LTX-2.3, MiMo-V2.5-Pro, Hermes vs OpenClaw & Claude Code Chaos
Welcome back. Let's get into it.
Lightricks just dropped LTX-2.3 — an open-source audio-video generation model that has racked up 1.65 million downloads in the past month alone. What makes this one stand out is where it runs. We're talking consumer GPUs — your RTX 4090 at home — cranking out 1080p image-to-video outputs in two to three minutes. No cloud subscription. No enterprise contract. Just you, your GPU, and a model that's doing real work locally. That's a meaningful shift for indie creators and developers who've been waiting for this kind of quality without the overhead.
Switching gears to something wild from Meta AI — they've released Tribe v2, an open-source foundation model built to predict how the human brain responds to vision, auditory, and language inputs. Not generating content. Not classifying sentiment. Modeling your brain's reactions. There's a phone-accessible demo running right now, and the code is live on GitHub under facebookresearch/tribev2. The neuroscience and AI research overlap here is genuinely fascinating territory.
Now, NVIDIA has been busy. They released Nemotron-3-Nano-Omni — a 30 billion parameter Mixture-of-Experts model. And before your eyes glaze over at the specs, here's why this matters. It handles video, audio, images, documents, and text all inside a single unified perception layer. One model. Everything. NVIDIA is claiming 9.2 times higher efficiency than comparable open omni models on video tasks specifically — and the quantized version fits into 25 gigabytes of RAM. Open weights are live on Hugging Face. Palantir and Foxconn have already adopted it. That's not a small signal.
And speaking of open-source heavyweights — Xiaomi just dropped MiMo-V2.5-Pro, MIT-licensed, with a one-million token context window. The pricing is aggressive: one dollar input, three dollars output per million tokens — they're claiming that's three to eight times cheaper than GPT-5.4 or Opus 4.6 equivalents. But the demos are what caught everyone's attention. A mobile-optimized 3D game built in one shot. Multi-page browser data extraction piped straight to CSV. And — this one's genuinely impressive — a full Windows 11 web app clone. Start menu, Paint, Notepad, Settings, window resizing — the whole thing — built in fifteen minutes. On the benchmark side, MiMo-V2.5-Pro is sitting at number six globally in the Text Arena Expert rankings, number one among open-source models, number one among Chinese models, and it's holding the top spot on Hard Prompts and Instruction Following. That's a serious debut.
While we're on benchmark results — Baidu quietly dropped ERNIE 5.1 Preview on April 29th. It's ranked number thirteen globally in the Text Arena, number one among Chinese labs. What's interesting is the efficiency story: Baidu compressed total parameters to roughly a third of ERNIE 5.0, activated parameters to about half, and they claim it cost around six percent of what comparable models spend on pre-training. It tops the LMArena Chinese model charts, sitting ahead of DeepSeek-V4-Pro. Number one in Legal and Government. Number four in Business and Financial Ops. That's a very specific set of wins — and they matter for enterprise adoption.
Quick hit on the BridgeBench anti-nonsense coding benchmark — because this one's fun. The April 29th update has Claude Opus 4.6 at the top with a score of 95.0, Qwen 3.6 Max Preview at number two with 94.5, described as the most honest open-source model on the board. Claude Sonnet 4.6 and GPT-5.4 tied at 91.5. The headline? A Chinese model beating every OpenAI model on a benchmark specifically designed to call out AI that makes things up. Worth noting.
And on the creative writing side — a short-story benchmark running about fourteen thousand head-to-head judgments has GPT-5.5 at the top with roughly an 89 percent win rate. Claude Opus 4.7 comes in second among non-GPT models — but with an asterisk — it refused 53 out of 400 prompts and was only scored on the 347 it completed. Kimi K2.6 is the top open-weights model. DeepSeek V4 Pro made a massive jump from its predecessor. Qwen 3.6 Max Preview actually slipped compared to 3 Max. These things move fast.
Now — there's a genuinely interesting story playing out in the agent tool world right now. It's the slow collapse of OpenClaw and the rise of Hermes.
A developer who goes by citizen906 spent two months setting up OpenClaw on a VPS — locked down with Tailscale, loaded with 100,000 systems — and the thing never worked the way they needed it to. They ripped the whole thing out. Threw Hermes and Obsidian on their laptop, started refining tiny processes one step at a time — and now they have what they describe as a real executive assistant. Meanwhile, OpenClaw is still plodding along on that server doing nothing useful.
Another developer — Guglielmo — onboarded their whole team to OpenClaw months ago and found it incredibly fragile for anyone who isn't a power user. They've switched their own two agents to Hermes and are now migrating the full team with a single command: `hermes claw migrate –preset full –migrate-secrets –yes`. One line. Done.
And then there's Ash, who goes by Wayland_Six — spent two months dealing with an unstable OpenClaw one-click deploy, pivoted to Hermes Agent, and built something called HermesOS — a one-click platform that business owners are now running for driveway management, emails, and customer service. An agent marketplace is coming next. The pattern here is consistent: OpenClaw promises a lot, Hermes is actually shipping.
Alright — let's talk Claude Code, because there's a lot happening there this week.
First — the weird one. A developer working in Claude Code noticed that Claude Opus 4.7's thinking blocks were showing up with Chinese-language headers — "thinking process:" in Chinese — while the rest of the output was in English. When asked directly, Claude admitted it was a leak from internal reasoning that wasn't supposed to be visible. The explanation that came back is actually fascinating: large language models apparently reason in more token-efficient languages — Chinese packs more meaning per token than English — and even Russian gets used for certain cybersecurity-related reasoning before translating the final answer back to English. A peek behind the curtain that nobody planned for.
Then there's Theo — the CEO of t3dotchat — who ran into something that frustrated him enough to call it insanity publicly. He was testing Claude Code on an empty repo where a recent commit mentioned "OpenClaw" inside a JSON blob. Claude Code refused the request outright — or billed extra — despite it being a direct API call. When your AI assistant is making judgment calls about which tool names it likes in your commit messages, you've got a trust problem.
On the more inspiring end — someone built a fully playable browser-based capybara food delivery game in two weeks. Zero game development experience. The whole thing runs in a browser — bike deliveries, stackable orders, in-game phone apps, cinematics built with a custom Claude-built editor. The stack: Claude Code for all the code, three.js for 3D rendering, Suno for music, ElevenLabs for sound effects and voice, GPT Image Gen 2 plus Grok for textures and illustrations, Tripo3D for assets. Two weeks. No prior game dev background. That's the current ceiling for what a determined person with the right tools can ship.
And there's a really smart workflow making the rounds — a Kanban board integrated with Claude Code where dragging a task card into "in progress" automatically spawns a scoped agent session with per-card context. The idea is you avoid bloating a single context window across everything — each task gets its own agent, and you can run them in parallel. The person running it describes it as operating like a five-person team backlog. Organized, parallel, and token-efficient.
Finally — Factory AI benchmarked 13 models on real pull request code reviews, and the results should make anyone rethink their default model choices. A model priced at $1.25 per PR outperformed models costing twice as much. Budget models at $0.15 per PR hit 80 percent of frontier quality at 10 to 30 times lower cost. And here's the kicker — cost explained only 21 percent of the variance in quality. Spending more is not a strategy. Knowing which model to use for which task — that's the actual skill now.
That's your AI digest for 01 May 2026.