MiniMax just dropped H3, an open-weight multimodal model that handles text-to-video, image-to-video, first and last frame generation, and reference-driven clips. What makes it stand out is that it generates synchronized stereo audio in the same pass. Dialogue, effects, music, all together. You can run it locally through ComfyUI checkpoints up to fifteen seconds at 768p, or go hosted for 2K output. Workflow templates are included for joint visual-audio production.
Alibaba is also moving. Wan 3.0 just went into public beta. It handles video with audio up to thirty seconds, across 480p, 720p, and 1080p. The pricing structure is worth knowing if you're building with it. Four cents per second at 480p, eight cents at 720p, seventeen cents at 1080p. There's a catch though. If you use a reference video, you're charged for its full duration on top of your output. Ten seconds of reference footage plus thirty seconds of 1080p output lands you around six dollars and seventy cents. The platform also handles character, object, and scene consistency, plus video continuation and automatic duration selection.
And Dreamina launched Seedance 2.5 globally with something called Smart Edit tools, plus native three-minute video generation. They're positioning it as the first US platform offering the full Seedance lineup, and they're doing it at the lowest pricing available for that model family.
Moving into the business layer of open-source AI, Alibaba is planning to take a revenue cut from heavy users of the next open-source Qwen version. Moonshot is doing something similar with Kimi K3, seeking up to thirty percent from partners. The open-source model landscape is starting to look a lot more like a licensing negotiation than a free-for-all.
Speaking of models worth watching, polls circulating among developers are naming DeepSeek v4, Kimi K2.6, GLM-5, and Qwen 3.5 as the top open-source options right now. That's the short list people are actually reaching for.
Switching to tooling, OpenRouter shipped something called ORI Harness. It auto-configures Codex, Claude Code, OpenCode, and Hermes with zero manual environment variable setup. Any model on the platform works with it. You can swap models without changing your workflow. That's the whole pitch, and it's a practical one.
There's also a strange and interesting project from the co-founder of Octra. A graph-based LLM, version 0.0.1 pre-alpha, is now live on the Circles devnet. The weights are 409 megabytes in int8. The architecture uses a single root circle graph that anchors thirteen resource circles, across 291 tensors and 313 source chunks. It supports public verifiable inference, agent tools, and private chat via something called PVAC private flow. It also supports Tor-based internet access and customizable privacy settings through HFHE and CircleGate. Pre-alpha is not an understatement here, but the architecture is genuinely unlike anything in the mainstream right now.
Poolside's Laguna S 2.1 crossed two trillion tokens served in fourteen days after release, running at around three hundred billion tokens per day. Total across all Poolside models, ten trillion tokens in under three months through OpenRouter, Vercel, and direct API. Those are not small numbers.
Vercel also pushed two meaningful updates. The AI Gateway now emits per-request OpenTelemetry traces covering routing, timing, token usage, and attribution. You can export them via Drains to OTLP endpoints with sampling controls. It deliberately omits prompt and completion content. Separately, a project called rareui.com hit Vercel's free bandwidth limit and Vercel stepped in with six months of Pro credits and an OSS Program invite.
On the agent side, Claude Code sessions can now message each other. One session working on a feature can send a plain summary to a second session working on a different part of the same project. The receiving session picks up mid-task from that summary alone. No file transfers, no shared history. It's a clean coordination pattern that sidesteps a lot of context bloat.
Teknium added a command called /learn to Hermes Agent. You feed it any PDF or book, and it converts the content into a structured technical skill file the agent can work from. Practical and quietly powerful.
Untrivial AI open-sourced an agent orchestrator that now has over eight thousand seven hundred stars. It runs Claude Code, Codex, Cursor, and Kimi Code as a parallel fleet. One supervisor spawns workers, each gets its own branch, its own terminal, and an isolated git worktree. When CI fails, the failure routes back to whichever specific agent wrote the problematic code. That feedback loop is the part worth paying attention to.
Mark Jeffrey ran Hermes Agent on Engy with Kimi K3 underneath it and produced a report covering eighteen of Bittensor's one hundred and twenty-eight subnets. One prompt, one model chain, a structured research output across a complex ecosystem.
A developer named Da7em built a complete game inside Hermes Agent. MiniMax M3 generated the backgrounds, DeepSeek Flash handled the design and code for everything else. The whole thing stayed inside one agent environment.
And finally, both Kiro and Hermes Agent have added support for the new Agent Plugins 1.0.0 spec. That spec covers plugin.json, a skills directory, and mcp.json. AWS, Cursor, Microsoft, OpenAI, and Vercel are all aligned on it. Standardized plugin architecture across that many platforms is worth tracking.
That's your AI digest for 08 Aug 2026.