flow matching objective with a self-supervised feature reconstruction objective. The reference implementation on GitHub is Apache-2.0 and uses SiT-XL/2 with per-token timestep conditioning. It trains with a 25% per-token mask ratio and self-distillation from an EMA teacher at layer 20 to a student at layer 8.
That released checkpoint is an ImageNet 256×256 research model, not FLUX 3. BFL states that it ‘significantly scaled up compute and data resources’ on the same approach to train FLUX 3 across video, images and audio simultaneously. Self-Flow itself was introduced in March 2026, so it is not new to this launch. What is new is the scale.
What FLUX 3 Video does
FLUX 3 Video generates clips up to 20 seconds long in a single generation, with native audio. The supported modes cover text-to-video, image-to-video, video-to-video from a reference clip, keyframe-to-video for controlled transitions, and generative video-audio continuation from input video and audio.
BFL also lists multilingual dialogue, agentic chaining of clips into multi-shot sequences, and strong typography generation with animated designs. The BFL team reports particular strength in human facial expressions and in associating sounds with physical events.
Performance
BFL team published preliminary human preference results. The setup was 10-second text-to-video clips at 720p with audio. FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons and over Runway Gen-4.5 in 77%. Against Grok Imagine Video the figure is up to 69%, then Kling v3 Pro at 60%, Happy Horse v1 at 59% and Happy Horse 1.1 at 57%. Against Seedance 2.0 and Gemini Omni Flash the result is 52%, close to a coin flip.
Interactive Explorer
:root{
–bg:#0a0a0a; –panel:#121212; –panel2:#181818; –line:#282828;
–fg:#f4f4f4; –dim:#8b8b8b; –orange:#ff7a18; –orange2:#ffab5e; –orange-dim:#7a3a08;
–mono: ui-monospace, “JetBrains Mono”, “SFMono-Regular”, Menlo, Consolas, “Liberation Mono”, monospace;
}
*{box-sizing:border-box;}
html,body{margin:0;padding:0;background:transparent;}
body{font-family:var(–mono);}
#mtp-flux3-term{
background:var(–bg); color:var(–fg); border:1px solid var(–line);
border-radius:10px; overflow:hidden; font-family:var(–mono);
font-size:13.5px; line-height:1.55; max-width:100%;
}
/* title bar */
.tbar{display:flex;align-items:center;gap:10px;padding:9px 14px;background:var(–panel2);border-bottom:1px solid var(–line);}
.dots{display:flex;gap:6px;}
.dot{width:10px;height:10px;border-radius:50%;background:#333;}
.dot.a{background:var(–orange);}
.tbar .path{color:var(–dim);font-size:12px;letter-spacing:.02em;}
.tbar .path b{color:var(–fg);font-weight:600;}
.tbar .tag{margin-left:auto;font-size:10.5px;letter-spacing:.12em;text-transform:uppercase;color:var(–orange);border:1px solid var(–orange-dim);padding:2px 7px;border-radius:3px;}
.wrap{display:grid;grid-template-columns:206px 1fr;min-height:430px;}
/* sidebar */
.side{background:var(–panel);border-right:1px solid var(–line);padding:12px 10px;}
.side h4{margin:0 0 8px 4px;font-size:10px;letter-spacing:.16em;text-transform:uppercase;color:var(–dim);font-weight:600;}
.cmd{display:block;width:100%;text-align:left;background:transparent;border:1px solid transparent;color:var(–fg);
font-family:var(–mono);font-size:12.5px;padding:6px 8px;border-radius:4px;cursor:pointer;margin-bottom:2px;}
.cmd span{color:var(–orange);margin-right:6px;}
.cmd:hover{background:var(–panel2);border-color:var(–line);}
.cmd[aria-current=”true”]{background:#1d1207;border-color:var(–orange-dim);}
.cmd:focus-visible{outline:2px solid var(–orange);outline-offset:1px;}
.side .hintbox{margin-top:12px;padding:8px;border:1px dashed #2e2e2e;border-radius:4px;color:var(–dim);font-size:11px;line-height:1.5;}
/* output */
.main{padding:14px 16px 12px;overflow:auto;}
.line{white-space:pre-wrap;word-break:break-word;}
.c-o{color:var(–orange);} .c-d{color:var(–dim);} .c-w{color:var(–fg);}
.cmdline{margin:0 0 10px;color:var(–dim);}
.cmdline b{color:var(–orange2);font-weight:600;}
.cur{display:inline-block;width:7px;height:14px;background:var(–orange);vertical-align:-2px;animation:bl 1s steps(1) infinite;}
@keyframes bl{50%{opacity:0;}}
h5.sec{margin:14px 0 6px;font-size:10.5px;letter-spacing:.14em;text-transform:uppercase;color:var(–orange);font-weight:600;}
h5.sec:first-child{margin-top:0;}
.note{color:var(–dim);font-size:11.5px;margin:8px 0 0;line-height:1.5;}
/* key-value rows */
.kv{display:grid;grid-template-columns:minmax(120px,190px) 1fr;gap:2px 14px;}
.kv dt{color:var(–dim);}
.kv dd{margin:0;color:var(–fg);}
.kv dd em{font-style:normal;color:var(–orange2);}
/* bars */
.bar{margin:0 0 7px;}
.bar .top{display:flex;justify-content:space-between;gap:10px;font-size:12.5px;}
.bar .top .val{color:var(–orange2);font-variant-numeric:tabular-nums;}
.track{height:9px;background:#1c1c1c;border:1px solid var(–line);border-radius:2px;overflow:hidden;margin-top:3px;}
.fill{height:100%;width:0;background:repeating-linear-gradient(90deg,var(–orange) 0 6px,#e06a10 6px 8px);transition:width .7s cubic-bezier(.2,.8,.2,1);}
.fill.mute{background:repeating-linear-gradient(90deg,#4a4a4a 0 6px,#3a3a3a 6px 8px);}
button.bar{display:block;width:100%;text-align:left;background:transparent;border:0;padding:2px 0;cursor:pointer;font-family:var(–mono);color:var(–fg);}
button.bar:hover .top .lbl{color:var(–orange2);}
button.bar:focus-visible{outline:2px solid var(–orange);outline-offset:2px;}
.readout{margin-top:8px;border-left:2px solid var(–orange);padding:6px 10px;background:var(–panel);color:var(–dim);font-size:12px;min-height:32px;}
.readout b{color:var(–fg);font-weight:600;}
.toolbar{display:flex;flex-wrap:wrap;gap:6px;margin-bottom:10px;}
.chip{background:var(–panel2);border:1px solid var(–line);color:var(–dim);font-family:var(–mono);font-size:11.5px;
padding:3px 9px;border-radius:3px;cursor:pointer;}
.chip[aria-pressed=”true”]{border-color:var(–orange);color:var(–orange);background:#1d1207;}
.chip:focus-visible{outline:2px solid var(–orange);outline-offset:1px;}
/* status table */
table.st{width:100%;border-collapse:collapse;font-size:12.5px;}
table.st th,table.st td{text-align:left;padding:6px 8px;border-bottom:1px solid var(–line);vertical-align:top;}
table.st th{color:var(–dim);font-weight:600;font-size:10.5px;letter-spacing:.1em;text-transform:uppercase;}
.pill{display:inline-block;padding:1px 7px;border-radius:3px;font-size:11px;border:1px solid;}
.pill.live{color:#0a0a0a;background:var(–orange);border-color:var(–orange);font-weight:600;}
.pill.soon{color:var(–orange2);border-color:var(–orange-dim);background:#1d1207;}
.pill.later{color:var(–dim);border-color:var(–line);background:var(–panel2);}
/* slider */
.slab{display:flex;align-items:center;gap:12px;flex-wrap:wrap;margin:6px 0 2px;}
input[type=range]{-webkit-appearance:none;appearance:none;flex:1 1 200px;height:22px;background:transparent;cursor:pointer;}
input[type=range]::-webkit-slider-runnable-track{height:6px;background:#1c1c1c;border:1px solid var(–line);border-radius:3px;}
input[type=range]::-webkit-slider-thumb{-webkit-appearance:none;width:16px;height:16px;margin-top:-6px;border-radius:2px;background:var(–orange);border:0;}
input[type=range]::-moz-range-track{height:6px;background:#1c1c1c;border:1px solid var(–line);border-radius:3px;}
input[type=range]::-moz-range-thumb{width:16px;height:16px;border-radius:2px;background:var(–orange);border:0;}
input[type=range]:focus-visible{outline:2px solid var(–orange);outline-offset:3px;}
.bignum{font-size:26px;color:var(–orange);font-variant-numeric:tabular-nums;letter-spacing:-.01em;}
.bignum small{font-size:12px;color:var(–dim);letter-spacing:.08em;text-transform:uppercase;margin-left:6px;}
/* prompt */
.promptrow{display:flex;align-items:center;gap:8px;border-top:1px solid var(–line);padding:9px 16px;background:var(–panel);}
.promptrow .ps1{color:var(–orange);}
.promptrow input[type=text]{flex:1;background:transparent;border:0;color:var(–fg);font-family:var(–mono);font-size:13px;padding:2px 0;}
.promptrow input[type=text]:focus{outline:none;}
.promptrow .kbd{color:var(–dim);font-size:11px;border:1px solid var(–line);border-radius:3px;padding:1px 5px;}
footer.ft{display:flex;justify-content:space-between;gap:10px;flex-wrap:wrap;
padding:8px 16px;background:var(–panel2);border-top:1px solid var(–line);color:var(–dim);font-size:11px;}
footer.ft a{color:var(–orange);text-decoration:none;}
footer.ft a:hover{text-decoration:underline;}
footer.ft .brand{color:var(–orange);font-weight:600;}
@media (max-width:640px){
#mtp-flux3-term{font-size:12.5px;}
.wrap{grid-template-columns:1fr;min-height:0;}
.side{border-right:0;border-bottom:1px solid var(–line);display:flex;flex-wrap:wrap;gap:4px;padding:10px;}
.side h4{width:100%;}
.cmd{width:auto;flex:1 1 44%;font-size:11.5px;padding:5px 6px;}
.side .hintbox{display:none;}
.main{padding:12px;}
.kv{grid-template-columns:1fr;gap:0 0;}
.kv dt{margin-top:6px;}
table.st th:nth-child(3),table.st td:nth-child(3){display:none;}
}
@media (prefers-reduced-motion:reduce){
.fill{transition:none;} .cur{animation:none;}
}
enter
(function(){
var out = document.getElementById(‘out’);
var input = document.getElementById(‘cmdin’);
var hist = [], hp = -1;
var reduce = window.matchMedia && window.matchMedia(‘(prefers-reduced-motion: reduce)’).matches;
var BENCH = [
{n:’Luma Ray 3.2′, v:93, note:’Highest reported preference margin in BFL’s early evaluation.’},
{n:’Runway Gen-4.5′, v:77, note:’Second-highest margin reported by BFL.’},
{n:’Grok Imagine Video’, v:69, note:’BFL reports “up to” 69% — the phrasing marks an upper bound, not a mean.’},
{n:’Kling v3 Pro’, v:60, note:’Modest but consistent preference margin.’},
{n:’Happy Horse v1′, v:59, note:’Near the middle of the reported field.’},
{n:’Happy Horse 1.1′, v:57, note:’Near the middle of the reported field.’},
{n:’Seedance 2.0′, v:52, note:’Statistically close to a coin flip.’},
{n:’Gemini Omni Flash’, v:52, note:’Statistically close to a coin flip.’}
];
function esc(s){return String(s);}
function bars(list, mode){
var arr = list.slice();
if(mode === ‘sort’) arr.sort(function(a,b){return a.v – b.v;});
var h = ”;
for(var i=0;i<arr.length;i++){
var b = arr[i];
h += '‘;
}
return h;
}
var VIEWS = {
help: {cmd:’help’, body:function(){
return ‘
Available commands
‘
+ ‘
- ‘
- spec
- Model shape, modalities and video limits
- bench
- Preliminary human-preference win rates
- compute
- Where the training budget actually goes
- action
- Cost of adding action prediction to the backbone
- mimic
- FLUX-mimic latency and success rates
- rollout
- What ships when, and to whom
- selfflow
- The training method underneath FLUX 3
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘;
}},
spec: {cmd:’flux3 –spec’, body:function(){
return ‘
One backbone, four output types
‘
+ ‘
- ‘
- Model class
- Multimodal flow matching foundation model
- Joint training modalities
- image + video + audio, one architecture
- Extended output
- Action prediction (robot state)
- Max video length
- 20 seconds in a single generation
- Audio
- Native, generated jointly with video
- Video modes
- text-to-video, image-to-video, video-to-video, keyframe-to-video, video+audio continuation
- Other capabilities
- multilingual dialogue, agentic multi-shot chaining, typography and animated design
- Announced
- 23 July 2026, Black Forest Labs
- Pricing
- not published
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
BFL describes each modality as a lossy projection of one underlying reality. Training on all of them at once lets their mutual constraints supervise each other.
‘;
}},
bench: {cmd:’flux3 bench –video’, body:function(state){
var mode = (state && state.mode) || ‘raw’;
return ‘
Human preference win rate — FLUX 3 vs. baseline
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
Setup: 10-second text-to-video clips at 720p with audio. BFL labels these results preliminary and notes the model and evaluation harness were still in development. Win rates are vendor-run and not independently reproduced.
‘;
}},
compute: {cmd:’flux3 stack –compute’, body:function(){
return ‘
Where the training budget goes
‘
+ ‘
‘
+ ‘
‘
+ ‘
Training corpus
‘
+ ‘
- ‘
- General video
- tens of millions of hours
- Manipulation-focused video
- hundreds of thousands of hours, human and robot
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
BFL’s argument: video is the expensive modality because rendering the world correctly forces the model to learn contact, motion, weight and cause. Audio and actions are low-dimensional signals that attach to a world model already paid for.
‘;
}},
action: {cmd:’flux3 action –transfer’, body:function(){
return ‘
Adding action prediction to a video backbone
‘
+ ‘
+ ‘
‘
+ ‘
‘
+ ‘
+ ‘step 0 / 3500
‘
+ ‘
‘
+ ‘
Two reported endpoints: an initial drop of up to 10% in human ratings, and full recovery of prior video quality after 3,500 steps. The slider interpolates linearly between those two points — it is an illustration of the reported result, not BFL’s published curve.
‘;
}},
mimic: {cmd:’flux3 mimic –deploy’, body:function(){
return ‘
Latency budget on-robot
‘
+ ‘
- ‘
- Backbone, input → world representation
- < 80 ms on a single NVIDIA RTX 5090
- Full self-contained robot system
- 101 ms reaction time
- Inference trick
- action decoder reads latent features, so no video rollout is generated at inference
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
Soft-body kitting, autonomous success rate
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
Numbers reported by mimic robotics for a preview build on their own robot. With the backbone frozen, mimic reports FLUX-mimic stayed competitive while the pi0.5 action decoder dropped to 0%. Deployed tasks include kitting into structured trays, inserting electronic control units, assembly, and handling seals and cables. Audi is testing and deploying the system on production lines.
‘;
}},
rollout: {cmd:’flux3 rollout –status’, body:function(){
return ‘
Launch plan
‘
+ ‘
| Capability | Access | Channel |
|---|---|---|
| FLUX 3 Video | early access | API + private weights |
| FLUX-mimic / FLUX 3 Action | early access | selected research and commercial partners |
| FLUX 3 Image | weeks | API + private weights |
| FLUX 3 Dev | later | open-weight multimodal backbone |
‘
+ ‘
Each tier goes through an early access phase before wider release. BFL has said it will publish more technical detail on the underlying approach. No pricing has been announced for any tier.
‘;
}},
selfflow: {cmd:’flux3 selfflow –paper’, body:function(){
return ‘
The method underneath
‘
+ ‘
- ‘
- Paper
- Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis (ICML 2026)
- arXiv
- 2603.06507
- Authors
- Chefer, Esser, Lorenz, Podell, Raja, Tong, Torralba, Rombach
- Idea
- flow matching objective + self-supervised feature reconstruction, in one framework
- Why it matters here
- generation quality and representation quality improve together, so actions are decodable from the backbone
- Reference code
- Apache-2.0, SiT-XL/2 with per-token timestep conditioning, 25% per-token mask ratio, EMA teacher at layer 20 and student at layer 8
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
‘
+ ‘
The released checkpoint and configuration are the ImageNet 256×256 research model, not FLUX 3 itself. BFL says FLUX 3 is the same method with compute and data scaled up. In Self-Flow experiments, action prediction reached a given success rate in half the training steps of a video model without Self-Flow.
‘;
}}
};
var benchState = {mode:’raw’};
function animateFills(){
var f = out.querySelectorAll(‘.fill’);
requestAnimationFrame(function(){
for(var i=0;i<f.length;i++){ f[i].style.width = f[i].getAttribute('data-w') + '%'; }
setTimeout(sendHeight, 780);
});
}
function render(key, opts){
var v = VIEWS[key];
if(!v){
out.innerHTML = '
command not found: ‘+key+’ — type help for the list.
‘;
sendHeight(); return;
}
var head = ‘
flux3 $
‘;
out.innerHTML = head + ‘
‘;
var target = out.querySelector(‘.cmdline b’);
var bodyEl = out.querySelector(‘#vbody’);
var txt = v.cmd, i = 0;
function finish(){
target.textContent = txt;
bodyEl.innerHTML = v.body(opts || benchState);
bodyEl.style.opacity = ‘1’;
bind(key);
animateFills();
sendHeight();
}
if(reduce){ finish(); return; }
(function type(){
target.textContent = txt.slice(0, i);
i++;
if(i <= txt.length){ setTimeout(type, 16); } else { finish(); }
})();
var btns = document.querySelectorAll('.cmd');
for(var b=0;b<btns.length;b++){
btns[b].setAttribute('aria-current', btns[b].getAttribute('data-k') === key ? 'true' : 'false');
}
}
function bind(key){
if(key === 'bench'){
var chips = out.querySelectorAll('.chip');
for(var c=0;c<chips.length;c++){
chips[c].addEventListener('click', function(){
benchState.mode = this.getAttribute('data-mode');
render('bench', benchState);
});
}
var bs = out.querySelectorAll('button.bar');
var ro = out.querySelector('#readout');
for(var j=0;j<bs.length;j++){
bs[j].addEventListener('click', function(){
ro.innerHTML = 'FLUX 3 vs. ‘ + this.getAttribute(‘data-n’) + ‘ — preferred in ‘
+ this.getAttribute(‘data-v’) + ‘% of comparisons. ‘ + this.getAttribute(‘data-note’);
sendHeight();
});
}
}
if(key === ‘action’){
var s = out.querySelector(‘#steps’);
var qv = out.querySelector(‘#qval’);
var sv = out.querySelector(‘#stepv’);
var ar = out.querySelector(‘#areadout’);
s.addEventListener(‘input’, function(){
var st = parseInt(this.value, 10);
var q = 90 + (st / 3500) * 10;
qv.textContent = q.toFixed(1);
sv.textContent = st.toLocaleString(‘en-US’);
if(st === 0){
ar.innerHTML = ‘At the moment action prediction enters the curriculum, human ratings on text-to-video and image-to-video fall by up to 10%.’;
} else if(st >= 3500){
ar.innerHTML = ‘After 3,500 steps the model has regained its full previous video quality — and now also predicts actions. BFL’s reading: the same backbone carries both.’;
} else {
ar.innerHTML = ‘Recovery phase. The model is learning the structure of the action space and aligning its world representation to it.’;
}
});
}
}
var buttons = document.querySelectorAll(‘.cmd’);
for(var i=0;i<buttons.length;i++){
buttons[i].addEventListener('click', function(){ render(this.getAttribute('data-k')); });
}
input.addEventListener('keydown', function(e){
if(e.key === 'Enter'){
var raw = input.value.trim().toLowerCase();
if(!raw) return;
hist.push(raw); hp = hist.length;
var word = raw.replace(/^flux3s+/, '').split(/[s-]+/)[0];
if(word === 'clear'){ out.innerHTML = '
cleared. type help for the command list.
‘; input.value=”; sendHeight(); return; }
if(raw.indexOf(‘–sort’) > -1){ benchState.mode = ‘sort’; }
render(VIEWS[word] ? word : (word || ‘help’), benchState);
input.value = ”;
} else if(e.key === ‘ArrowUp’){
if(hp > 0){ hp–; input.value = hist[hp]; } e.preventDefault();
} else if(e.key === ‘ArrowDown’){
if(hp < hist.length – 1){ hp++; input.value = hist[hp]; } else { hp = hist.length; input.value=''; } e.preventDefault();
}
});
function sendHeight(){
var el = document.getElementById('mtp-flux3-term');
if(!el) return;
var h = el.offsetHeight + 40;
if(window.parent !== window){
window.parent.postMessage({mtpFlux3Height: h}, '*');
}
}
window.addEventListener('resize', sendHeight);
render('spec');
setTimeout(sendHeight, 120);
setTimeout(sendHeight, 900);
})();
Key Takeaways
- FLUX 3 is one flow matching backbone trained jointly on image, video and audio.
- FLUX 3 Video generates up to 20 seconds with native audio in a single generation.
- Video prediction consumes over 95% of the training compute; audio is under 0.5% of tokens.
- The same backbone drives FLUX-mimic, a robot policy running under 80 ms on one RTX 5090.
- Access is gated: Video and Action are in early access, Image follows, open weights come last.
Check out the FLUX 3 announcement, the FLUX 3 x mimic technical post and the Self-Flow paper. All credit for this research goes to the researchers of this project.
The post Black Forest Labs Releases FLUX 3: A Multimodal Flow Model for Image, Video, Audio and Robot Action Prediction appeared first on MarkTechPost.
This article was originally published on MarkTechPost (AI research simplified). Click below to read the complete article.