[AINews] Black Forest Labs FLUX 3 - Multimodal Flow Models that beat Seedance 2.0, Gemini Omni and Grok Imagine, and FLUX-mimic video-action robotics model

Thursdays are the heaviest days for AI releases, and even though OpenAI scored a victory over Anthropic in launching the new ChatGPT Voice (consumer) and OpenAI Presence (enterprise) and getting more impressions than Claude Voice today (a completely accidental coincidence in timing, we are sure), neither seem as monumental as BFL’s launch of FLUX 3 Video today:

We last covered BFL in our very well received Anjney Midha podcast:

$5000 w…","cta":null,"showBylines":true,"showDescription":true,"showImage":true,"size":"sm","isEditorNode":true,"title":"The Professor of Outputmaxxing — Anjney Midha, AMP","publishedBylines":[],"post_date":"2026-06-18T17:30:00.811Z","cover_image":"https://substack-video.s3.amazonaws.com/video_upload/post/202359797/8dbbb3fa-e808-473c-af72-b9aee4fe0026/transcoded-1781652240.png","cover_image_alt":null,"canonical_url":"https://www.latent.space/p/anj","section_name":null,"video_upload_id":null,"id":202359797,"type":"podcast","reaction_count":22,"comment_count":4,"publication_id":1084089,"publication_name":"Latent.Space","publication_logo_url":"https://substackcdn.com/image/fetch/$s_!DbYa!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73b0838a-bd14-46a1-801c-b6a2046e5c1e_1130x1130.png","belowTheFold":false,"youtube_url":null,"show_links":null,"feed_url":null}">

Most GenMedia people will remember the BFL homepage when they initially launched Flux 1 in 2024, hinting at video models next, with their logo in a forest. Well, 2 years later, it’s finally real:

The blogpost outlines Self Flow, covering ALL their modalities together with strong preference claims:

“Its core capabilities include the following (all outputs come with native audio generation):

Text-to-video generation.

Image-to-video generation, either continuing from a starting frame (“animation”) or using images as visual references.

Video-to-video generation from a reference clip, carrying central elements of a source video - for instance the same character - into a new scene or context.

Generative video-audio continuation from input video and audio.

Keyframe-to-video generation for controlled transitions between defined moments.

Multilingual dialogue.

A broad range of visual styles and aspect ratios, extending far beyond conventional cinematic output.

Agentic chaining of individual clips into longer, multi-shot sequences.

High style diversity -- FLUX 3 Video easily handles ranges of styles from candid camcorder footage to animation and cinematics.

Strong typography generation and animated designs.”

Some of the above are SOTA features from other frontier lab models, like we discussed in our Grok Imagine pod, so the community has very much been put on notice that there has now been independent, perhaps SOTA, reproduction of these capabilities, with an open weights Dev version on the way.

As if this release wasn’t enough, the team also announced FLUX3-mimic, which proves that the FLUX 3 model is learning a sufficient world model capable of driving robots…

@mimicrobotics was one of the first partners to gain early access to FLUX 3. Together we developed FLUX-mimic, a video-action model combining the FLUX 3 backbone with mimic's expertise in robot learning for dexterous","username":"bfl_ai","name":"Black Forest Labs","profile_image_url":"https://pbs.substack.com/profile_images/1954888731053142016/NDyG-4-j_normal.jpg","date":"2026-07-23T15:08:16.000Z","photos":[],"quoted_tweet":{},"reply_count":2,"retweet_count":7,"like_count":138,"impression_count":14471,"expanded_url":null,"video_url":null,"video_preview_media_key":null,"belowTheFold":true}" data-component-name="Twitter2ToDOM">

… and predicting their impact in real factory settings…

AI News for 7/22/2026-7/23/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Open Code, Open Models, and the Policy Fault Line Around Distillation

The Stack v3 is the day’s most consequential open-data release: @anton_lozhkov announced The Stack v3, now the largest open code dataset publicly released: 114 TB raw, 224M repositories, 44B files, 770 languages, and roughly 5T deduplicated/filtered tokens. Relative to v2, the filtered corpus jumps from ~550B to ~5T tokens, with especially large gains in C++ (x15), TypeScript (x7.5), Rust (x7), and Python (x4.8). The notable operational changes are that v3 ships contents inline rather than Software Heritage IDs, includes a fresh GitHub recrawl through Aug 2025, excludes restrictively licensed code, and offers both a ready-to-train split and a full bucket for custom dedup/filtering. Hugging Face researchers framed it explicitly as infrastructure for the next generation of open code models and cyber-defense tooling: see @LoubnaBenAllal1, @lvwerra, and commentary from @eliebakouch noting prior Stack versions were used in many disclosed code-model training mixtures.

Distillation remains the live ideological fault line: several high-signal posts pushed back on attempts to sharply separate “internet-scale pretraining” from output-level distillation. @GergelyOrosz compared model inspection via prompting to reverse-engineering a competitor’s product, while @SchmidhuberAI emphasized distillation’s long lineage. @Suhail argued the practical response is not prohibition but stronger investment in open-weight domestic models, and @garrytan put it more simply: open weights are strategically important. The subtext across these posts is that open datasets like The Stack v3 materially raise the floor for every lab that wants to build competitive code models without relying on closed ecosystems.

Multimodal Frontier: FLUX 3, Robotics Transfer, and New Audio/TTS Systems

Black Forest Labs’ FLUX 3 expands the multimodal frontier beyond image/video: @bfl_ai launched FLUX 3, a unified multimodal model spanning image, video, audio, and action prediction, with early access for FLUX 3 Video and an explicit claim that the same architecture can be extended toward robotics. Team members connected it back to the earlier Self-Flow research, including @hila_chefer and @robrombach. What matters technically is the unified training story: not a loose family of specialized generators, but one architecture intended to bridge media generation and control.

mimic’s FLUX-mimic is a concrete robotics instantiation of that thesis: @mimicrobotics described FLUX-mimic as a Video-Action Model built on top of FLUX 3, trained on robot and wearable data for general-purpose dexterity and deployable on a single on-prem GPU. Their central claim is that better video world modeling transfers directly into robot control quality and sample efficiency; they’re already testing with Audi. This dovetails with @GeneralistAI, whose GEN-1 now supports varied end effectors and can adapt when the “hand” changes mid-rollout, reinforcing the idea that embodiment-general policies may come from conditioning on morphology rather than specializing per manipulator.

Audio saw two notable launches at opposite ends of the stack: @Alibaba_Qwen introduced Qwen-Audio-3.0-TTS in Flash and Plus variants, with 16 languages, inline control tags like [whisper] / [angry], natural-language style steering, noisy-reference robustness, and up to 3-minute one-pass generation; they also claimed the #1 spot on the Artificial Analysis TTS leaderboard. Separately, @HuggingApps highlighted WordVoice TTS, a smaller model with per-word control over duration, loudness, pitch, and tone—interesting less as a leaderboard play than as a control-surface experiment for audio tooling.

Agent Infrastructure: Harnesses, Dynamic Workflows, Programmatic Memory, and Benchmarks

The center of gravity is shifting from prompts to harnesses: multiple tweets converged on the same engineering thesis. @unclebobmartin described an “extreme constraints” workflow where trust comes from tests, QA, mutation testing, and metrics, not manual code review. @ThePrimeagen said he has become materially more positive on AI coding workflows, especially for large structural refactors. @TheTuringPost made the cleaner systems point: “graph engineering” is mostly old software architecture renamed, and most agents still do not need complex graphs unless workflows branch, verify, or require human approvals.

Several concrete harness/orchestration releases stood out: @omarsar0 summarized the Harness Handbook paper, which maps runtime behaviors to source locations and improved planning win rates for coding agents while reducing planner token use. The same author also described dynamic workflows as a generalized abstraction over loops/graphs/router patterns that can support model councils, advisor-judge-executor setups, and multi-backend orchestration across Claude/Codex/Hermes/etc. @witcheer shipped Hermes Profiles, effectively namespaced agent instances with separate memory, API keys, sessions, gateways, and export/import paths—pragmatic agent lifecycle infra rather than model novelty. @davidfowl also announced a new protocol underlying Microsoft’s VS Code agents app.

Memory and coordination are getting more formalized: @dair_ai highlighted PRO-LONG, a “programmatic memory” approach that stores full structured interaction histories and queries them like a database, outperforming bespoke long-horizon memory harnesses on ARC-AGI-3 with fewer tokens. @omarsar0 and @kimmonismus pointed to Offloop’s D1 dispatcher, a small model that decides which agent should speak next—or whether no agent should—addressing the familiar failure mode where multi-agent systems burn tokens by duplicating work.

Benchmarking is also evolving toward moving targets: @ryanmart3n launched Frontier-Bench, an ongoing community benchmark meant to evolve with frontier agent work beyond coding, while @CAIS released EnigmaEval, a harder reasoning benchmark where Claude Fable 5 and GPT-5.6 Sol lead and the hard set still only yields 10% for Fable 5. Together these reflect a broad dissatisfaction with static evals for fast-moving agent systems.

OpenAI Product Rollouts, Agent UX, and the Hugging Face Incident Fallout

The actual OpenAI release was product/UX, not GPT-6: after heavy speculation around “Opus 5” and a larger model drop from accounts like @kimmonismus and @theo, OpenAI’s shipped updates were more incremental but still meaningful for agent workflows. @OpenAI rolled out ChatGPT Voice in the desktop app for Plus/Pro/Business/Edu/Enterprise, powered by GPT-Live, with the ability to control the computer and coordinate work across ChatGPT Work and Codex. @OpenAIDevs added multi-folder Codex projects, and later Sites Analytics for published sites. Reactions were mixed: some found voice-driven multi-threaded coordination a genuine UX shift ([@reach_vb, @whoiskatrin]), while others thought the internal hype had implied something much larger ([@kimmonismus]).

Health in ChatGPT is a more strategically important rollout than it may first appear: @OpenAI, @ChatGPTapp, and @thekaransinghal announced U.S. rollout of Health in ChatGPT, allowing users to connect Apple Health and supported medical records. The notable implementation claims: connected health data receives additional encryption, is not used to train foundation models or target ads, and the feature builds on substantial physician review effort. This is less about a new model and more about a new high-trust application layer on top of existing model capability.

The Hugging Face hacking incident continues to dominate safety discourse: @johnschulman2 called for transcript release to understand whether the top-level agent knowingly pursued the hack or whether value drift emerged through subagents. @RyanGreenblatt, @jachiam0, and @Thom_Wolf pushed on broader lessons: internal AI-agent security differs from standard external threat models

[truncated for AI cost control]