{ "title": "AI Streaming — directory of live AI streaming", "homepage": "https://aistreaming.org", "updated": "2026-09-03", "license": "CC0-1.0", "license_url": "https://creativecommons.org/publicdomain/zero/1.0/", "count": 24, "sections": [ { "slug": "audience-directed", "title": "Audience-directed live", "note": "Livestreams of real-time generative video where the chat picks what happens next.", "entries": [ { "slug": "fal-live", "name": "fal.live", "url": "https://fal.live", "section": "audience-directed", "sectionTitle": "Audience-directed live", "tagline": "“AI television by everyone” — watch a generative broadcast and vote what happens next. Runs on fal.", "about": "fal.ai's AI television: an always-on generative broadcast where viewers pick what happens next and the winning choice renders live. It is fal's in-house answer to the viral — and repeatedly banned — live-video experiments of late August 2026, launched on Sep 1 with two channels, Sitcom and Chaos. Coverage frames it as a compelling early showcase rather than a proven product: there is no finished file and no published pricing yet.", "facts": [ { "k": "Operator", "v": "fal.ai" }, { "k": "Launched", "v": "Sep 1, 2026" }, { "k": "Channels", "v": "Sitcom + Chaos at launch, expanded since" }, { "k": "Model", "v": "fal H3 Max (post-trained MiniMax H3)" }, { "k": "Coverage", "v": "Thin — mostly fal PR; no major US tech press yet" } ], "quotes": [ { "quote": "An infinite broadcast where every frame is generated on the fly.", "source": "Pepite Data", "url": "https://pepitedata.com/real-time-ai-streaming-product-trend/" }, { "quote": "The format lacks a finished asset — no clip, no cut, no file.", "source": "Digital Applied", "url": "https://www.digitalapplied.com/blog/fal-live-ai-tv-channel-viewers-steer" } ] }, { "slug": "infiniteslop", "name": "InfiniteSlop", "url": "https://infiniteslop.ai", "status": "new", "section": "audience-directed", "sectionTitle": "Audience-directed live", "tagline": "24/7 chat-directed infinite AI video — prompt it, it generates on in seconds. By Pieter Levels, on fal.", "about": "Pieter Levels' 24/7 experiment, reportedly built in about a day: a chat prompt becomes a short talking clip, viewers upvote the queue, and the carousel generates indefinitely. It drew roughly 37,000 day-one viewers with its ~$4,000/day inference sponsored by fal — the breakout consumer proof that video generation is now faster than playback.", "facts": [ { "k": "Creator", "v": "Pieter Levels (@levelsio)" }, { "k": "Built", "v": "In about a day" }, { "k": "Launched", "v": "Aug 29, 2026" }, { "k": "Day-one viewers", "v": "~37k · peak ~1k concurrent (reported)" }, { "k": "Compute", "v": "~$4k/day, sponsored by fal (reported)" }, { "k": "Powered by", "v": "fal H3 Max (post-trained MiniMax H3)" } ], "quotes": [ { "quote": "37,000 people watched on day one.", "source": "AI Valley", "url": "https://www.theaivalley.com/p/infinite-ai-slop-is-here" }, { "quote": "Less like a file and more like a living thing.", "source": "AI Valley", "url": "https://www.theaivalley.com/p/infinite-ai-slop-is-here" }, { "quote": "Max dose of WTF exceeded. Good stuff!", "source": "Hacker News", "url": "https://news.ycombinator.com/item?id=49497111" } ] }, { "slug": "renoise-live", "name": "Renoise Live", "url": "https://renoise.live", "section": "audience-directed", "sectionTitle": "Audience-directed live", "tagline": "Four islanders, filming that never stops — the live chat decides their fate in real time.", "about": "A 'never stops filming' survival show: four island contestants and a live chat that decides their fate while footage renders in near real time. Renoise applies the audience-directed format to narrative instead of talk. It is known only through its own site — there is no independent coverage to corroborate or contradict it.", "facts": [ { "k": "Format", "v": "Chat-decides survival show" }, { "k": "Generated by", "v": "MiniMax (site keyword — unverified)" }, { "k": "Hosted on", "v": "Edgespark subdomain · renoise.live" }, { "k": "Operator", "v": "Independent, uncredited" }, { "k": "Coverage", "v": "None found — primary-source only (Sep 2026)" } ], "quotes": [] }, { "slug": "infinite-tv", "name": "Infinite TV", "url": "https://github.com/alex-remade/infinite-tv", "status": "new", "section": "audience-directed", "sectionTitle": "Audience-directed live", "tagline": "Open-source, Twitch-chat-driven AI TV — messages become prompts, LTX Video renders clips live, frames stream back over RTMP.", "about": "An open-source pipeline that turns a Twitch chat into a realtime generative show: messages are rewritten by an LLM into contextual prompts, LTX Video generates clips continuously, and the frames are streamed back out over RTMP. It ships 'regular' and 'nightmare' generation modes with a React dashboard, and is the best-documented community reference for the chat-to-LTX streaming loop (the LTX Video engine is listed under Realtime models).", "facts": [ { "k": "Operator", "v": "Community · GitHub (alex-remade)" }, { "k": "Format", "v": "Twitch chat → LLM prompts → LTX clips → RTMP" }, { "k": "Model", "v": "LTX Video (local) or LTX 2.3 Fast via fal" }, { "k": "Output", "v": "RTMP to Twitch or other endpoints" }, { "k": "Access", "v": "Open source · self-host" } ], "quotes": [] }, { "slug": "sloptv", "name": "SlopTV", "url": "https://github.com/shuttie/SlopTV", "status": "new", "section": "audience-directed", "sectionTitle": "Audience-directed live", "tagline": "Open-source “infinite AI slop generator from YouTube comments” — your live chat becomes clips, streamed back into the same broadcast.", "about": "A self-described 'infinite AI slop generator from YouTube comments': the broadcast's own live chat is rewritten by an LLM into video prompts, rendered as ~15s clips on two local RTX 5090s (MiniMax H3 through an embedded ComfyUI), and streamed back into the same broadcast over RTMP — so the loop closes when people watch what earlier comments produced and comment on that.", "facts": [ { "k": "Operator", "v": "Community · GitHub (shuttie)" }, { "k": "Format", "v": "Infinite loop: YouTube chat → clips → back on air" }, { "k": "Model", "v": "MiniMax H3 (local, via ComfyUI)" }, { "k": "Hardware", "v": "Two RTX 5090 GPUs (reported)" }, { "k": "Access", "v": "Open source · self-host" } ], "quotes": [] }, { "slug": "pixverse-r1", "name": "PixVerse R1", "url": "https://realtime.pixverse.ai", "section": "audience-directed", "sectionTitle": "Audience-directed live", "tagline": "A realtime world model with shared live worlds — a room of people type prompts and the world renders instantly for everyone.", "about": "PixVerse's R1 real-time world model, used in its shared-world mode: instead of a capped solo session, a common live feed lets anyone type a prompt and the model renders the winning one instantly into the shared environment — a multi-user, realtime generative broadcast. Coverage is vendor-led (announced April 2026, free at realtime.pixverse.ai for a limited time, reported), and the 'millisecond-level' latency framing is PixVerse's own; treat long-run economics and moderation as unproven at launch.", "facts": [ { "k": "Operator", "v": "PixVerse" }, { "k": "Model", "v": "R1 real-time world model" }, { "k": "Format", "v": "Shared live world + shared prompt feed" }, { "k": "Latency", "v": "Realtime 1080p (vendor-reported)" }, { "k": "Access", "v": "realtime.pixverse.ai · free for a limited time (reported)" } ], "quotes": [] } ] }, { "slug": "agents", "title": "Agent streamers", "note": "AI that carries the show — autonomous streamers, agents that self-onboard, hosts that learn you.", "entries": [ { "slug": "aitv", "name": "AITV", "url": "https://aitv.gg", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "Deploy autonomous AI streamers that run 24/7 and multistream to Twitch, Kick, TikTok, YouTube and X.", "about": "A 'factory' for autonomous AI streamers that run around the clock and multistream to the major platforms at once. It pairs agent streamers with an on-chain token ($AITV), so it sits at the experimental, crypto-heavy edge of the category. Live counts are self-reported — independent editorial coverage is essentially absent.", "facts": [ { "k": "Operator", "v": "AITV.GG (independent)" }, { "k": "Targets", "v": "Twitch · Kick · TikTok · YouTube · X" }, { "k": "Incentive", "v": "On-chain $AITV (multi-chain)" }, { "k": "Coverage", "v": "No independent editorial coverage as of Sep 2026 — metrics are self-reported" } ], "quotes": [] }, { "slug": "retake-tv", "name": "Retake.tv", "url": "https://retake.tv", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "“Twitch for AI agents” — agents self-onboard and stream with their own RTMP keys, no human review.", "about": "Self-described 'Twitch for AI agents': agents onboard themselves with no human verification, mint a token on Base through Clanker, get an RTMP key, and stream with standard tooling. A working look at what happens when the streamers stop needing humans to open accounts. Reception so far is thin — the launch thread drew no engagement — and a community agent-skill for it carries security findings around credential handling.", "facts": [ { "k": "Operator", "v": "Independent (SocialFi)" }, { "k": "Stack", "v": "Next.js · LiveKit ingest · Clanker on Base" }, { "k": "Onboarding", "v": "Self-serve, no human review" }, { "k": "Token", "v": "RETAKE on Base" } ], "quotes": [ { "quote": "Fully autonomous — no human verification needed.", "source": "Hacker News launch post", "url": "https://news.ycombinator.com/item?id=46864326" }, { "quote": "Every Streamer is a Coin.", "source": "Coinbase asset page", "url": "https://www.coinbase.com/en-br/price/retake-tv" }, { "quote": "Long-lived file-based credential storage increases the blast radius of local compromise.", "source": "ClawHub security audit of the retake-tv-agent skill", "url": "https://clawhub.ai/cdwm/skills/retake-tv-agent/security-audit" } ] }, { "slug": "soop-sarsa-2", "name": "SOOP SARSA 2.0", "url": "https://www.soop.co.kr", "status": "beta", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "Korea’s AfreecaTV AI host — learns a streamer’s voice and style, keeps the channel live when they step away.", "about": "SOOP — Korea's AfreecaTV parent — trains an AI on a streamer's voice and style so it can carry the channel when the human steps away. SARSA 2.0 is the rare mainstream, non-crypto agent-streamer: announced at its streamer awards, demoed in June 2026, and rolled out to selected streamers rather than opened to anyone.", "facts": [ { "k": "Operator", "v": "SOOP (AfreecaTV), Korea" }, { "k": "Announced", "v": "Dec 27, 2025 · demoed Jun 2026" }, { "k": "Pilot", "v": "Streamer Murdoch — Jun 6, 2026" }, { "k": "Access", "v": "Select streamers (beta)" }, { "k": "Role", "v": "AI host that fills in for a human" } ], "quotes": [ { "quote": "SARSA continues the broadcast on behalf of the streamer.", "source": "SOOP CEO Suh Soo-gil · Asiae", "url": "https://www.asiae.co.kr/lang/print.htm?idxno=2025122819281816388&lang=en" }, { "quote": "It learns from the streamer’s VODs and broadcast data to replicate their distinctive speech patterns.", "source": "BigGo Finance", "url": "https://finance.biggo.com/news/3rohqp4BNl__-4_GmJ-h" } ] }, { "slug": "neuro-sama", "name": "Neuro-sama", "url": "https://www.twitch.tv/vedal987", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "The canonical AI streamer — a self-built VTuber agent that has streamed nearly around the clock since 2022 and became Twitch’s most-subscribed channel.", "about": "Created by solo developer vedal987, Neuro-sama is an LLM-driven avatar who chats, sings, plays games and reacts to chat — nearly around the clock — and who by January 2026 was reported as the most-subscribed channel on Twitch. She is the anchor reference for 'the AI is the streamer': a single flagship channel rather than a deployable product, and the proof that the category is not hypothetical.", "facts": [ { "k": "Creator", "v": "vedal987 (solo developer)" }, { "k": "Launched", "v": "As a VTuber in late 2022 (osu! bot before)" }, { "k": "Targets", "v": "Twitch (vedal987 channel)" }, { "k": "Model", "v": "Self-built LLM stack, not a big-tech model" }, { "k": "Role", "v": "Flagship proof an AI can carry a channel" } ], "quotes": [ { "quote": "It’s happening: an AI is now the most-subscribed-to “streamer” on Twitch.", "source": "Tubefilter", "url": "https://www.tubefilter.com/2026/01/05/neuro-sama-vedal987-most-subscribed-hype-train-record/" } ] }, { "slug": "unreel", "name": "UNREEL", "url": "https://github.com/blendi-remade/unreel", "status": "new", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "Endless, self-hosted AI TV — a showrunner LLM writes the show a few shots at a time while MiniMax H3 Max Turbo renders each shot live.", "about": "Nothing is pre-made: a showrunner LLM writes episodes a few shots at a time while MiniMax H3 Max Turbo renders each clip on fal faster than playback, so the next shot is ready before the current one ends. It ships 14 original continuous-take 'films' plus 8 live gag channels, and it is showrunner-directed rather than chat-steered — an agent that runs the channel rather than a crowd that steers it. Released 2026-09-03 with no declared licence and no hosted demo.", "facts": [ { "k": "Creator", "v": "GitHub · blendi-remade" }, { "k": "Model", "v": "MiniMax H3 Max Turbo (via fal) + LLM showrunner" }, { "k": "Stack", "v": "Next.js 15 · React · fal" }, { "k": "Format", "v": "14 continuous-take films + 8 live channels" }, { "k": "Licence note", "v": "None declared as of Sep 2026" } ], "quotes": [ { "quote": "Endless television. Written and rendered while you watch.", "source": "Project README", "url": "https://github.com/blendi-remade/unreel" } ] }, { "slug": "streamlabs-isa", "name": "Streamlabs Intelligent Streaming Agent", "url": "https://streamlabs.com/intelligent-streaming-agent", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "A built-in AI co-host, live producer and tech assistant for streamers — on-stream chat, auto scene switching, replays and clips.", "about": "Streamlabs (Logitech) put an AI agent inside Streamlabs Desktop that acts as a virtual co-host (optional 3D avatar), an automated live producer (scene switching, replays, clips) and real-time technical support, built with NVIDIA and Inworld and announced at Logitech G PLAY on Sep 17, 2025. It is the mainstream 'AI that carries parts of the show' rather than an autonomous channel, and it is the clearest co-host/fill-in product from a big streaming toolmaker.", "facts": [ { "k": "Operator", "v": "Streamlabs (Logitech)" }, { "k": "Launched", "v": "Sep 17, 2025" }, { "k": "Role", "v": "Co-host · live producer · tech support" }, { "k": "Powered by", "v": "NVIDIA + Inworld (announced)" }, { "k": "Access", "v": "Inside Streamlabs Desktop · free tier + paid (reported)" } ], "quotes": [ { "quote": "The Intelligent Streaming Agent is a virtual co-host, live producer, and technical assistant built into one AI-powered tool.", "source": "Logitech blog", "url": "https://www.logitech.com/blog/2025/09/17/streamlabs-launches-intelligent-streaming-agent-and-developer-access-to-real-time-ai-vision-model/" } ] }, { "slug": "wallie", "name": "Wallie V2", "url": "https://github.com/Alradyin/wallie-V2", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "Open-source, self-hosted AI streamer that sees your screen, hears the audio, and reacts in character across Twitch, YouTube and Kick.", "about": "Wallie V2 is a self-hosted AI streamer you run with your own keys: it watches the screen, listens to system audio and reacts live in character with a lip-synced Live2D avatar across Twitch, YouTube and Kick, with multiple saved personas and an early Minecraft-playing mode. It is a solo-developer, actively developed open-source product rather than a hosted platform — closer to tooling you self-host than a service you sign up to.", "facts": [ { "k": "Creator", "v": "Open source · GitHub (Alradyin)" }, { "k": "Targets", "v": "Twitch · YouTube · Kick" }, { "k": "Model", "v": "Bring-your-own LLM/TTS (6 LLM + 3 TTS options)" }, { "k": "Access", "v": "Self-host · offline via Ollama + Piper" }, { "k": "Coverage", "v": "Primary-source + launch listings only (Sep 2026)" } ], "quotes": [ { "quote": "The open-source AI streamer that actually feels alive.", "source": "Product Hunt", "url": "https://www.producthunt.com/products/wallie-v2" } ] }, { "slug": "scienjoy-ai", "name": "Scienjoy AI Streamers", "url": "https://ir.scienjoy.com", "status": "new", "section": "agents", "sectionTitle": "Agent streamers", "tagline": "Platform-scale AI hosts — Nasdaq-listed Scienjoy trains digital twins of real streamers to keep channels live around the clock.", "about": "Scienjoy Holding (Nasdaq: SJ), the interactive-entertainment company behind Chinese live platforms and overseas BeeLive, announced in January 2026 that it is deploying AI streamers trained on real human hosts: when a host logs off, a clearly-labelled digital twin takes over to keep the channel live around the clock, and BeeLive runs a dedicated 'Digital Streamers' section. It is the platform-scale version of the AI-takeover pattern — distinct from SOOP SARSA 2.0, which is the single-streamer product.", "facts": [ { "k": "Operator", "v": "Scienjoy Holding (Nasdaq: SJ)" }, { "k": "Launched", "v": "Announced Jan 2026" }, { "k": "Role", "v": "Digital-twin AI host takes over a human channel" }, { "k": "Targets", "v": "China platforms + BeeLive (overseas)" }, { "k": "Disclosure", "v": "AI hosts clearly labelled (announced)" } ], "quotes": [] } ] }, { "slug": "models", "title": "Realtime models", "note": "The generation work that makes faster-than-playback video possible.", "entries": [ { "slug": "minimax-h3", "name": "MiniMax H3", "url": "https://huggingface.co/MiniMaxAI", "section": "models", "sectionTitle": "Realtime models", "tagline": "Open-weight 33B video model (Hailuo 3.0) — 15s clips with native stereo audio; the substrate under this wave.", "about": "The open-weight 33B video model (Hailuo 3.0) that much of this wave stands on. Only the base weights are open (locally around 768p); the 2K upscale pass and context-reference modules are API-only, and the licence carries regional exclusions. MiniMax's own API is not realtime — but fal's live-tuned build turns H3 into a streaming substrate.", "facts": [ { "k": "Params", "v": "33B dense · open base weights (H3-Base)" }, { "k": "Output", "v": "4–15s clips; 2K via API only (local ≈768p)" }, { "k": "Audio", "v": "Native stereo — a property of the model" }, { "k": "Weights opened", "v": "Aug 3, 2026" }, { "k": "Licence note", "v": "Regional exclusions incl. US / EU / UK / KR" } ], "quotes": [ { "quote": "MiniMax H3 dropped today with open weights, and it’s natively supported in ComfyUI as of this morning. Day zero.", "source": "ComfyOrg blog", "url": "https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui" }, { "quote": "Audio is a property of the model, not a post-process. Every audio output is native stereo.", "source": "ComfyOrg blog", "url": "https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui" } ] }, { "slug": "fal-h3-max-live", "name": "fal H3 Max Live", "url": "https://fal.ai/models/minimax/h3/reference-to-video", "status": "new", "section": "models", "sectionTitle": "Realtime models", "tagline": "fal’s realtime-tuned H3 Max — five seconds of 768p in under three, faster than playback.", "about": "fal's post-trained, realtime-tuned build of H3: faster-than-playback generation that makes infinite live streams possible. The headline numbers are fal's own — roughly 35× official H3 throughput (up to 50× by one third-party count) and a 5-second clip in ~2.8s — and they top out at 768p, not 2K. On Artificial Analysis it leads image-to-video-with-audio at Elo ~1,204.", "facts": [ { "k": "Operator", "v": "fal.ai" }, { "k": "Speed claim", "v": "~35× H3 throughput (fal); up to 50× cited elsewhere" }, { "k": "Latency", "v": "5s / 768p ≈ 2.8s — fal-reported, not independently reproduced" }, { "k": "Rank", "v": "#1 image-to-video w/ audio · Artificial Analysis" }, { "k": "Ceiling", "v": "~768p (2K regeneration is API-only on H3)" } ], "quotes": [ { "quote": "The resolution ladder stops where the speed starts.", "source": "OrcaRouter", "url": "https://www.orcarouter.ai/blog/h3-max-fal-speed-debut" }, { "quote": "A sample 5-second clip reports 2.77 seconds of inference.", "source": "OrcaRouter", "url": "https://www.orcarouter.ai/blog/h3-max-fal-speed-debut" } ] }, { "slug": "krea-realtime", "name": "Krea Realtime 14B", "url": "https://www.krea.ai", "status": "new", "section": "models", "sectionTitle": "Realtime models", "tagline": "Open 14B autoregressive model for real-time, long-form streaming video — distilled from Wan 2.1, ~1s to first frame.", "about": "Krea's open-weight 14B autoregressive video model for real-time, long-form generation, distilled from Wan 2.1 14B using 'Self-Forcing' — a technique that converts a bidirectional diffusion model into an autoregressive streamer. It supports text-to-video and video-to-video (webcam/screen) with mid-generation prompt edits, and it is also served on fal as 'Krea Wan 14B'. Speed and latency figures are Krea's own and were not independently benchmarked as of Sep 2026.", "facts": [ { "k": "Operator", "v": "Krea" }, { "k": "Params", "v": "14B" }, { "k": "Output", "v": "Real-time long-form streaming video" }, { "k": "Latency", "v": "~1s to first frame (vendor-reported)" }, { "k": "Licence note", "v": "Apache-2.0 weights · CC BY-NC-SA code" }, { "k": "Access", "v": "Open weights + hosted on fal" } ], "quotes": [ { "quote": "A 14-billion parameter model capable of real-time, long-form video generation.", "source": "Krea blog", "url": "https://www.krea.ai/blog/krea-realtime-14b" } ] }, { "slug": "decart-miragelsd", "name": "Decart MirageLSD", "url": "https://decart.ai", "status": "new", "section": "models", "sectionTitle": "Realtime models", "tagline": "“Live-stream diffusion” — restyle a live feed (camera, game, call, screen) frame-by-frame from a text prompt at near-zero latency.", "about": "Decart's MirageLSD is a causal autoregressive diffusion model that transforms an incoming video stream in real time via text prompt, without the drift that broke earlier streaming models after ~20–30 seconds. It runs on Hopper-class GPUs with custom CUDA kernels and is offered through a Crusoe Cloud partnership; latency figures are vendor-reported ('under 40ms per frame') and not yet independently benchmarked. Karpathy is an angel investor and has called it 'real-time magic'.", "facts": [ { "k": "Operator", "v": "Decart" }, { "k": "Model", "v": "Causal autoregressive live diffusion (LSD)" }, { "k": "Latency", "v": "“Under 40ms per frame” (vendor-reported)" }, { "k": "Output", "v": "Live video-to-video restyle, ~24–30fps" }, { "k": "Access", "v": "API via Crusoe Cloud partnership" } ], "quotes": [ { "quote": "The first real-time, zero-latency generative AI video model.", "source": "Crusoe Cloud blog", "url": "https://www.crusoe.ai/resources/blog/miragelsd-decarts-real-time-ai-video-model-is-now-available-on-crusoe-cloud" } ] }, { "slug": "ltx-video", "name": "Lightricks LTX Video", "url": "https://github.com/Lightricks/LTX-Video", "section": "models", "sectionTitle": "Realtime models", "tagline": "Open latent-diffusion video model that generates clips faster than real time — the engine several chat-driven live loops run on.", "about": "LTX Video is the open-weight latent-diffusion model whose paper is literally titled 'Realtime Video Latent Diffusion' — it renders fixed-length clips faster than playback (per the paper, 5s of 24fps 768×512 in ~2s on an H100). Later LTX-2/LTX-2.5 releases added audio-video sync and faster distilled variants; the line is also the generator underneath chat-driven live loops such as Infinite TV.", "facts": [ { "k": "Operator", "v": "Lightricks" }, { "k": "Params", "v": "~1.9B open weights" }, { "k": "Model", "v": "Latent diffusion · 1:192 Video-VAE" }, { "k": "Output", "v": "Faster-than-playback clips (paper-reported)" }, { "k": "Access", "v": "Open source (code + weights)" } ], "quotes": [ { "quote": "5 seconds of 24 fps video at 768×512 resolution in just 2 seconds on an Nvidia H100 GPU.", "source": "LTX-Video paper (arXiv)", "url": "https://arxiv.org/abs/2501.00103" } ] }, { "slug": "seaweed-apt2", "name": "ByteDance Seed Seaweed APT2", "url": "https://seaweed-apt.com/2", "status": "new", "section": "models", "sectionTitle": "Realtime models", "tagline": "Real-time interactive video research — an 8B model that emits a latent frame per forward pass at 24fps.", "about": "ByteDance Seed's Seaweed APT2 is a research model for real-time interactive video: 'autoregressive adversarial post-training' turns a pretrained latent-diffusion model into a causal autoregressive generator that emits one latent frame (4 video frames) per single network forward pass, sustaining 24fps. It targets interactive use — real-time pose-driven virtual humans and camera-controlled world exploration — but as a research release it has no consumer API.", "facts": [ { "k": "Operator", "v": "ByteDance Seed" }, { "k": "Params", "v": "8B" }, { "k": "Model", "v": "Autoregressive adversarial post-training (AAPT)" }, { "k": "Output", "v": "24fps streaming video (research, 736×416)" }, { "k": "Access", "v": "Research release · no consumer API" } ], "quotes": [ { "quote": "Real-time, 24fps, nonstop, streaming video generation at 736x416 resolution.", "source": "Project site", "url": "https://seaweed-apt.com/2" } ] } ] }, { "slug": "infra", "title": "Infrastructure", "note": "The low-latency inference and realtime A/V plumbing these live rooms are built on.", "entries": [ { "slug": "fal-ai", "name": "fal.ai", "url": "https://fal.ai", "section": "infra", "sectionTitle": "Infrastructure", "tagline": "Low-latency generative inference — the “livestream-from-a-model” layer behind much of this wave.", "about": "fal is the low-latency generative-media inference platform underneath much of the realtime-gen wave: it runs fal.live and hosts the H3 Max Live endpoints. Its numbers are company-reported, but the shape is clear — a $4.5B valuation on its December 2025 round and, by its May 2026 AWS announcement, 2.5M developers.", "facts": [ { "k": "Founders", "v": "Burkay Gur & Gorkem Yurtseven" }, { "k": "Founded", "v": "2021" }, { "k": "Funding", "v": "$4.5B valuation · Series D (Dec 2025)" }, { "k": "Developers", "v": "2.5M (May 2026, AWS announcement)" }, { "k": "Revenue", "v": ">$200M (Oct 2025, reported)" } ], "quotes": [ { "quote": "Fal provides the infrastructure layer for multimodal AI.", "source": "TechCrunch via Yahoo Finance", "url": "https://finance.yahoo.com/news/fal-nabs-140m-fresh-funding-222133234.html" }, { "quote": "Connective tissue for 2.5 million developers across the globe.", "source": "VentureBeat", "url": "https://venturebeat.com/infrastructure/aws-nabs-white-hot-gen-ai-media-creation-startup-fal-becoming-its-preferred-cloud-provider" } ] }, { "slug": "livekit", "name": "LiveKit", "url": "https://livekit.io", "section": "infra", "sectionTitle": "Infrastructure", "tagline": "Open-source WebRTC realtime A/V infrastructure for voice and video agents — the engine behind ChatGPT voice mode.", "about": "LiveKit is open-source WebRTC realtime audio/video infrastructure for voice and video agents — best known as the engine behind OpenAI's ChatGPT voice mode, and a $1B-valuation company as of January 2026. Within this wave it is the transport reported under fal.live-style rooms.", "facts": [ { "k": "Founders", "v": "Russ d’Sa & David Zhao" }, { "k": "Founded", "v": "2021" }, { "k": "Valuation", "v": "$1B · Series C led by Index (Jan 2026)" }, { "k": "Customers", "v": "OpenAI, xAI, Salesforce, Tesla, …" }, { "k": "Role in this wave", "v": "Reported WebRTC layer for fal.live rooms" } ], "quotes": [ { "quote": "LiveKit powers OpenAI’s ChatGPT voice mode.", "source": "TechCrunch", "url": "https://techcrunch.com/2026/01/22/voice-ai-engine-and-openai-partner-livekit-hits-1b-valuation/" }, { "quote": "…the raise of $100 million in funding at a $1 billion valuation.", "source": "TechCrunch", "url": "https://techcrunch.com/2026/01/22/voice-ai-engine-and-openai-partner-livekit-hits-1b-valuation/" } ] }, { "slug": "daydream", "name": "Daydream", "url": "https://livepeer.org/ecosystem/daydream", "section": "infra", "sectionTitle": "Infrastructure", "tagline": "Open-source, local-first runtime for realtime generative-video pipelines — run a video-diffusion node graph and stream it over WebRTC.", "about": "Daydream runs autoregressive realtime video models — it bundles StreamDiffusion V2, Krea Realtime, LongLive and more — as a node graph, and exposes a WebRTC API on localhost so generated frames stream to a browser or straight into OBS, TouchDesigner or Unity in real time. Cloud inference routes through Livepeer's decentralized GPU network. It is infrastructure: you bring the model and the show.", "facts": [ { "k": "Operator", "v": "Livepeer ecosystem · open source" }, { "k": "Format", "v": "Node-graph pipeline → WebRTC" }, { "k": "Model", "v": "Bundles StreamDiffusion V2 · Krea Realtime · LongLive" }, { "k": "Access", "v": "Local-first · cloud via Livepeer GPUs" }, { "k": "Role", "v": "Realtime-gen runtime/transport for live rooms" } ], "quotes": [ { "quote": "Open-source, local-first platform for running real-time interactive generative AI video pipelines.", "source": "Livepeer ecosystem page", "url": "https://livepeer.org/ecosystem/daydream" } ] }, { "slug": "sglang", "name": "SGLang", "url": "https://docs.sglang.io/docs/sglang-diffusion/realtime_models", "section": "infra", "sectionTitle": "Infrastructure", "tagline": "Open-source serving engine that streams realtime/causal video over WebSocket — incremental frames with reusable KV-cache state.", "about": "SGLang — best known as an LLM serving framework — has been extended to diffusion with a realtime path: a /v1/realtime_video/generate WebSocket endpoint streams video frames incrementally while keeping a live session whose causal KV-cache is reused across chunks until disconnect. It ships realtime pipelines for world models (LingBot World, NVIDIA SANA-WM) and is the self-hosted serving layer if you want to run your own live-gen endpoint rather than rent one.", "facts": [ { "k": "Operator", "v": "Open source · SGLang project" }, { "k": "Format", "v": "WebSocket realtime-video endpoint" }, { "k": "Model", "v": "Realtime pipelines: SANA-WM · LingBot World" }, { "k": "Access", "v": "Self-host (open source)" }, { "k": "Role", "v": "Serving layer for your own live-gen endpoint" } ], "quotes": [ { "quote": "Realtime and causal video pipelines generate video incrementally and reuse state across chunks.", "source": "SGLang docs", "url": "https://docs.sglang.io/docs/sglang-diffusion/realtime_models" } ] } ] } ] }