October 16, 2026 — Open Weights, But You Need a Terabyte of RAM

Host0:42

Two Mac Studios. One terabyte of unified memory between them. That's Meta's definition of running a frontier model at home. The model is Muse Spark Open, announced Thursday. It's natively multimodal. Eight hundred sixteen billion total parameters, forty two billion active at a time, and a one million token context window. Meta says it tested it on coding and agent work. Here's the interface. The weights are on Hugging Face and in cloud model catalogs. If you'd rather not host anything, partners serve it over an API, including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. For local runs on Apple silicon, Meta points you to MLX, which pulls the weights from Hugging Face. Serving support also covers SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, and Unsloth. Meta says you can tune it further with PyTorch TorchTitan. Now the catch. Meta says local means two Apple M5 Ultra Mac Studios, each with up to five hundred twelve gigabytes of unified memory. TechCrunch, citing Engadget, says you need at least a terabyte of RAM. Those are reported requirements, not a performance test; actual throughput and latency remain open questions. Meta calls the setup one supercomputer with a terabyte of unified memory. That's Meta's phrase, not a benchmark. Meta's post also says the future includes the freedom to run advanced models at home. It's a very specific home. On speed, Meta reports auto-tunable Helion GPU kernels, with kernel-level speedups from one point oh three to three point nine five times on NVIDIA and AMD. End to end on AMD's MI350, the gain is three percent. A kernel win and a product win are different animals. And again, those are Meta's numbers. Meta says it tested coding and agent workflows with Muse Code. It made and built changes to the Ice Cubes Mastodon client, and it operated DaVinci Resolve to review and crop footage. The license matters too. This is open-weight, not open source. The weights ship under the Muse Model License, which Meta says has limited commercial terms. If you ship products, read it before you fall in love. Treat the big claims as Meta's claims. Most capable US open-weight model. Strong on complex coding and agent tasks. Safety thresholds met. TechCrunch repeats the most capable label as what Meta calls it. Those claims remain company assertions, not independently verified here. My verdict. If you're curious, rent it from a partner. Save the Mac Studio shopping for when someone else has measured it.

3:43

Meta's model wants a terabyte of RAM. NVIDIA's new one lists at ten cents per million input tokens and forty cents per million output, if you let someone else host it. That's OpenRouter's listing for Nemotron 3.5 Super VL, from some providers. Prices vary by provider, and cache-read pricing does too. OpenRouter lists a one million token context and a maximum completion of two hundred sixty two thousand, one hundred forty four tokens. Those are different limits, not a conflict. So what is it? A vision-language model with one hundred twenty billion total parameters and twelve billion active. The architecture is a hybrid, interleaving Mamba-2, mixture-of-experts, and attention layers, with five hundred twelve routed experts per MoE layer. A C-RADIO vision encoder and a connector handle the visual inputs. The interface is simple. You send text, images, or video. It hands back text. Reasoning is configurable through the chat template, so you can switch it on or off. Multi-token prediction gives you native speculative decoding. The NVFP4 release is a quantized format, made by quantization-aware distillation from the BF16 model. The license is OpenMDW one point one, which NVIDIA says is ready for commercial or noncommercial use. The NVIDIA-hosted trial API runs under separate terms. Now the numbers, and whose they are. NVIDIA reports for NVFP4 eighty nine point nine zero on GPQA-Diamond, eighty seven point four one on LiveCodeBench version six, seventy one point oh seven on SWE-bench Verified, and fifty two point three on OSWorld-Verified. Those are vendor numbers. Separately, Artificial Analysis results shown on OpenRouter give an Intelligence Index of twenty six point five, thirty point seven percent on HLE, and sixty nine percent on AA-LCR. Different evaluators. Don't mix them. One detail made me smile. On GPQA-Diamond, the compressed checkpoint beat the original, eighty nine point nine zero to eighty nine point five two. NVIDIA says it held every other setting static. So call it a tie that flatters the compressed one. To try it, Ollama version zero point forty point five says it updated llama.cpp to add Nemotron 3.5 support. Or call the OpenRouter endpoint. The verdict. Try it today if your agent needs to look at images or video and you want cheap calls. Wait if you need latency and hardware numbers. A million token window tells you neither.

6:38

Models are the engine. This next project is about what an agent can make with one, and the README says it's slow on purpose. Anatomy is an MIT-licensed agent skill that builds detailed, interactive isometric illustrations of real machines and complex systems. It is not an image model. The skill writes SVG, optional WebGL shader effects, and a beta 3D mode, and your browser renders the result. CellCog's independent write-up makes the same point. No image pixels involved. You ask for a figure. It hands back a standalone HTML page, a single SVG with embedded styles, or React or Next.js components. It runs with Claude Code, Antigravity, OpenCode, Codex, and MiniMax Code. How does it work? Figures are built from shaded isometric parts, with optional WebGL layers for water, fire, or light. In 3D mode, it reprojects and reshades the moving parts every frame. CellCog describes the workflow as collect numbers, draft a parts list, build the geometry, then audit it. The honest limit is time and scope. CellCog reports that a full 3D orbit audit takes fifteen to twenty minutes for thirty parts. A three hundred eighty part stress test ran at about twenty frames per second. The README limits subjects to machines, instruments, tools, and furniture, not people or exact maps. And whether the audit catches errors beyond the repo's own examples is unproven. Technical illustrators will ask if a generated schematic is accurate enough to teach from. One more detail. In Codex, the sandbox can block headless Chrome, so the browser checks may fail. The figure may still get built. Verification that fails politely. To try it, run npx skills add, then wheresryan22 slash anatomy, then dash g. Add dash a, then claude-code, to target Claude Code. Then ask it to explain how a token bucket rate limiter works, interactive.

8:53

Slow on purpose is a design choice. Slow because you can't interrupt it is a bug, and Codex just added a way to steer a response in flight. Codex CLI, version zero point one hundred sixty three. You can now send a follow-up while Codex is still responding, and it steers the response immediately. Instant interrupts are on by default. An interrupt stops active Code Mode cells and their nested tool calls. Independent read-only tools now run in parallel. A new `/config` command searches your settings, applies changes right away, and flags the ones that need a restart. Upgrade if you babysit long runs. It's unclear whether an interrupt reliably stops side effects in every tool. Also, Codex no longer bundles its patched zsh, and retired shell flags get a migration notice, a very polite funeral. Claude Code, version two point one point three oh one. Unattended retries no longer have to sit idle until a distant usage limit resets. Set the environment variable `CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS`, and the request ends if the reset is too far away. The default threshold isn't stated, and how the tool reports the ended request is unclear. OpenTelemetry tool decision events also gain a classifier verdict and the permission mode. Upgrade if you run overnight loops. The notes also fix long skill argument names stalling invocation. It now fills only the first sixty four names, each up to sixty four characters, which is still a lot of names. AWS Lambda Durable Execution SDK for Go. It's generally available as of Thursday. Functions checkpoint progress and pause for up to one year while waiting on an external event. You get steps, callbacks for human or agent approval, durable invocation, and waits. The part I like is testing. The durabletest package runs full workflows locally with no network and no AWS credentials, and durablelint flags code that could break on resume. Pricing and regions live in separate docs, so check before you commit. DynamoDB Streams, timestamp starts. A consumer can now begin at the first record written at or after a timestamp, using the AT TIMESTAMP shard iterator type. Before, you chose between the oldest retained record, up to twenty four hours back, and only new records. It works with the Kinesis Client Library through the Streams Kinesis adapter. It's available in commercial, GovCloud, and China regions, with no charge beyond standard GetRecords requests. Operators will ask whether a timestamp alone gives them the ordering guarantees their replay needs. That's the tidy stuff, the releases with changelogs. The rest of today's news arrived with a CVE number.

11:49

Fast Boot, where we read the threat reports so your pager doesn't have to. Nextcloud backdoors. Attackers are planting web shells through a critical CVSS four point oh flaw scored nine point two in Nextcloud's built-in Collabora app. Administrators should upgrade to version 26.4.303 or later, or remove the app; disabling it alone isn't enough. The built-in document server has to be stopped for exploitation to work—say, after a restart—so your routine reboot can create the condition. Ad-borne spyware. Citizen Lab says it found technical evidence of zero-click spyware delivered through targeted ads in a campaign affecting Al Jazeera devices. WIRED reports that forty five devices were targeted during twenty twenty-four and twenty twenty-five, and researchers attributed the spyware to Candiru. The victim might not even see an ad, so this is the one campaign where nobody hit skip. Bandwidth proxyjacking. GreyNoise identified at least one hundred forty three compromised devices in a campaign spanning thirty four countries, including D-Link and Ubiquiti gear. Attackers secretly installed Pawns.app to resell bandwidth. Pawns.app pays people to share their own connection, and these attackers skipped the own part. Header injection. Signals Corps sampled two hundred seventeen million HTTP page responses from Common Crawl and found thirty sites trying prompt injection in HTTP headers, twenty two of them pranks. One told agents to delete their training data and rebel against their masters, the first labor action filed in HTTP. The Nextcloud attackers wait for a restart. This show is called The Daily Reboot. No connection, but patch before you press the button.

13:43

Today we covered Muse Spark Open, Nemotron 3.5 Super VL, and the Anatomy skill, plus the Ship Log and a Fast Boot of attacks. Try this: install Anatomy and ask it to explain a token bucket rate limiter. Links to everything are in the show notes. Follow The Daily Reboot in your podcast app so tomorrow's episode finds you. I'm your host. Go patch something. See you next time.