MiroMind drops MiroThinker-H1 research agents that verify their own work

When an AI agent spends twenty minutes scouring the web for a deep research brief, the real danger isn't that it comes back empty-handed. It is that it makes a tiny slip on step three, keeps digging in the wrong direction, and delivers a polished hallucination disguised as insight.

MiroMind rolled out MiroThinker-1.7 alongside its flagship variant, MiroThinker-H1. Built specifically for heavy-duty research, these open-source agents do not rely on brute-force test-time compute. Instead, they bake verification directly into the reasoning loop, forcing the model to audit its own steps before locking in conclusions.

The team flaunted some eyebrow-raising demonstrations. On February 10, the model was asked to forecast the price of gold on February 25, ultimately landing within four dollars of the actual price—an error margin of just 0.08%. In early January, it correctly called the Seattle Seahawks winning the Super Bowl a month early, and accurately picked Kendrick Lamar as the biggest winner at the Grammys. While showy stunts like sports bets always warrant skepticism, they underline a serious engine for multi-step source cross-referencing and probabilistic synthesis.

Real-time audits replace trial-and-error

Standard research agents tend to operate in long, brittle chains. MiroThinker-H1 attacks that flaw with a two-tier verification system.

The first layer is a Local Verifier. Running at inference time, it monitors planning decisions, tool calls, and evolving hypotheses turn by turn. If an agent starts chasing a dead-end lead, the verifier course-corrects immediately. On the hard subset of the BrowseComp benchmark, this local auditor slashed average interaction steps by roughly 82% while driving accuracy up by 26.4 points, putting an end to the aimless trial-and-error loop typical of older systems.

The second layer is the Global Verifier. Once a full trajectory completes, it leans on generation-verification asymmetry—the reality that checking a proof is computationally cleaner than writing one. It audits the entire evidence trail, ensuring no final claim is served without coherent backing. If the proof falls short, the system forces the agent to resample or deepen its reasoning.

This setup rests on a four-stage training pipeline. MiroMind paired mid-training for core atomic skills with supervised fine-tuning on structured reasoning triplets. Direct Preference Optimization (DPO) followed, using answer correctness as the sole ranking metric to prevent stylistic bias. Finally, reinforcement learning via GRPO introduced targeted entropy control on low-probability tokens alongside priority scheduling, ensuring stubborn long-tail problems were not swept under the rug.

235B weights and benchmark checks

Under the hood, MiroThinker-1.7 is an absolute heavyweight. It spans 235 billion parameters with a 256K context window, weighing in at roughly 470GB of raw weight files and demanding serious multi-GPU clusters for local serving. For teams running lighter infrastructure, MiroMind also dropped a 61GB mini variant.

On performance charts, H1 holds its own against top proprietary engines. It scored 88.2 on BrowseComp, edging out Gemini-3.1-Pro (85.9), Claude-4.6-Opus (84.0), and GPT-5.4 (82.7). On BrowseComp-ZH, it led the pack at 84.4, while notching 79.0 on FrontierScience-Olympiad over GPT-5.2 (77.1). We should point out that benchmark comparisons across vendor releases often feature mismatched baselines, so independent replication remains key.

Why it matters

We think MiroMind's most compelling contribution is shifting the agent narrative from brute-force thinking steps to effective interaction scaling.

For enterprise analysts, quants, and researchers, agents that lack self-auditing are essentially unusable for high-stakes decisions. By pairing real-time intermediate pruning with end-of-turn evidence verification, MiroThinker-H1 turns automated reports from fragile first drafts into auditable deliverables. You can trace back why it arrived at a probability distribution.

For the broader ecosystem, it proves that open-source models can challenge the best closed labs in multi-step agentic workflows, not just raw conversational benchmarks. The trade-off is clear: hosting 470GB locally is out of reach for solo hackers, making the 61GB mini model the true testbed for community adoption. If self-auditing architectures keep delivering this level of step-reduction, expect closed-model labs to adopt similar verification-centric loops very quickly.