Gemini 3.5 Transcribe and Muse Image on AI Gateway, Qwen 3.8 Flash, and agent deployment tools
TL;DR
New transcription and image models landed on Vercel AI Gateway alongside Qwen variants and infrastructure for running coding agents at scale.
What shipped
On 26 August several vendors released production-ready models and platforms aimed at builders who need reliable audio, image, and agent capabilities. Vercel expanded its AI Gateway with Google and Meta models while Alibaba added a long-context option suited to coding workflows. Separate launches in agent infrastructure and trending open models on Hugging Face show the shift from chat prototypes toward deployable systems.
Vendor launches
Vercel added three new models to its AI Gateway while NVIDIA detailed memory hardware for large agent clusters. The Gemini and Muse releases give immediate access to transcription and image editing without separate providers. Qwen 3.8 Flash brings a 1-million-token window aimed at agent and coding tasks.
- •NVIDIA NVLink Fusion with NVHBM NVIDIA released custom high-bandwidth memory hardware to support trillion-parameter agent workloads. Hyperscalers can now combine compute, memory, and networking in one system for lower latency on large models.
- •Gemini 3.5 Transcribe on AI Gateway Google released Gemini 3.5 Transcribe on Vercel AI Gateway in both batch and live WebSocket versions. It handles 85 languages and custom vocabulary for meeting notes or support calls without switching providers.
- •Muse Image on AI Gateway Meta released Muse Image on Vercel AI Gateway for prompt-to-image and image-editing in one model. Teams can generate or edit visuals inside existing Vercel projects instead of managing separate image services.
- •Vercel Security Dashboard Vercel made its Security Dashboard generally available on all plans with CLI access. It flags missing 2FA and long-lived keys as teams add more coding agents to projects.
- •Qwen 3.8 Flash on AI Gateway Alibaba released Qwen 3.8 Flash on Vercel AI Gateway with a 1-million-token context. It targets coding agents and multi-step tool use inside the AI SDK.
- •Gemini 3.5 Transcribe from DeepMind Google DeepMind released an updated Gemini 3.5 Transcribe model focused on accurate speaker tracking. Developers can integrate it for real-time captioning in apps without extra language detection code.
Hugging Face trending
Five models climbed the Hugging Face charts, led by GLM-5.3-Flash and several Qwen variants. Builders can download and fine-tune them locally or run inference directly from the hub. The mix shows continued demand for fast text and multimodal checkpoints.
- •GLM-5.3-Flash by zai-org zai-org released GLM-5.3-Flash, a text-generation model now trending on Hugging Face. It supports quick local inference or fine-tuning for chat and agent prototypes.
- •Qwen3.8-Flash-Next by Qwen Qwen released Qwen3.8-Flash-Next, an image-text-to-text model trending on the hub. It fits visual question answering or document analysis inside existing transformers pipelines.
- •MiniMax-H3-Fun-Controlnet-Union by alibaba-pai alibaba-pai released a text-to-video ControlNet model trending on Hugging Face. Teams can fine-tune it for short video clips from text prompts.
- •Qwen3.8-27B-DFlash2 by incoai incoai released Qwen3.8-27B-DFlash2, a text-generation model now trending. It offers a larger Qwen variant for local coding or agent experiments.
- •Ornith-1.5-9B-GGUF by ornith-ai ornith-ai released Ornith-1.5-9B-GGUF, a quantized text model trending on the hub. It runs on modest hardware for on-device chat or tool calling.
Product Hunt picks
Three agent-focused tools appeared on Product Hunt, covering workload infrastructure, persistent agent hiring, and page inspection. Each targets teams moving from prototype agents to production runs.
- •Warren Warren launched infrastructure built for coding-agent workloads. Teams can run multiple agents without managing separate servers or scaling scripts.
- •DeployHermes DeployHermes released a platform to hire persistent agents with defined roles and memory. SMBs can assign tasks to agents that retain context across sessions.
- •LoupeKit LoupeKit launched a tool that shows what any webpage uses and how much is AI-generated. Builders can audit competitor sites for AI components in minutes.
Industry news
Anthropic signed a $45 billion compute deal while OpenAI stated it expects AGI under its definition by end of 2026. Political candidates also formed a pact on data-center rules. These moves signal rising infrastructure costs and regulatory attention.
- •OpenAI AGI timeline Sam Altman said OpenAI expects AGI by end of 2026 under its definition. The Astra model already acts as an automated research intern on internal tasks.
- •China Robot Games Humanoid robots at Beijing's Robot Games showed advanced fine-motor control beyond running speed. The demonstrations highlight progress in precise manipulation for warehouse or assembly use cases.
- •Anthropic Nscale deal Anthropic signed a $45 billion compute agreement with Nscale. The deal continues the pattern of large model labs locking in capacity years ahead.
Other
LangChain published multiple updates on agent evaluation, latency reduction, and its new generally available platform. Additional posts covered GitHub Copilot automation and a Thai-language corpus built with Dolma. The volume shows maturing tooling around agent reliability.
- •Podium LangSmith case study Podium used LangSmith to cut engineering intervention on its AI employee agent by 90 percent while reaching 98 percent F1 quality. Other teams can apply the same dataset curation steps.
- •LangChain second birthday reflections LangChain marked two years with notes on product and community growth. The post outlines the shift from open-source library to paid platform offerings.
- •LangChain State of AI 2024 Report LangChain released usage data showing how teams build and monitor LLM apps. The report highlights rising adoption of evaluation and tracing features.
- •LangChain three-year journey Harrison Chase detailed LangChain's path to a $1.25 billion valuation and announced LangChain 1.0 plus new funding. The update includes LangSmith expansion plans.
- •LangChain chatbot rebuild LangChain rebuilt its own chatbot with Deep Agents and subgraphs for sub-15-second responses and citations. Builders can copy the subgraph pattern for faster support agents.
- •January 2026 LangChain Newsletter LangChain published its January 2026 newsletter with product and event updates. It covers recent additions to LangGraph and evaluation tools.
- •AI Agent Latency 101 LangChain posted practical steps to reduce agent latency through parallelism and fewer LLM calls. Teams can test the listed optimizations on existing agents this week.
- •LangGraph Platform GA LangChain made LangGraph Platform generally available for deploying stateful agents. It handles long-running workflows without custom orchestration code.
- •Amazon Bedrock AgentCore Evaluations AWS released AgentCore Evaluations to test any agent framework. Developers can benchmark open-source agents against Bedrock baselines.
- •Dolma Thai adaptation Thai researchers adapted Ai2's Dolma toolkit to create a 47-billion-token Thai corpus. The work shows how to filter web data while preserving cultural knowledge in smaller languages.
- •GitHub Copilot Dependabot triage GitHub posted a beginner guide to using Copilot for automating Dependabot pull request reviews. Teams can reduce manual library update work with the provided prompts.
- •Qwen3.8 Flash on OpenRouter Qwen3.8 Flash became available on OpenRouter with 1,000k context at $0.16 per million input tokens. It supports coding and long-document agent tasks at lower cost than prior options.
Replicate new models
p-video on Replicate: prunaai released p-video on Replicate with built-in draft mode for fast video generation. Vibe Builders can call it via the existing Replicate token for quick creative tests.
What this means for you
For Vibe Builders: You can now call Gemini 3.5 Transcribe and Muse Image directly inside Vercel projects for audio notes and image edits without new accounts. Qwen 3.8 Flash and the trending Hugging Face checkpoints give longer context for agent experiments. Warren and DeployHermes on Product Hunt plus LangGraph Platform GA let you run persistent agents with less custom setup this week.
For Non-techies: Transcription and image tools from Google and Meta now sit inside tools you already use, so meeting notes and product visuals take fewer steps. Agent platforms like DeployHermes and Warren mean you can assign ongoing tasks to AI helpers that remember prior work. Watch the new security dashboard to keep projects safe as you add more of these tools.
For Developers: Vercel AI Gateway now hosts Gemini 3.5 Transcribe, Muse Image, and Qwen 3.8 Flash with documented SDK calls for direct integration. LangGraph Platform GA and Amazon Bedrock AgentCore Evaluations give concrete options to deploy and benchmark agents against current stacks. NVIDIA's memory hardware and Anthropic's $45 billion deal signal continued capacity pressure; test local Qwen and GLM checkpoints before scaling cloud spend.
What to watch next
Track rollout of Gemini 3.5 Transcribe live mode and any new Qwen variants on additional gateways. Watch LangGraph Platform adoption metrics and the first public benchmarks from AgentCore Evaluations. Note regulatory signals from the AI Pact signatories on data-center rules.
Harsh’s take
The day shows clear separation between flashy model drops and the unglamorous work of making agents reliable at scale. Most new checkpoints still require careful prompt and evaluation work, while the real production wins come from tracing and deployment platforms that reduce manual fixes. The $45 billion compute commitments highlight that raw scale remains expensive even as open models improve.
Builders should pick one agent workflow, run it through LangSmith or AgentCore this week, and measure intervention rate before adding another model. Skipping that step leaves teams paying for capacity they cannot yet use effectively.
by Harsh Desai
Sources
Vendor launches
- •NVIDIA NVLink Fusion Expands With NVHBM Custom High-Bandwidth Memory
- •Gemini 3.5 Transcribe now available on AI Gateway
- •Muse Image now available on AI Gateway
- •Vercel Security Dashboard is now generally available
- •Qwen 3.8 Flash now available on AI Gateway
- •Intelligent transcription with Gemini 3.5 Transcribe
- •28 startups using AI to transform the energy sector
- •Google at the Global Forum on Intellectual Property
Hugging Face trending
- •GLM-5.3-Flash by zai-org trends on HuggingFace
- •Qwen3.8-Flash-Next by Qwen trends on HuggingFace
- •MiniMax-H3-Fun-Controlnet-Union by alibaba-pai trends on HuggingFace
- •Qwen3.8-27B-DFlash2 by incoai trends on HuggingFace
- •Ornith-1.5-9B-GGUF by ornith-ai trends on HuggingFace
Product Hunt picks
Industry news
- •Candidates Are Signing a Pact Promising Action on Data Centers and AI Safety
- •Sam Altman says OpenAI will have AGI by the end of 2026 if you accept his definition
- •The Humanoids at China’s Robot Games Were Faster Than Usain Bolt, but I’m More Impressed by Their Tweezer Mastery
- •Anthropic continues compute-gobbling streak in $45 billion deal with Nscale
Other
- •How Podium optimized agent behavior and reduced engineering intervention by 90% with LangSmith
- •LangChain's Second Birthday
- •LangChain State of AI 2024 Report
- •Reflections on Three Years of Building LangChain
- •Why We Rebuilt LangChain’s Chatbot and What We Learned
- •January 2026: LangChain Newsletter
- •AI Agent Latency 101: How do I speed up my AI agent?
- •LangGraph Platform is now Generally Available: Deploy & manage long-running, stateful Agents
- •Evaluate any agent framework with Amazon Bedrock AgentCore Evaluations
- •How researchers adapted Dolma for better Thai language models
- •GitHub Copilot app for Beginners: Automate Dependabot pull request triage
- •Qwen: Qwen3.8 Flash now available on OpenRouter (1,000k context, $0.16/M in, $0.47/M out)
Replicate new models
More AI news
- Daily RoundupVercel Connect GA and Run SDK, Wan 3.0 on AI Gateway, and agent tools for production
Vercel expanded agent connectors and security tools while new video models and debugging platforms rolled out, giving builders more ways to ship reliable agents without managing credentials or long timeouts.
- Daily RoundupAudio8 TTS and Wan 3 video model launch, plus agent efficiency benchmarks (sandbox tools to test today)
Hugging Face and Replicate added new text-to-speech, video, and image models while NVIDIA and Vercel released efficiency and sandbox updates that affect how agents run in production.
- Weekly DigestHermes Agent v0.20 voice and A2A launch, OpenClaw GPT-5.6 support, and agent stability fixes (two flagships + practical hook)
Hermes Agent rolled out v0.20 series with conversational voice, A2A protocol, and desktop UI updates while OpenClaw added model support, plugin safeguards, and mobile caching across the week.