Claude Desktop now integrates Ollama as gateway
Claude Desktop can now route inference through Ollama, letting you swap providers without code changes; KV cache bugs fixed to prevent token reprocessing.
Reduces friction for developers running local models or multi-provider setups. Cache restore point fixes eliminate wasteful reprocessing on cancelled requests, cutting latency and compute cost on long prefills.
Replaces manual gateway configuration for Ollama users. Requires Claude Desktop update (v0.33.0+) and Ollama configured as third-party provider. Worth adopting now if you're already running Ollama; otherwise standard Claude Desktop path unchanged.
- “Developers can now easily configure Claude Desktop to seamlessly work with Ollama as a third-party gateway provider”
- “a cancelled prefill keeps every restore point it crossed, so retries resume where they stopped instead of restarting from scratch”
- “on models with recurrent layers this previously forced a request matching 46k of 47k tokens to reprocess from zero”
claudeollamakv-cacheinference-optimizationlocal-models
Vercel Connect replaces long-lived tokens with scoped runtime requests
Request ephemeral, scoped credentials at runtime via OIDC identity instead of storing long-lived secrets—eliminates credential rotation and leakage risk for agents hitting external APIs.
Agents need access to Slack, GitHub, databases, and internal APIs without the operational burden of rotating shared tokens or the security liability of compromised standing credentials. This removes the audit and compliance surface for credential management across environments.
Replaces environment-stored API keys and bot tokens for Vercel-deployed workloads. Requires registering connectors once per provider, then calling `getToken()` at runtime—no additional secrets needed because deployments carry OIDC identity. Ready now: GA with 100+ preset connectors, audit logs, and RBAC. Worth adopting if you're already on Vercel and running agents; friction if you're not.
- “Vercel Connect replaces long-lived tokens with ones your code requests at runtime, scoped to the task and expiring on their own”
- “Every deployment on Vercel carries an OIDC identity, and the SDK uses it to prove who's asking”
- “Vercel Connect now ships with 100+ preset connectors for developer tools and SaaS providers”
- “Minting short-lived tokens instead of keeping provider credentials in paused sandboxes has removed a whole class of security risk for us”
credential-managementagentsoidcvercelsecurity
Wan 3.0 video model ships on AI Gateway
Unified text/image/audio-to-video model replaces separate Wan 2.7 endpoints; generates 30-second clips at up to 1080p with async webhooks.
Consolidates multi-modal video generation into one model ID and doubles max duration, reducing endpoint management overhead. Async generation with webhooks eliminates long-polling on video render tasks.
Direct replacement for Wan 2.7's `-t2v` and `-r2v` variants. Requires webhook URL for async mode (shown in docs); ready to adopt now if you're already on AI Gateway. Breaking change: Wan 2.7 users must migrate model IDs and handle longer generation windows.
- “Wan 3.0 combines text-to-video, image-to-video, first- and last-frame conditioning, and reference-based generation in one model”
- “generates clips up to 30 seconds at 30 fps in 480p, 720p, or 1080p, with synchronized audio”
- “Wan 3.0 supports asynchronous generation, so no HTTP request needs to remain open for the entire render”
video-generationai-gatewayasync-apisalibaba
Granite 4.2 adds reasoning and tool use to dense models
Three open-source reasoning models (3B/8B/30B) trained with 15T tokens, chain-of-thought, and agentic RL in sandboxed environments—8B and 30B learn tool calling and code execution.
Native tool calling in OpenAI-compatible format and agentic RL training mean you can run self-contained reasoning agents locally without custom glue code. 512K context window supports long-horizon tasks.
Replaces GPT-4o for reasoning workflows where you control the compute budget. Requires vLLM or SGLang for serving; agentic features (8B/30B only) need sandboxed tool environments. Try 8B now if you have inference infrastructure; 3B is viable for latency-critical edge cases.
- “Granite 4.2 is our first family of dense, decoder-only reasoning LLMs, released in three sizes: 3B, 8B, and 30B”
- “pre-trained from scratch on roughly 15T tokens”
- “extends the context window to 512K tokens”
- “the 8B and 30B models learn to act with tools inside real sandboxed environments”
- “All Granite 4.2 models are released under the Apache 2.0 license”
- “emits tool calls in the OpenAI function-calling format”
open-source-llmreasoningtool-useagentic-ailocal-inference
Next.js August 2026 vulnerabilities patched
Two critical RCE vulnerabilities in Next.js 15.x and 16.x require immediate upgrades for self-hosted apps; Vercel-hosted apps are already protected.
Self-hosted Next.js deployments face unauthenticated remote code execution via AVIF image processing and Windows filesystem paths. Vercel customers see zero operational overhead, but self-hosted teams must upgrade or disable AVIF handling immediately.
For Vercel deployments: nothing required, AVIF disabled server-side. For self-hosted: upgrade to next@15.5.24 (15.x) or next@16.3.3 (16.x) without workaround for Windows variant—patched versions serve AVIF as-is pending libheif fix. Action required now if self-hosted on Windows.
- “Applications hosted on Vercel are protected. No upgrades, configuration changes, or redeploys are required.”
- “can lead to unauthenticated remote code execution when Image Optimization processes a crafted AVIF input”
- “can lead to unauthenticated remote code execution on Windows-hosted Next.js servers”
- “AVIF images are not resized or optimized. They are served as-is until a fixed libheif version is available.”
next.jssecurityrceself-hostedimage-optimization