Black Forest Labs opens FLUX 3 Image Klein weights for 8GB GPUs
A 5.6-billion-parameter open-weight image model shares native 4K generation and pixel-perfect editing with the full FLUX 3 line, and runs on consumer GPUs.
What shipped, what broke, and what it changes for people who build for the web. A few paragraphs each, newest first.
A 5.6-billion-parameter open-weight image model shares native 4K generation and pixel-perfect editing with the full FLUX 3 line, and runs on consumer GPUs.
CISA added CVE-2026-21589 to its exploited vulnerabilities catalog on 15 October, confirming in-the-wild attacks against the Jira, Confluence and Bitbucket bug Atlassian patched last week.
A new /v1/audio/voices endpoint turns a prompt like "calm, warm, mid-range pitch" into a reusable voice ID for GPT-Live. Creating the voice is free.
IFM and Inception42 release an open Arabic-native model family from 900 million to 375 billion parameters, plus a dedicated speech model, with weights and checkpoints on Hugging Face.
A new RDMA driver over Thunderbolt 5 lets MLX split a model across several Mac Studios, running Meta's 816B-parameter Muse Spark Open at over 36 tokens per second on four M5 Ultras.
A new validator type lets a Cloudflare Worker inspect a DLP match and adjust its confidence score, or veto it outright, using whatever logic the organisation wants.
A new trust-band system watches whether bots that already passed Cloudflare's identity check still honour robots.txt and back off on HTTP 429s, with scores heading to Cloudflare Radar.
Enterprise customers can now run Vercel Functions inside their own AWS account instead of Vercel's, while Vercel still handles deployment and routing.
GitHub Enterprise Cloud can now block pull requests over non-compliant dependency licences, with a dedicated review role and an API coding agents can query before adding a package.
A new sideEffectsCache.exclude setting stops pnpm reusing a cached native build across machines where the system library it links against might differ.
cf docs search returns five ranked results, with title, description and URL, straight from the terminal. Works for you and for coding agents.
A drive is storage you mount into a Vercel Sandbox that survives after the sandbox stops, so an agent or a dependency cache can pick up where it left off.
The default jumps automatically for every account. A new Tunnels for Platforms option also lets you run more than 1,000 tunnels per account, at $4.50 a tunnel a month.
You no longer need a waitlist or the fast-mode-2026-02-01 header to use fast mode on Opus 5.5. Opus 5 and Opus 4.8 stay in research preview.
The release lets you profile where a build spends its time, use different image optimisers for build and runtime, and raises the minimum Node.js version to 22.22.0.
Chrome 155.0.8059.84/.85 fixes ten security bugs, including a critical use-after-free in Prerender. No sign of in-the-wild exploitation yet.
CVE-2026-93953 lets a subscriber plant SQL in their own bio that fires when an admin restores a backup. Fixed in Backuply 1.5.9.
OpenAI wires its Decisions API into its realtime voice API, so a voice agent can act on what a caller is saying before they finish speaking.
Four open-weight models forecast in probabilities rather than single numbers, and the largest, at 209 million parameters, still runs on a CPU.
A new public beta lets agents use services such as Slack on a signed-in employee's behalf, routed through the company identity provider instead of a separate consent screen per service.
CVE-2026-102764 let attackers skip authentication and reach root on Remote Support and Privileged Remote Access appliances for weeks before the 15 October patch.
Gateway and DLP policies can now call out to a Worker mid-request, turning fixed rule configs into code you write yourself.
Pro teams can subscribe for monthly AI Gateway credit covering GLM 5.3 and Kimi K3, with zero data retention by default and no overage billing once the credit runs out.
Daily built PhoneLLM Mini for low-latency phone and support calls; Cloudflare now serves it on Workers AI with tool calling and a 262,140-token context window.
CVE-2026-96951 lets a crafted file skip the Mark of the Web warning entirely, and Microsoft confirmed exploitation before the fix even shipped.
Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra lose access across Copilot Chat, inline edits and agent mode, with one narrow carve-out for annual individual subscribers.
The feature backfills then continuously syncs charges, customers, subscriptions and invoices into a dedicated schema, authenticated with short-lived OAuth tokens rather than a stored database password.
AI Scan now covers vulnerabilities in PHP, Shell, Terraform, Dockerfiles, JSP and Blazor, languages CodeQL does not scan, with findings shown inline on every pull request.
The open-weight release of Meta's Muse Spark model ships with 42B active parameters, a 1M-token context window and a free licence for most commercial use.
Python support goes GA, Docker containers reach public beta, and a new Daemons feature runs background workers on Vercel without an HTTP server.