How do I integrate GLM-5.3 Flash into my AI Gateway setup?

Set the model parameter to `zai/glm-5.3-flash` in API calls. GLM-5.3 Flash is a multimodal model with 1M token context that routes through the unified AI Gateway API. For coding agents like Claude Code or Cursor, use the Gateway's agent setup CLI command for direct integration.

What are the setup requirements for Gemini 3.5 Transcribe on AI Gateway?

Install ai@latest and @ai-sdk/gateway@latest. Gemini 3.5 Transcribe requires 16kHz PCM audio input format and integrates via WebSocket for live transcription. It supports language detection across 85+ languages with custom vocabulary support and is production-ready.

How does Qwen 3.8 Flash compare to other long-context models on AI Gateway?

Qwen 3.8 Flash offers 1M-token context with a 65k output window as a multimodal model with no platform fee. It provides cost-competitive inference for coding agents and tool use. Only a model string swap in `streamText()` or agent setup CLI is required—no lock-in or vendor markup.

What transcription accuracy improvements does Gemini 3.5 Transcribe deliver?

Gemini 3.5 Transcribe achieves 4.0% WER on streaming transcription and 2.6% WER non-streaming, representing 70% latency reduction over Chirp 3. Two APIs available: Live for real-time voice agents and Interactions for post-recorded audio with speaker attribution.

Why should I benchmark vision models for spatial reasoning before production?

StateSight benchmark reveals that models like GPT-4.5 and Claude Sonnet 5 produce format-valid but spatially incoherent outputs on spatial tasks. Human baselines (80.8%) far exceed model performance (59.3% GPT-4.5, 53.3% Claude), so task-specific benchmarks are needed to catch failures before deployment.

Gateway Rush: GLM-5.3 Flash, Gemini Transcribe, Qwen 3.8

Tool of the Week

GLM-5.3 Flash launches on Vercel AI Gateway

Multimodal model with 1M token context now routable through unified AI Gateway API; set model to `zai/glm-5.3-flash` to use.

Adds another provider option for vision-capable inference without managing separate API keys or retry logic. Integrates directly with coding agents (Claude Code, Cursor, etc.) via Gateway's agent setup CLI.

Replaces direct Z.ai API calls with Gateway routing. Requires `model` string change and optional agent setup CLI command. Worth trying now if already on AI Gateway; otherwise evaluate against existing provider stack for cost/latency tradeoff.

“GLM-5.3 Flash is a multimodal model that supports text and vision input, with a 1M token context window”
“set `model` to `zai/glm-5.3-flash`”
“AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”

multimodal-llmai-gateway

Gemini 3.5 Transcribe now available on AI Gateway

WebSocket-based live transcription with language detection across 85+ languages and custom vocabulary support, integrated into AI SDK v7.

Eliminates the need to manage separate transcription endpoints and provider auth. Live streaming reduces latency for real-time applications like meeting transcription or live captioning.

Replaces direct Google Speech-to-Text calls. Requires npm install of ai@latest and @ai-sdk/gateway@latest, plus 16kHz PCM audio input format. Production-ready—browser sandbox available for immediate testing.

“covers 85+”
“google/gemini-3.5-transcribe-live transcribes audio over a WebSocket”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”
“streamTranscribe opens the socket and takes a ReadableStream of raw audio chunks”

speech-to-textwebsocketai-sdkstreamingvercel

Qwen 3.8 Flash now live on AI Gateway

1M-token context, 65k output window multimodal model integrated into Vercel's unified inference API with no platform fee.

Adds a cost-competitive, long-context alternative for coding agents and tool use without switching deployment infrastructure. Streamlines model selection across providers through single SDK call.

Replaces manual Alibaba API integration for teams already on AI Gateway. Requires only model string swap in `streamText()` call or agent setup CLI. Worth trying now if you're evaluating long-context models for agents—no lock-in, no markup.

“takes text and images as input, serves a context window of 1 million tokens, and can return up to 65k tokens in a response”
“Alibaba recommends it for coding, tool use, and multi-step agent workflows”
“set `model` to `alibaba/qwen3.8-flash` in the AI SDK”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”

model-releaselong-contextcoding-agentsinference-apimultimodal

Node.js 24.20.0 LTS ships async scope, stream iteration

AsyncLocalStorage gains using scopes for cleaner context binding, stream/iter moves to stable API, and JSPI WebAssembly support lands in production.

Using scopes eliminate manual context tracking boilerplate in async code; stream iteration standardization reduces fragmentation across async iteration patterns. For AI workloads, better async context isolation improves observability in concurrent operations.

Using scopes replace manual AsyncLocalStorage.run() wrapping—try it immediately if you manage deep async call stacks. Stream iteration replaces custom async iterator polyfills. Requires Node 24.20+; production-ready for LTS workloads, but audit permission audit mode needs explicit opt-in testing.

“add using scopes to AsyncLocalStorage”
“add node:stream/iter implementation”
“enable JSPI”
“update root certificates to NSS 3.125”
“add permission.drop”

node-ltsasync-contextstreamswebassemblypermissions

Muse Image launches on AI Gateway with unified API

Meta's Muse Image handles both generation and editing in a single model call via AI SDK, replacing separate model switches.

Eliminates context switching between generation and editing models. Unified API reduces integration friction and simplifies image manipulation workflows for developers using Vercel's AI Gateway.

Replaces dual-model workflows. Requires: Vercel AI Gateway access, AI SDK integration, `meta/muse-image-1.0` model endpoint. Ready now—use `generateImage()` with `prompt.images` for reference blending or instruction-based editing. Worth trying if already on AI Gateway stack.

“It is their first image model and a separate family from Muse Spark, returning images rather than text.”
“One model does both, so you don't switch models to move from generating to editing.”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”

image-generationapi-gatewaymeta-musevercelsdk

Gemini 3.5 Transcribe delivers sub-second streaming transcription

Two separate APIs—Live for real-time voice agents (gemini-3.5-transcribe-live), Interactions for post-recorded audio with speaker attribution—both handle custom vocabulary and 85+ languages natively.

Replaces Chirp 3 with 70% latency reduction and measurable accuracy gains (4.0% WER streaming, 2.6% non-streaming). Enables voice-first workflows in existing Gemini API integrations without architectural changes.

Public preview now in Google AI Studio and Gemini API. Requires switching from Chirp 3 model identifiers and choosing between Live (streaming) or Interactions (batch) paths. Ready to try for voice agents and transcription pipelines; production readiness depends on your WER tolerance and language mix.

“Delivers continuous, bidirectional streaming with sub-second latency for interactive voice apps via the Live API using gemini-3.5-transcribe-live”
“achieves an average Word Error Rate (WER) of 4.0% for streaming and 2.6% for non-streaming use-cases”
“time to final transcription, for example, improves by 70%”
“Automatically detects and transcribes over 85 languages”
“In public preview in the Gemini API via Google AI Studio and Google Antigravity”

speech-to-textreal-time-apigeminivoice-agentstranscription

Data Point

Vision models fail spatial reasoning despite format validity

StateSight benchmark exposes that GPT-4.5 and Claude Sonnet 5 produce syntactically correct but spatially incoherent outputs on cube-net and tower-counting tasks, masking fundamental reconstruction failures.

If you're building multimodal QA systems that depend on spatial understanding from single images, format-valid API responses hide real failures in image-state reconstruction—you need task-specific benchmarks to catch this gap before production.

Doesn't replace existing vision pipelines yet; reveals that broad evals mask spatial reasoning brittleness. Requires custom procedural benchmarks for your domain. Worth running StateSight on your target model before shipping spatial-inference features—human baselines (80.8% on cubes) far exceed current model performance (59.3% GPT-4.5, 53.3% Claude).

“Format-valid responses can mask failures to recover the spatial structure required for verifiable visual inference”
“OpenAI GPT-5.5, using the API model identifier gpt-5.5, achieved 59.3%, 33.3%, and 28.3% accuracy across the three tasks”
“A 30-participant human baseline on 60 items exceeded both models on every task, with mean accuracies of 80.8%, 68.8%, and 64.3%”

vision-language-modelsspatial-reasoningbenchmarkevaluationmultimodal