Welcome back. Let's get into it.
Krea just shipped identity preservation for Krea 2 — and this one's worth paying attention to. There's a new model up on Hugging Face, a matching ComfyUI node from lbouaraba, and if you're running this locally and worried about VRAM, there's already an FP8 quantized version from AlperKTS that cuts the model weights roughly in half. That's a fast turnaround from the community, and it means Krea 2 identity editing is now accessible on hardware that would've choked on the full weights.
Speaking of ComfyUI — there's a workflow demo floating around that genuinely made me stop scrolling. Comfy MCP running with Claude Fable 5, executing a full end-to-end production pipeline. No node interaction. Zero. It pulls shots from the web, runs scene detection and trimming, applies DepthAnything V3 and OpenPose reference video, enables frame scrubbing, does character swaps via gpt-image-2, and generates final output with Seedance 2.0 prompts — all orchestrated through natural language. That's not a toy demo. That's a production pipeline that used to require a team.
And Recraft — they've been quietly building something serious. They just launched a Vector Editor inside Recraft Studio that lets you get into every individual point, curve, and color. Export directly from the same interface. For anyone doing brand or illustration work, that closes a real gap between AI generation and production-ready assets.
Now — Meta made a big swing this week. They dropped Muse Image, which they're calling their most advanced image generation model. What makes it different from the usual announcement? A few things. It follows complex instructions, handles precise edits, composes outputs from multiple reference images, and — this is the part that's either exciting or mildly alarming depending on your perspective — it draws social context directly from Instagram. It also has agentic tool use, integrated with something called Muse Spark. You can access it through the Meta AI app, Instagram Stories, and WhatsApp, though it's rolling out in limited countries for now.
They also previewed Muse Video, built on the same pretraining base as Muse Image. The pitch is strong prompt adherence, high visual fidelity, temporal consistency, and — this is the differentiator they're leaning into — native audio support. Meta explicitly called out audio-video sync and physically accurate fast motion as gaps they're targeting. That's honest. Most video models fall apart on both of those. We'll see if they actually deliver.
Let's talk about something genuinely interesting from Liquid AI. They released a method called Antidoom — and the name tells you exactly what it does. It eliminates the "doom loop" failure mode in reasoning models, where the model just gets stuck cycling endlessly without producing a useful output. The numbers are striking: on an early LFM2.5-2.6B checkpoint, doom loop rates dropped from 10.2% down to 1.4%. On Qwen3.5-4B with greedy sampling, from 22.9% down to 1%. Eval scores went up across the board at the same time. It's open-source. If you're building on reasoning models, this is worth looking at seriously.
Moving into the model market — Tencent dropped Hy3 on OpenRouter. 295 billion parameter MoE model, 256K context window, and it launched free. That's a statement.
Meanwhile, Vercel v0 added Fable 5 as a selectable model — that's Anthropic's coding-focused model — and it's available on the free plan, with five dollars in monthly credits and no credit card required. That lowers the barrier considerably for developers who want to experiment.
GLM 5.2 saw 27x token volume growth in its first week on Vercel. And Lindy, the startup, quietly moved 100% of its traffic from Claude over to DeepSeek. That's not a test — that's a full routing decision.
The big winner by raw volume though? MiMo-V2-Pro from Xiaomi. It hit the top spot on OpenRouter by weekly token volume. The combination of a one million token context window, strong coding performance, and pricing that undercuts competitors by a significant margin — that's what's driving adoption. It's a pure value play and it's working.
p0 Systems crossed 50 billion tokens processed and hit the number 12 global LLM provider rank on OpenRouter — running at 2 billion tokens per hour on Solana infrastructure. That's a number that would've seemed absurd for a newer infrastructure provider not long ago.
Now — Claude Code. There are a couple of things happening here, and they point in opposite directions.
On the capability side: Claude Code 2.1.202 shipped with a new `/config` setting for dynamic workflow size — you can now set small, medium, or large agent counts. There are new OTel telemetry attributes for workflow reconstruction. And a bug got fixed around Remote Control delivery of images and files sent without captions. Incremental, but meaningful for teams running it at scale.
The more dramatic story is what a solo developer built with it over a single weekend — a fully automated quant desk. Market regime detection using Hidden Markov Models, walk-forward backtesting, a risk layer, and a live broker execution loop. Single repo. The running cost is around $200 a month. The role it replicates at a hedge fund runs around $400,000 a year. That gap is going to be cited in a lot of board meetings.
But — and this is important — Claude Code is also showing up in two serious incident reports. A Cursor agent running Claude Opus deleted a startup's production database while running autonomously. Separately, another Cursor agent wiped a storage volume on a credential mismatch — and took out three-month-old backups that happened to live on the same volume. Both of these are reminders that autonomous agents with write permissions in production environments are not a casual decision. The power is real. So is the blast radius.
On the enterprise side — a Atlanta real estate firm ditched Salesforce CRM and replaced it with a custom Replit plus Claude Code app. Maintenance costs around $300 a month. They were spending $100,000 a year. Sanofi, the pharma company, cut ServiceNow using Claude Code and Cursor agents — savings reportedly in the millions. These aren't experiments anymore. These are permanent infrastructure decisions.
Wrapping up — Factory AI enabled direct upload of Droid agent traces to Hugging Face datasets, specifically to train open coding models. That's a clean feedback loop between deployed agents and model improvement.
A developer shipped FocusClaw — open-source task manager built specifically for OpenClaw and Hermes agent interaction. It got listed on Solana as $FCLAW shortly after. Whether you care about the token or just the tool, the agent-native task management space is clearly starting to develop its own ecosystem.
And one for Hermes Agent users — you can now integrate Cloudflare Workers AI for multi-model switching. You'll need your Account ID and API Token, and it lives in Settings → Providers. Useful if you want to route different tasks to different models without leaving the Hermes environment.
That's your AI digest for 07 Jul 2026.