MiroThinker 1.5 - The 30B That Outperforms 1T Models

MirrorThinker 1.5 is positioned as a practical shift in agent design: instead of relying on a single, information-heavy model, it’s built to repeatedly call tools—up to 400 tool calls—so it can research, verify, and generate multi-step outputs. The core claim is that this tool-centric “agentic” capability lets a comparatively smaller model compete with far larger systems on tasks that demand sustained tool use, not just raw language fluency. That matters because many real workflows—web research, code execution, report writing, and even podcast-style or slide-deck generation—depend on retrieving new information and transforming it through multiple steps.

The lineup centers on two open models under an MIT license: MirrorThinker 235B (22B active parameters) and a smaller 30B variant (3B active parameters). Both are derived from the Qwen mixture-of-experts base models and then improved for long-horizon tool calling. They come with a 256,000-token context window and are intended to support up to 400 tool calls, though the transcript notes that reaching 400 elegantly without loops or repetition remains a broader challenge across agent systems.

Benchmark comparisons are used to argue that MirrorThinker’s tool-use focus narrows the gap with much larger models. On some evaluations—such as humanities-oriented tasks—the larger MirrorThinker model is described as close to Gemini 3 Pro and nearer to GPT-5’s high-level model and GLM 4.7. In browser-oriented benchmarks, MirrorThinker is described as state-of-the-art. The emphasis, however, is less on “climbing” a single leaderboard and more on overall competitiveness against other tool-using families, including DeepSeek V 3.2, MiniMax models, GLM, and Qwen’s own K2 thinking model.

Under the hood, the agent setup relies on a suite of tool interfaces: code execution in sandboxes, file management, information retrieval, web search, and page fetching. Long multi-step runs require context management—deciding what to keep, what to truncate, and how to retain recency—so the model can keep track of what matters as tool results accumulate. The transcript highlights that this kind of context retention and compaction is increasingly being “baked into” agent systems, citing similar directions from Anthropic (Claude models and Claude Code) and OpenAI, with early examples also appearing in Google products.

A hands-on walkthrough shows how the 30B model can be run locally, but with real hardware costs: full-resolution use reportedly needs an NVIDIA A100 with 80GB RAM, and model weight downloads take several minutes. Instead of using Hugging Face Transformers directly, the setup runs a VLM server so the model can be accessed via an OpenAI-style API, leveraging built-in function-calling support. Tool calls are executed through a custom agent loop built from scratch (not LangChain), with explicit handling for tool invocation, tool results, and iteration limits.

In practice, the agent can solve math via calculate tools and perform web research by searching and fetching pages, but the transcript flags two recurring failure modes: hitting the maximum iteration budget before producing a final answer, and producing answers influenced by irrelevant retrieved content. A time-zone example (Singapore time) illustrates another friction point—correct results may require many tool steps, especially when the runtime isn’t in the target region. The takeaway is that MirrorThinker 1.5 looks useful for local, non-real-time workflows that tolerate multi-step latency, while smaller models still struggle with efficiency and step quality.

Finally, the transcript raises a forward-looking question: if quantized versions (MLX 8-bit or 4-bit) can preserve reasoning quality, MirrorThinker-style tool agents could become more accessible for local deployment using tools like Llama.cpp or LLM Studio. An online demo is offered for experimentation, with the expectation that it may use the larger model given the longer “thinking” behavior.