#llm

Google has released Gemini 4 Hydrogen to replace Gemini 3.8 Flash. Its overall score is slightly above Argon, it runs about 2.5 times faster, and it has a free tier in AI Studio and the Gemini API, with paid use at $1.50 per million input tokens. We cover the benchmarks, pricing, where to use it, how to migrate from older models, and a prompt you can give your AI agent to switch models for you.

Meta's smallest and cheapest Muse model, Muse Ember, is built for repetitive work: tool calling, small coding tasks, routing, and extraction. It costs $0.10 per million input tokens and offers a 1M-token context window. Covers pricing and specs, how it compares with Claude Haiku 5.5 and GPT-6 Luna, and where it fits in an AI agent's workflow.

Gemini 4 Carbon is now open to everyone through the paid API and Google AI Ultra. It's built for long-running agent work and cybersecurity. Here's how it splits work with Argon, when the price doubles, and what Antigravity's /swarm does.

ERNIE 5.5 Flash Preview is live on Baidu's Qianfan platform and OpenRouter, with open weights promised by November 25. What it's good at, where it lags on benchmarks, and how to try it for free.

Beginner tutorial: first have your agent sort each step in a run log into four categories. Pick one judgment that repeats dozens of times a day and hand it to TypeSafe's Jev. Write the question by the question-writing rules, route by confidence, send uncertain cases to a large model, and do a one-week shadow run before switching. Includes 3 prompts you can copy directly.

A beginner tutorial that explains in plain language what an AI agent is, what parts it is made of, the common types, and which tasks suit an agent and which do not need one.

Intermediate tutorial: metacognition means thinking about how you think. Have the Agent evaluate its own performance, catch mistakes, and adjust its approach, with a travel Agent example.

Advanced developer tutorial: running agents with small models on your own computer, the three benefits of privacy, cost savings and offline use, what small models are good and bad at, and how to pair local tools, retrieval and MCP.

A list of the models we currently find most useful, by use case: general, coding, complex reasoning, deep research, web search, long documents, slides, image drawing and editing, AI video, digital humans, voice and music, value for money, and open weights. Personal experience; models change fast, for reference only.

Advanced tutorial: add claude_cli, codex_cli and fallback LLM providers to xiaozhi-esp32-server, calling the locally logged-in official CLIs, with all tokens handled by the CLI itself. Tool calls use a <tool_call> text protocol and are not spoken aloud. If Claude errors, it automatically switches to Codex, and the failing provider cools down for 300 seconds. Includes full code, configuration, testing methods, and measured latency.

Feedback from learner Feiwen: before, every problem meant searching everywhere, costing hours in total; now, opening an LLM clears up doubts in three minutes, and learning is much faster. We break down why it's faster, and how to use it to be both fast and accurate: give full context, ask one thing at a time, ask about the principle behind it, and verify the key points.