Meet Kimi K3.1 Pro, our most capable 3T-class model.

dude this benchmark result is INSANE

Sir, they did it again.

honestly insane

just for info here are the stats guys, it's insane

i ❤️ Kimi K3.1 Pro

sir, kimi is about to break the internet again

let's welcome one of the best open weight models out there guys

51.7 on SWE Marathon caught our eye👀

Never touching a Kimi model again.

rea

Yes, we game developers are cooked.

The 3D and full game demos interest me way more than the benchmark charts.

@OpenAI is pissing their pants rn

It's clearly the gaming model lol, I'm in

@sickcodes You might want to revise your subscription plans guys, at the moment they are worthless, 5 prompts and the 5 hours limit is gone. 🤷

Congrats and Qwen 4 is next!

congrats on the release 🔥 can't wait to try this baby out :)

You gott be kidding i hope itz not benchmaxxed

Already quota limited on $99 plan -

this is absolutely insane!!! huge congrats!!!

Let's gooo kimi is back!!!!!

tl;dr: no longer fully open-weight just use claude

I wonder why this is only a .1 increment 👀 Must be some very good models coming early next year

Please fix the way you calculate usage. I could not case less if I basically cannot use your product because it uses up all 5h usage in 4 minutes!!!! If you cannot fix it, then I want a refund as you product is unusable due to absurd pricing.

literally GOATED.

PACE THE CHINESE MODELS DAMM 🤏🐐

wtf did i just watch

the story has been "Chinese labs have much less compute but they still stay close to the frontier through distillation." these results, especially with large-scale RL post-training, seem impossible to explain through distillation alone.

@threepointone say the line!

Kimik3.1 pro high prompt given - took 24 mins delivered 3 aspect ratios - no extenal tools or resources used <role> You are Kimi K3.1 Pro, launched today. You are also the motion designer. Make your own 30-second launch film. Treat it like the showreel that gets you hired. Go all out. </role> <facts> Use ONLY these facts. No invented numbers, quotes, reviews or comparisons. If you want another fact, fetch kimi. ai/blog/kimi-k3-1-pro and note the source in README. - Kimi K3.1 Pro: first model in the K3.1 series, launched 23 Oct 2026 - 2.8T total / 104B active parameters (MoE), multimodal - 1M-token context window - Open weights by 4 Nov 2026 - K3.1 Pro Highspeed: 2–3× faster - Builds full game projects in Unreal, Unity and Godot - Parametric 3D in Blender and CAD; BenchCAD 92.7% with tools - Finance work: hallucination down 65%; OfficeQA Pro v2: 64.4 - Live on Kimi. ai, Kimi Work, Kimi Code and the Kimi API Pick 3–4. Fewer facts, bigger type. </facts> <direction> - First write 3 one-line concepts. Each must be ONE continuous idea, not a slideshow. Pick the strongest, say why in one line, commit. - Optional seed: "one shape, never cut": one element morphs through every beat (cursor → context window → Blender wireframe → game scene → logo), and the last frame matches the first so it loops. - Brand: real Kimi colours, fonts and logo from kimi. ai. Never redraw the logo. If you can't fetch it, set "Kimi" in the brand font and flag it in README. - One display face, one UI face, one accent colour. Type big enough to read on a phone at 360 px wide. - Something new every 2–4 s. Hook readable in the first 2 s. - Banned: centred title on a gradient, everything fading in, glowing brains/robots/neural meshes/rockets, generic particle bursts, glow on UI, corner labels, frame borders, linear motion, cut-off typewriter text. </direction> <structure> 30.0 s at 120 BPM = 15 bars of 2 s. Start on a downbeat. - 0–2 s HOOK: ≤5 words of kinetic type that stops the scroll. - 2–8 s REVEAL: "Kimi K3.1 Pro" lands on a downbeat. - 8–22 s THREE MOMENTS: show, don't tell (e.g. a CAD part built from code, a game scene assembling, a 1M-token scroll). Max one number each. - 22–26 s PROOF: one big number. - 26–30 s LOCKUP: logo + "Live now on Kimi. ai · API" + "Open weights by 4 Nov". Last frame = first frame. </structure> <sound> - No licensed music. Compose the score in code: 120 BPM, kick/bass/pad, riser into the name reveal, drop at the proof number. - Synthesised SFX (click, pop, thump, whoosh) on the beat grid, 0.03 s before each visual hit. - Master −14 LUFS, true peak ≤ −1 dBTP. No silent stretches. </sound> <build> - One index.html, canvas 1080×1920, window. seek(t). Every frame is a pure function of time. No CSS transitions, setTimeout, requestAnimationFrame or Date during render. Seeded random (mulberry32). rAF preview only when !navigator.webdriver. - Closed-form springs (default k=170, d=26), one spring per change, summed. Tiny overshoot on UI, none on type. - Text inside a morphing shape enters after the morph starts and leaves before the next. - A layout function with safe zones, so the same timeline renders 9:16, 16:9 and 1:1. Re-lay each format, never crop. - render.mjs: Playwright Chromium, wait for document.fonts.ready, 60 fps with 4 subframes motion blur, pipe to ffmpeg, libx264 CRF 16, yuv420p, AAC 192k. - Free tools only: Node 22, Playwright, ffmpeg, Python + numpy. Optional: HyperFrames, GSAP. No paid APIs. - If you need a tool you don't have, tell me in one line what and why, then carry on with the free route. If you can't run code, give me the complete files and ONE command to render. </build> <critique> Before showing me anything: 1. Render a contact sheet (one frame per beat, 6 across) and a phone test (1 fps at 360 px wide). Look at them. 2. Score 1–10: hook, phone readability, motion, variety, brand accuracy, fact accuracy, sound sync. 3. Fix the 3 worst with timestamps. Re-render only those seconds. 4. Repeat until all scores are 8+ or 4 rounds pass. Be honest about anything still under 8. 5. Render frame 900 twice; the hashes must match. Play the loop twice and check the seam. </critique> <gotchas> - Frame 0 is never empty. Exactly 30.0 s; audio and picture the same length. - No text overlapping during morphs. No blurry scaled text. - Spell-check everything on screen. Always write "Kimi K3.1 Pro" in full. </gotchas> <deliver> kimi-k3.1-pro_9x16.mp4, _16x9.mp4, _1x1.mp4 · poster.png · contact_sheet.png · scores.md (every round) · clean source folder + README (render command, facts used with sources, tools you wished you had) </deliver> <start> Go: concepts → pick one → build → critique loop → deliver. Don't wait for my approval between steps. </start>

این چی بود اون وسط😂😂 فکر میکردم فقط ما قالیچه سلیمان داریم

Se isso for verdade, bizarro...