AI Daily · 2026-08-03

Qwen launched Qwen3.8-Max, a 2.4-trillion-parameter model that excels at coding and autonomous long-horizon tasks, and plans to open-source the weights next week; MiniMax open-sourced MiniMax-H3, a multimodal video generation system supporting text, image, video, and audio inputs with stereo audio. LangChain rolled out an agent-first data stack, a public beta for its LLM gateway, and cost-control tools for coding agents, while CEO Harrison Chase argued that enterprises should 'own their intelligence' to build lasting advantage. Separately, Simon Willison released condense-json 1.0 for compressing JSON strings.

East Asia · First-hand

Qwen

⭐⭐⭐⭐⭐ [Model Release] Qwen3.8-Max: A New Bar for Coding and Cowork

Qwen officially releases Qwen3.8-Max, the most capable model in the Qwen family, with 2.4 trillion parameters (95B active) and open weights coming next week. Built on Qwen3.5 architecture, it delivers comprehensive improvements in coding, work, research, and long-horizon tasks, enabling reliable end-to-end complex task completion. In demonstrations, the model autonomously coded for over 10 days to build the self-evolving oh-my-cli harness from scratch, accumulating 265 commits, 127 PRs, and 151 issues, and also reproduced and improved a research paper on AI data selection. The model is available via QwenCloud API.

MiniMax

⭐⭐⭐⭐ [Model Release] MiniMaxAI/MiniMax-H3

MiniMax releases MiniMax-H3, an omni-modal generative system that takes text, image, video, and audio as input and outputs video with native stereo audio at up to 2K resolution and 15 seconds duration. The 33B model comes in two variants: a first/last-frame mode and an omni-reference mode supporting mixed inputs of up to 9 images, 3 video clips, and 3 audio clips. Its pipeline consists of a context-IR preprocessor, a 768p base generator, and a 2K super-resolution module. The model is open-sourced on HuggingFace, with API and apps available on MiniMax and Hailuo platforms.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐⭐ [Product Update] How LangChain Built an Agent-First Data Stack

LangChain rebuilt its data stack from a traditional BI-centric setup to an agent-first architecture, equipping the data agent with clear models, metric definitions, business context, and trust signals. The shift enabled the agent to handle roughly 40 times the request volume a three-person data team could manage directly, with nearly 100% of provisioned users engaging over 2,200 conversations in the last month. The data team now focuses on improving models, context, guardrails, and feedback loops rather than answering every one-off request. The experience underscores that reliable data agents need business-specific context beyond raw table access.

⭐⭐ [Product Update]

LangSmith LLM Gateway enters public beta as a centralized governance layer between agents and model calls. It enforces spend caps, rate limits, user- and customer-level policies, and model fallbacks across providers, helping prevent cost overruns and outages. The gateway also redacts sensitive data like PII and secrets before they reach model providers. These runtime controls eliminate the need to embed governance logic in each agent, offering flexibility and avoiding vendor lock-in.

⭐⭐ [Product Update] Your coding agent bill doubled. Here’s how to fix it.

Coding agent bills have spiked for many companies due to lack of oversight and fragmented cost visibility across tools. LangChain addresses this by unifying session tracing from Claude Code, Cursor, Copilot Chat and others into a single LangSmith model for standardized cost comparison. Its Engine provides automated optimization recommendations, while its LLM Gateway offers tiered cost caps and will soon route eligible tasks to cheaper open-source models.

⭐⭐ [Product Update] condense-json 1.0

Simon Willison released condense-json 1.0, a Python library that condenses JSON by replacing duplicated substrings with a special reference syntax. The tool uses a replacements object to shrink JSON size, supports reversing the condensation, and is designed to save space in SQLite logs generated by LLMs.

Opinion

⭐⭐ [Opinion] Own Your Intelligence: The Key to Lasting AI Advantage

Harrison Chase argues that generic AI cannot create lasting advantage; companies must own their intelligence by controlling models, agent systems, context, and memory, as well as managing cost, quality, risk, and behavior. Using an insurance claims example, he shows that off-the-shelf models cannot grasp company-specific rules, and real defensibility comes from systems that improve with feedback. He advises buying generic infrastructure but owning the part that compounds intelligence for the business.