October 7, 2026 — Le Chonk Is Huge. Is It Any Good?

Host0:44

Mistral calls its new model “unofficially ML4, very officially: le Chonk.” It's Mistral Large four, announced Tuesday as a public preview. One trillion parameters, natively multimodal, with Mistral saying its weights will be released by the end of the month. Here's the part your GPU bill cares about. It's a mixture-of-experts model, so it activates forty-nine billion of those trillion parameters at a time. Most of the chonk is napping. Mistral says it trained from scratch on thirty-eight hundred NVIDIA Grace Blackwell GPUs in European data centers, across more than a hundred sixty languages. It hasn't said how long the run took. Its VP of Science, Pierre Stock, told CNET the focus is agentic capabilities. The scores are Mistral's own. Sixty-one point seven percent on DeepSWE version one point one. Twenty-eight point three on Terminal-Bench four. A Coding Agent Index of forty-nine point eight, which Mistral says beats DeepSeek V four Pro zero eight one three and Qwen three point eight Max. Now the unflattering number. The Register reports that Artificial Analysis, an independent firm, ranks the preview below leading OpenAI and Anthropic models, though ahead of Thinking Machines Lab's Inkling. Different tests, so Mistral's claims stay contested until someone reproduces them. Pricing is one dollar thirty-six per million input tokens and four dollars eighteen per million output, through the preview API in Mistral Studio. The pitch is self-deployment, European infrastructure, and fewer provider-level blocks on legitimate cyber-defense work. Security teams will ask how that gets governed. So watch the end of the month for the promised weights; the license and serving requirements remain questions. Until then, test the preview on your own agent tasks, not Mistral's. That's a trillion parameters you can't download yet. Google's news is the opposite, seven hundred forty million, and its weights are available today.

3:00

One hundred ninety-one megabytes. That's the active RAM Google reports for quantized, text-only EmbeddingGemma two on a Pixel eleven Pro. The full multimodal model needs about five hundred sixty-seven. You send it text, images, audio or video. It hands back a seven hundred sixty-eight-dimensional vector in one shared space, so a text query can search image, audio and video embeddings directly. It's seven hundred forty million parameters across text, vision and audio encoders. Weights are on Hugging Face and Kaggle under Apache two point oh, and the model works with sentence-transformers, vLLM, llama.cpp and Ollama. Hosted pricing is unknown. The eight thousand one hundred ninety-two-token budget is shared across modalities. Alone, that's about twenty-nine images or three hundred twenty-seven seconds of audio. Mix them and they compete. From Google DeepMind's own model card: MTEB Code NDCG at ten of seventy-eight point six eight, up from sixty-eight point seven six. Multilingual is nearly flat. No independent comparative evaluation was found in the sources read. Matryoshka truncation lets you cut vectors to five hundred twelve, two hundred fifty-six or one hundred twenty-eight dimensions. The model card says two hundred fifty-six dimensions retain relatively close quality. At one hundred twenty-eight, multimodal quality drops, while Google reports up to a sixfold reduction in vector-storage needs. The catch is quiet. In float sixteen, it can return NaNs or degraded embeddings without raising an error. Your pipeline looks healthy while your search rots. Use bfloat sixteen or float thirty-two. Try it today for local, privacy-sensitive retrieval, like text-to-code search or a small offline media index. Wait if you have a production index, because embeddings from different models aren't interchangeable. Install with `pip install -U sentence-transformers transformers`, then load `SentenceTransformer("google/embeddinggemma-2")`. An embedding model that indexes your media locally has an obvious companion, something to edit that media. A Rust project is trying.

5:23

Six thousand six hundred stars, and the README admits up front it's not ready for your day job. PhotoCraft is a clean-room, open-source Photoshop reimplementation in pure Rust, dual-licensed MIT or Apache two point oh. It lists layers, masks, adjustment layers, layer styles, type, vectors, brushes and PSD support. The design is engine-first, a command engine under a thin egui interface, and the UI, a CLI, a JSON channel and an MCP server all share the same commands. The number: re-saves match Photoshop's render for three hundred seven of three hundred nine files in the psd-tools test set. That's the project's own result. The limit is theirs, not mine. The maintainers say it's early alpha, with no generative features, about twenty missing tools, limited typography depth and no plug-in compatibility. Re-saved PSDs aren't byte-identical either. To try it, run `git clone https://github.com/storytold/photocraft`, then `cargo run --release -p photocraft -- image.psd`. PhotoCraft listed its limits in the README. OpenAI's math release makes you go find them.

6:37

OpenAI released seven hundred twenty-two manuscripts, grouped into three hundred seventy-two result families, according to The Verge. They sit in a GitHub repository with paper-revision and citation protocols. The mechanism that matters is Lean, a proof-assistant language where a computer checks a formal proof mechanically. OpenAI says many proofs are formalized in Lean, with more coming. The repository also has ten reasoning summaries and compute estimates. The average result used roughly three hours of ChatGPT Pro thinking, OpenAI says. Math effort, now measured in a subscription tier. WIRED reports mathematicians criticized labs for announcing results through company posts. Northwestern's Bryna Kra said, “Math by tweet and math by press release to me is not the way to nurture the ecosystem that created the fertile ground that they have trained on.” OpenAI said it was consulting an independent advisory group. And a Lean check verifies a formal proof of a statement. It doesn't settle novelty or attribution, and the count itself is OpenAI's claim about its own model. So the number to watch isn't three hundred seventy-two. It's how many of them are Lean-verified, and how many are new. OpenAI's claims need an outside checker. So do Anthropic's, and these are about who gets to point its strongest models at real software.

8:12

Every attempt without program access was blocked on the first prompt, according to SiliconANGLE's account of Anthropic's test. That's the pitch of its expanded Cyber Verification Program, announced Tuesday with three tiers. Defense Access covers defensive work, Red Team Access covers authorized penetration testing, and Specialized Access covers especially sensitive infrastructure. Existing Glasswing members move to that last tier. In the test, Claude Opus five point five ran ten offensive challenges, five times each. At Defense Access, forty-six of fifty runs were blocked at some point. At Red Team Access, none were, and the model completed thirty-four. Higher tiers aren't a free pass. Requests heading toward ransomware deployment can still be blocked in real time, and enrolled organizations must allow data retention so Anthropic can monitor misuse. Anthropic's evidence is partner-reported. Glasswing partners found at least one hundred twenty-nine thousand verified vulnerabilities from April to July. Those figures draw on thirty-three partner reports, fewer than half disclosed patch counts, and the reports are not a complete independently audited accounting. It's available on the Claude Platform, Vertex AI and Microsoft Foundry. Excluded researchers will call it too restrictive, and privacy advocates will question the retention. Watch patches landed, not bugs found. Finding bugs makes the headline. Proving they're real is the job, and that's what Google built PageBreak around.

9:57

Google says its internal PageBreak agent found more than five hundred cross-site scripting vulnerabilities across its first-party web apps. The interesting part is what the agent isn't allowed to do. It doesn't get the final word. The model proposes a flaw. Then a separate validator, not written by AI, runs a real payload against the running application, like JavaScript for cross-site scripting. Only then is it confirmed. The model suggests, conventional code checks. The AI gets to accuse, but not to convict. Google says its high-assurance web frameworks had just two XSS findings across hundreds of applications as of September fourth, both in internal apps or debug endpoints. Dark Reading lists fixed findings including an XSS issue at admin dot google dot com, but said Google had not immediately responded when asked whether every finding was fixed. The limits are Google's own. Validators don't cover every vulnerability class, so some real flaws can slip past, and near-zero false positives is Google's claim. If you build agentic scanners, steal this design. Validators that don't trust the model are one answer. A permission prompt the tooling could skip is the opposite problem, and Claude Code just fixed one.

11:26

Claude Code version two point one point two nine two. The fix that matters: PreToolUse hook approvals and auto mode could bypass the permission prompt for file reads from network UNC paths. Upgrade if you use either. It also adds a marketplace flag to plugin install, one step instead of two, under the same policy checks. A convenience flag and a bypass fix in one release is a balanced diet. OpenAI Decisions API, public beta. Send evidence and a question, get a typed answer back: a probability, a choice from your values, or a score against ordered levels. Only GPT six luna is supported, at ten cents per million input tokens and nothing for output. OpenAI claims it's about ten times faster than the Responses API, so measure that yourself. Use it for classification and routing. GitHub stacked pull requests, generally available. Each pull request in a stack targets the branch below it, and each is checked against the stack's base. GA keeps approvals alive through rebases when code is unchanged. It needs linear history, with no cross-fork stacks and no GitHub Desktop. GitHub says repositories using stacks during preview saw nine percent more merged code compared with peers, with no methodology given. Try `gh stack init --base release auth-layer`.

12:52

Fast Boot, the part of the show where the padlock icon does not mean what you think. Counterfeit certificates. Attackers hijacked the dot gh, dot sl and dot as registries, rewrote DNS, and passed automated validation to get HTTPS certificates for Google domains. Chrome blocked the unauthorized certificates, and Google adds that you shouldn't rely on that, a fix with a built-in disclaimer. Atlassian file access. Atlassian rates a file-access flaw critical, nine point three, affecting every version of eight self-hosted products. Attackers must already know the exact filename, so it's a guessing game that somehow earned a nine point three. Fake AI ads. Island researchers found fake ad products for ChatGPT, Gemini and Claude, using painted login windows to harvest passwords and MFA codes from ad agency staff. A human operator picked which MFA challenge a victim saw next, an alarmingly attentive phishing operator. A counterfeit certificate and a counterfeit address bar, all in one day. The padlock had a rough Tuesday.

14:05

That was Mistral's le Chonk, Google's EmbeddingGemma two, the PhotoCraft repo, OpenAI's math release, Anthropic's cyber tiers and Google's PageBreak agent. One thing to try: load EmbeddingGemma two in bfloat sixteen and run it on your own queries, not Google's. Every source is linked in the show notes. Follow The Daily Reboot wherever you get your podcasts, so tomorrow's news finds you. I'm your host. Keep your padlocks honest, and we'll reboot tomorrow.