Meta opens 816B Muse Spark Open, Claude Opus 5.5 fast mode leaves preview, Anthropic embeds METR evaluators
Today’s threads are openness and verification: Meta and NVIDIA release large model weights, GitHub has agents rerun the scan to prove a security fix worked, and Anthropic embeds outside evaluators who publish on their own. Capability keeps growing, and so do the ways to prove it holds up.
About 816B total parameters with about 42B active per token; 256 experts per layer, 8 picked per token; up to 1M tokens of context.
Weights come in BF16 (about 1.5 TB) and NVFP4; NVFP4 runs in Transformers on a single B200 node, while BF16 needs four 8×H200 nodes.
Reasoning cannot be turned off; effort levels are low, medium, high and xhigh, with xhigh as the default.
A 2-bit llama.cpp quant ran at 18 tokens/s on a 512 GB Mac M3 Ultra using about 304 GB; the post gives no benchmark scores and does not state the license.
Builder's takeThe Hugging Face post calls it the most capable US open-weights model to date, but for a small team like mine it is a frontier model you can self-host rather than one you casually run: a single B200 node is the entry point. I’ll wait for a clear license and independent evals before comparing it with closed APIs on AI Cloud Drive’s long-document Q&A; and since reasoning can’t be turned off and defaults to xhigh, work out latency and token cost first.
Fast mode for Claude Opus 5.5 no longer needs an access request, and requests no longer need the fast-mode-2026-02-01 beta header; requests that still send it keep working.
Claude Opus 5.5 now has a default fast mode rate limit in every usage tier.
Fast mode for Claude Opus 5 and Claude Opus 4.8 stays in research preview and still needs the beta header.
The same day, Opus 5.5 fast mode became available on Claude in Microsoft Foundry for deployments hosted on Azure.
Builder's takeFast mode used to require an application, so I only tried it in demos; now that it’s on by default, it can go into production paths. I’d turn it on only where a user is waiting live, such as follow-up question generation in AI Interview, and keep batch jobs on the standard tier; the release note doesn’t mention price, so check billing in the Console first. Read alongside the 10-11 GPT-6.1 Sol Ultrafast item, charging by speed tier is becoming standard across vendors.
Anthropic has onboarded six METR employees with company hardware and employee-like access; METR will publish its risk assessments.
Anthropic cannot veto publication and can only require redaction of narrow, predefined categories such as customer data, privileged information or model architectures; METR may comment publicly on whether redactions were justified.
Evaluator access has three tiers: SecureBio is in the first, reviewing the chemical and biological weapons chapter of the August Risk Report; Accenture with Faculty is working toward a partnership, and Hugging Face is in early discussions.
It follows Anthropic’s signing of the White House Accord on Super Intelligence, which included pledges to work with external auditors or evaluators.
Builder's takeFor product builders the practical value is one more risk report on the model you pick that the vendor didn’t write. When METR’s first report lands, I’ll check its findings on agents overstepping against our own permission design. Fixing in advance what may be redacted is also worth copying: when we go through customer security reviews, agree up front what won’t be disclosed.
Nemotron 3.5 Super VL is an open 120B-parameter model that NVIDIA says matches its largest reasoning model at about a fifth of the parameters and runs 5x faster than the next best in its class.
Super VL was trained on about 26 trillion tokens of pretraining and continued pretraining, then about 3.8 million post-training samples plus RL; weights, datasets and recipes are open, and it is on Hugging Face, OpenRouter, ModelScope and NVIDIA NIM.
Nemotron 3.5 Bolt 74B-A9B fits in 24GB using native 2-bit quantization and is on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
A smaller Bolt E4B 4GB arrives October 23; the article does not name a license.
Builder's takeIn the 10-14 issue this was still a preview; today Bolt is actually on Hugging Face, and a 74B agent model runs in 24GB. I’ll test Bolt locally on PandaClaws’ batch rewriting to see how much 2-bit quantization costs on long Chinese text; Super VL, with its datasets and recipes, suits teams that want to fine-tune. Until the license is clear, evaluate only and keep it out of production.
From November 16, Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra are retired across all Copilot experiences.
Suggested replacements: Claude Opus 5.5 for both Opus models, and GPT-6.1 Sol or GPT-6 Luna for GPT-5.6 Terra.
Claude Opus 4.8 stays available to individual Copilot subscribers on annual plans.
Under default model enablement, Business and Enterprise get the replacements turned on automatically unless an admin has turned off the global default or disabled the model.
Builder's takeTwo things to do now: search your repos for Copilot configs, custom agents and scripts that hard-code these three model names; and if your org turned off default enablement, turn the replacements on by hand, or after November 16 there will be nothing to fall back to. Before switching, compare old and new models on one or two real tasks for output and usage.
It scans pull request changes in languages and frameworks CodeQL doesn’t cover yet, including PHP, Shell/Bash, Terraform (HCL), Dockerfiles, JSP for Java and Blazor for C#.
Findings show on the Conversation and Files changed tabs with an AI indicator and can include a risk explanation and suggested autofix; they don’t fail the code scanning check and can’t block merges through rulesets.
It is available on public repositories owned by organizations or personal accounts; private and internal org repositories need GitHub Code Security or GitHub Advanced Security, and no Copilot license is required.
It doesn’t run on pull requests from forks or from Dependabot; it consumes Code Security AI Credits, with SKU-level budgets available.
Builder's takeOur deploy scripts, Dockerfiles and Terraform are exactly what CodeQL never scanned, and public repos can turn this on now. I’ll enable it on this site’s public repo first to see what it flags in the Shell steps of our GitHub Actions; since it only advises and never blocks, log false positives and misses for two weeks before deciding whether to wire its results into review rules.
The agent explores relevant files, proposes a fix, reruns CodeQL to confirm the alert is closed and opens a draft PR with a summary, why the fix works and validation steps; you can ask Copilot for more changes in PR comments or the Agents tab.
You can assign any alert on the default branch to Copilot (replacing the free Generate Fix button), select up to 25 alerts to fix in one PR, or set assignees to copilot-swe-agent[bot] via the REST API.
It is for GitHub Code Security and GitHub Advanced Security customers on GitHub Enterprise Cloud, including data residency; an enterprise admin must enable the policy and a Copilot billing method must be set.
It uses AI Credits from the Code Security budget plus GitHub Actions minutes; alerts found only on PRs or non-default branches can’t be assigned.
Builder's takeRerunning the scan to prove the alert is closed is the part I value most: it gives agent fixes a verifiable definition of done for the first time. But it replaces the free Generate Fix and bills AI Credits plus Actions minutes, so I’d batch-assign a set of low-risk old alerts first, check PR acceptance and cost per alert, and then decide whether new alerts go to it by default.
You can send a follow-up mid-response to redirect Codex right away; interrupting also stops active Code Mode cells.
Independent read-only tools such as skill, memory and history lookups can run in parallel; /config is searchable and flags settings that need a restart.
The network proxy checks permissions before DNS lookups, rejects duplicate Host headers and closes a credential-alias bypass; resuming a saved thread restores its permissions and network approvals.
Daybreak is on by default for eligible accounts.
Builder's takeIf you run Codex in CI, upgrade for the three network proxy fixes rather than the features: a DNS lookup can leak information on its own, and a credential-alias bypass is worse. Also, resuming a thread now brings back its network approvals, so after upgrading check which domains old threads had approved.
The pilot began in July 2026 with a cohort of about 1% of Cloudflare’s network and helped OpenAI discover and index hundreds of millions of new pages.
Signals filter out missing, redirecting and duplicate pages and indicate when a page is worth revisiting; harmful pages can be flagged individually instead of penalizing a whole domain.
OpenAI independently ranks how those pages appear in ChatGPT; the research uses web metadata, not page content for model training.
A signal doesn’t grant access, and site owners can still manage AI access with Cloudflare’s AI Crawl Control; the next phase will improve page selection and test faster change signals.
Builder's takeWith PandaClaws, what I care about most is how quickly published content becomes findable; this shows AI search indexing bypassing the classic sitemap and reading change signals from the CDN. If your site sits behind Cloudflare, check in AI Crawl Control that your policy for OpenAI’s crawler is what you intend; content sites should cut duplicate pages and pointless redirects, which directly hurt discovery.
Four sizes, 5M, 28M, 123M and 209M, share one architecture and recipe; they produce probabilistic multivariate forecasts with 21 quantiles, take past-only and known-future covariates, and have a 32,768-step context.
The 209M model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench with a single checkpoint and no ensembling; the 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME.
On an 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
The license is Modified MIT, with a GitHub repo and Hugging Face collection, and it can be used and fine-tuned through Mistral Forge.
Builder's takeCapacity planning and usage forecasting are things we’ve handled with rules or not at all. A zero-shot model that runs in about a hundred milliseconds on a CPU is worth a try: I’ll backtest it on AI Cloud Drive’s storage growth and model call volume from the past few months to see if it can replace our guesswork scaling thresholds. Read what the Modified MIT license changes before going to production.
Researched and drafted with AI assistance; editorial standards and views set by Darius.