Joining the ranks of a growing number of smaller, powerful reasoning models is MiroThinker 1.5 from MiroMind, with just 30 billion parameters, compared to the hundreds of billions or trillions used by leading foundation large language models (LLMs).
But MiroThinker 1.5 stands out among these smaller reasoners for one major reason: it offers agentic research capabilities rivaling trillion-parameter competitors like Kimi K2 and DeepSeek, at a fraction of the inference cost.
The release marks a milestone in the push toward efficient, deployable AI agents. Enterprises have long been forced to choose between expensive API calls to frontier models or compromised local performance. MiroThinker 1.5 offers a third path: open-weight models architected specifically for extended tool use and multi-step reasoning.
One of the biggest trends emerging in the industry is a move away from highly specialized agents toward more generalized ones. Until recently, that capability was largely limited to proprietary models. MiroThinker 1.5 represents a serious open-weight contender in this space. Watch my YouTube video on it below.
Reduced Hallucination Risk Through Verifiable Reasoning
For IT teams evaluating AI deployment, hallucinations remain the primary barrier to using open models in production. MiroThinker 1.5 addresses this through what MiroMind calls “scientist mode”—a fundamental architectural shift in how the model handles uncertainty.
Rather than generating statistically plausible answers from memorized patterns (the root cause of most hallucinations), MiroThinker is trained to execute a verifiable research loop: propose hypotheses, query external sources for evidence, identify mismatches, revise conclusions, and verify again. During training, the model is explicitly penalized for high-confidence outputs that lack source support.
The practical implication for enterprise deployment is auditability. When MiroThinker produces an answer, it can surface both the reasoning chain and the external sources it consulted. For regulated industries such as financial services, healthcare, and legal, this creates a documentation trail that memorization-based models cannot provide. Compliance teams can review not just what the model concluded, but how it arrived there.
This approach also reduces the “confident hallucination” problem common in production AI systems. The model is trained to seek verification rather than extrapolate when uncertain—a behavior that translates directly into fewer costly errors.
Benchmark Performance: Punching Above Its Weight
Under this framework, MiroThinker-v1.5-30B delivers performance comparable to models with up to 30× more parameters, including the trillion-parameter Kimi-K2-Thinking model.
On BrowseComp-ZH, a key benchmark for web research capabilities, the 30B model actually outperformed its trillion-parameter competitor with a score of 69.8.
The cost differential is equally notable. MiroMind reports inference costs as low as $0.07 per call for the 30B variant—roughly one-twentieth the cost of Kimi-K2-Thinking—along with faster inference speeds.
A larger 235B variant (with 22B active parameters in a mixture-of-experts architecture) ranks in the global top tier across multiple search-agent benchmarks. On general agentic search evaluations, these models hold their own against systems from DeepSeek V3.2, Minimax, GLM, and Kimi-K2.
In testing, the larger model approaches Gemini 3 Pro on several benchmarks and comes closer to GPT-5-class systems than its parameter count might suggest. While benchmark hill-climbing is increasingly common, what matters more is overall competitiveness—and MiroThinker holds up well.
Extended Tool Use: Up to 400 Tool Calls per Session
The defining capability of MiroThinker 1.5 is sustained tool use.
The models support up to 256,000 tokens of context and claim support for up to 400 tool calls per session—a critical requirement for complex research workflows involving extensive information gathering, synthesis, and cross-checking.
This places MiroThinker firmly in the emerging category of agentic models designed for autonomous task completion rather than single-turn Q&A. Practical applications include deep research workflows, content pipelines, report generation, and podcast-style outputs similar to NotebookLM.
Training Innovation: Time-Sensitive Sandbox
Another major innovation in MiroThinker 1.5 is its Time-Sensitive Training Sandbox.
Traditional model training operates from what MiroMind describes as a “God’s-eye view,” where the model has access to finalized outcomes within static datasets—creating hindsight bias. MiroThinker’s training removes that advantage.
During training, the model can only interact with information published before a given timestamp, preventing future leakage and forcing it to reason under realistic conditions of incomplete information.
The pipeline combines supervised fine-tuning with reinforcement learning using verifiable rewards via Group Relative Policy Optimization (GRPO), an advanced reinforcement learning algorithm popularized by DeepSeek,, encouraging the model to select the right tool at the right time.
This approach is especially relevant for enterprise use cases where models must reason about evolving situations rather than recall static facts.
Practical Deployment Considerations
For IT teams considering deployment, hardware requirements still matter. Even the 30B model requires a substantial amount of GPU memory, and smaller setups may struggle.
One advantage is compatibility. MiroThinker runs on vLLM servers with OpenAI-compatible API endpoints, making it easier to integrate into existing toolchains and function-calling workflows as a drop-in replacement.
Both model sizes are available under the permissive, enterprise-friendly MIT license on Hugging Face, and an online demo is available for evaluation. The permissive license removes major barriers to internal deployment and fine-tuning.
The Bigger Picture: Interactive Scaling vs. Parameter Scaling
MiroThinker 1.5 arrives as the industry confronts the limits of traditional scaling laws. Bigger models no longer guarantee better real-world performance. As Artificial Analysis has noted, many benchmarks are saturated, pushing the industry toward evaluations based on economic usefulness rather than abstract reasoning alone.
MiroMind’s bet is on interactive scaling—improving capability through deeper tool interaction rather than ever-larger parameter counts. If correct, this could enable sophisticated agents on infrastructure that does not depend on expensive frontier APIs.
The company, founded by Tianqiao Chen and AI scientist Jifeng Dai, describes its mission as building “Native Intelligence”—AI that reasons through interaction, not memorization.
Whether this approach becomes dominant or remains a specialized niche is still an open question. But for enterprises wrestling with cost-capability tradeoffs, MiroThinker 1.5 offers a compelling data point: sometimes, teaching a model how to research matters more than teaching it to remember everything.
