Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchAI, startups & growth insights. No spam.
TL;DR: MiroThinker 72B, a newly released open-source model, has posted an 81.9% score on the GAIA benchmark — a result that places it squarely in GPT-5 tier territory. With a 256K context window and the ability to execute up to 600 tool calls in a single task, it signals a turning point: open-source AI is no longer just "good enough." It is competitive with the best proprietary systems in the world.
There is a moment in every technology cycle when the open-source community catches up. It happened with databases, with operating systems, with machine learning frameworks. We are now watching it happen with frontier AI.
MiroThinker 72B is a new open-source large language model that has posted an 81.9% score on the GAIA benchmark — a result that puts it alongside GPT-5 in raw benchmark performance. The model has 72 billion parameters, a 256K token context window, and was engineered specifically for agentic tasks that require chaining hundreds of tool calls together to complete complex, multi-step reasoning problems.
To understand why this matters, you have to appreciate what was considered possible just twelve months ago. The conventional wisdom held that open-source models lagged proprietary frontier systems by two to three capability generations. You could run them locally, you could fine-tune them cheaply, and you could deploy them without per-token costs — but you could not use them for the hardest cognitive tasks. That was the trade-off.
MiroThinker 72B breaks that assumption. It is not a niche model that scores well on one narrow academic task. GAIA is specifically designed to test general AI assistants on real-world problems that require planning, tool use, and multi-hop reasoning. Achieving 81.9% on it is not a parlor trick. It is a demonstration of genuine frontier capability.
The implications cascade quickly. If an open-source model can match GPT-5 tier performance on one of the most demanding agentic benchmarks in existence, then the "proprietary frontier premium" — the idea that you have to pay OpenAI or Anthropic prices to get the best results — is no longer axiomatic. It becomes a choice, not a constraint.
GAIA (General AI Assistants) was introduced as a benchmark precisely because existing leaderboards had become easy to game. Models could score well on MMLU or HumanEval through rote pattern matching. GAIA is different. It consists of questions that require a model to browse the web, use calculators, interpret files, write and execute code, and chain all of these actions together in service of a single answer.
The benchmark has three difficulty levels. Level 1 questions can often be answered in one or two tool calls. Level 3 questions routinely require dozens of coordinated actions, and the model must track state across all of them without losing the thread. Average human performance on GAIA sits around 92%, which gives a useful ceiling. GPT-4 era models struggled to clear 30%. GPT-5 class systems — OpenAI's frontier models, the best available from Anthropic and Google — sit in the low-to-mid 80s.
MiroThinker 72B's 81.9% puts it in that same cluster.
That is not a small achievement. The gap between a model scoring 60% and one scoring 80% on GAIA is not linear. The additional 20 percentage points require qualitatively different capabilities: more robust planning, better error recovery when a tool call fails, the ability to backtrack and try a different approach without losing context accumulated in earlier steps, and a much stronger grounding in when to trust tool output versus when to sanity-check it against prior knowledge.
Every percentage point at the top of the GAIA leaderboard is harder to earn than the one below it. Reaching 81.9% as a 72B open-source model — a model that can run on a single server rather than a warehouse of GPUs — is a result the research community will be studying carefully.
It is also worth noting what GAIA does not measure. It does not test raw language fluency, creative writing, or broad cultural knowledge in the way that conversational benchmarks do. It tests something more specific and arguably more commercially important: can the model reliably complete the kind of complex, multi-step tasks that knowledge workers do every day? On that axis, MiroThinker 72B has now demonstrated parity with the best systems available.
Two architectural features define MiroThinker 72B's agentic capability: the 256K token context window and the ability to sustain up to 600 tool calls within a single task.
The context window matters because agentic workflows accumulate state. When a model is researching a topic, it might browse ten web pages, extract relevant passages from each, cross-reference them against a database, write and run code to analyze the data, and then synthesize a final answer. All of that accumulated context — the raw text from each tool call, the intermediate reasoning, the partial answers — has to live somewhere. A short context window forces the model to start dropping information or summarizing aggressively, both of which introduce errors.
At 256K tokens, MiroThinker can hold the equivalent of roughly 200,000 words in working memory — about two full-length novels, or a substantial corporate document repository. For practical agentic applications, this means the model can work through genuinely long, complex tasks without hitting a wall where it starts forgetting what it learned three steps ago.
The 600 tool call limit per task is even more striking. Most agentic frameworks today are optimized for tasks that require ten to fifty tool calls. The research that went into designing GAIA-competitive agents found that the hardest Level 3 questions require dozens or even over a hundred tool interactions to solve correctly. MiroThinker's 600-call ceiling is not arbitrary — it reflects a deliberate design choice to support the full complexity range of real-world agentic workflows.
Consider what 600 tool calls actually enables. A model could, in a single task, read and analyze a hundred-page PDF, run a web search to verify key claims, execute Python code to process a dataset referenced in that PDF, query a database to cross-reference the results, write a structured report summarizing everything, and then iterate on that report based on feedback — all within a single uninterrupted reasoning chain. That is not a toy demo. That is a description of what a skilled research analyst does over the course of a working day, compressed into an automated workflow.
The combination of 256K context and 600 tool calls makes MiroThinker not just a benchmark performer but a genuinely practical system for the kinds of long-horizon agentic tasks that enterprises actually care about.
The MiroThinker 72B result is the latest data point in a trend that has been accelerating for two years. The gap between open-source and proprietary frontier models is shrinking faster than most industry observers predicted.
The inflection point was the emergence of models like Meta's Llama series, Mistral's releases, and a growing ecosystem of high-quality base models that the community could build on. What followed was a massive distributed research effort — thousands of teams around the world fine-tuning, distilling, and extending these base models in ways that no single proprietary lab could replicate internally.
AI2's OLMo work on hybrid data efficiency showed that training efficiency gains could compress the gap further — getting more capability per training compute by being smarter about data curation and mixing. Sarvam AI's 30B and 105B open-source models for Indian languages demonstrated that open-source development was now happening at scale across geographies, not just in a handful of Western research labs. And Flash-MoE's work running 397B parameter models on consumer laptops showed that inference efficiency was improving rapidly enough that "frontier performance" was becoming accessible on commodity hardware.
MiroThinker 72B is the synthesis of all these trends. It is a model that benefits from better base models, better training recipes, better alignment techniques, and better inference optimization — all of which have been driven partly or entirely by open-source research.
The proprietary labs still hold advantages in a few specific areas. They have more compute for frontier-scale pretraining. They have proprietary data that the open-source community cannot access. And they often have months-long head starts on specific architectural innovations. But the compounding effect of open-source research is relentless, and MiroThinker 72B suggests the gap on agentic task performance has now effectively closed.
What remains to be seen is whether this is a one-time achievement by a particularly capable research team, or the beginning of a sustained period where open-source models routinely match or exceed proprietary frontier performance on specific capability dimensions. Based on the trajectory of the past two years, the latter seems more likely.
While the full technical paper for MiroThinker 72B details the architecture extensively, several key design decisions are worth highlighting for practitioners.
The model is built on a transformer-based architecture with modifications specifically designed to support long-context reasoning. Standard transformer attention scales quadratically with context length, which makes 256K context windows computationally expensive without careful engineering. MiroThinker employs a hybrid attention scheme that combines full attention for recent tokens with efficient approximations for distant context — a design that preserves recall of important information without the computational cost of naive full attention over 256K positions.
The training regime emphasizes tool-use from early stages rather than bolting it on as a post-training step. This is a significant architectural philosophy difference from many earlier agentic models, which were trained primarily on text and then adapted for tool use through instruction tuning and reinforcement learning from human feedback. MiroThinker was trained with tool-use data integrated throughout, which researchers believe contributes to more robust and reliable tool-call behavior.
Reinforcement learning played a substantial role in the final model. The team used outcome-based reward signals — did the model actually arrive at the correct final answer? — rather than process rewards that score intermediate steps. This approach encourages the model to be flexible about how it reaches the answer, rather than rigidly following a prescribed reasoning chain, which is crucial for GAIA-style tasks where there is often more than one valid solution path.
The 72B parameter scale is itself a deliberate choice. Larger models would likely score higher on GAIA, but they would also require more expensive hardware to run. The team optimized for the threshold where performance becomes frontier-competitive while remaining practically deployable on reasonable infrastructure. A 72B model, quantized appropriately, can run on a server with four to eight high-end GPUs — hardware that is accessible to well-funded startups and enterprise AI teams without requiring a dedicated AI supercomputer.
The economic implications of MiroThinker 72B matching GPT-5 tier performance are significant and likely underappreciated by the broader market.
Running GPT-5 tier tasks through OpenAI's API costs money at every inference call. For low-volume use cases — occasional queries, prototype applications, experimentation — this cost is negligible. But for production agentic applications that run thousands or millions of complex multi-step tasks, the cost picture changes dramatically.
Consider a company running automated research workflows that each consume 100K tokens of context and 50 tool calls. At typical frontier API pricing, that could cost several dollars per task. At scale — say, 10,000 tasks per day — you are looking at tens of thousands of dollars monthly just in inference costs. For many applications, this is not a business model that works.
With MiroThinker 72B, those same tasks can run on self-hosted infrastructure. The upfront capital cost of GPUs is significant, but the per-inference marginal cost approaches zero. For high-volume agentic applications, this changes the unit economics entirely. Workflows that were economically infeasible at GPT-5 tier performance become viable when you can run the model yourself.
This is not a hypothetical. The history of software infrastructure shows that when open-source alternatives reach parity with commercial offerings, adoption follows rapidly. It happened with Linux versus commercial Unix. It happened with PostgreSQL versus Oracle for many use cases. It is now beginning to happen with frontier AI models.
The proprietary labs are not standing still, of course. They have access to compute scales that will continue pushing capability frontiers beyond what open-source can immediately replicate. But MiroThinker 72B demonstrates that for the specific capability dimension of agentic task completion — which is arguably the most commercially important dimension for enterprise AI — open-source has reached the frontier. The economic moat around proprietary frontier AI for this category of use case is narrower than it has ever been.
For teams evaluating whether to deploy MiroThinker 72B, the practical hardware requirements are worth understanding in detail.
At full precision (BF16), a 72B model requires approximately 144GB of GPU memory — roughly eight NVIDIA A100 80GB GPUs or four H100 80GB GPUs. This is a significant hardware requirement, but it is within reach for enterprise AI teams and well-funded startups running their own infrastructure.
Quantization substantially changes the picture. At 4-bit quantization using techniques like GPTQ or AWQ, the memory footprint drops to approximately 36-40GB, making it feasible on a single 8xA10G server or even a pair of high-end consumer GPUs. The quantization does incur some performance penalty — benchmark scores typically drop by 1-3 percentage points — but a quantized MiroThinker still sits in the mid-to-high 70s on GAIA, which remains frontier-competitive.
Inference throughput at 72B scale is slower than smaller models, which matters for latency-sensitive applications. Agentic workflows, however, are often not latency-sensitive in the same way that conversational applications are. A research agent that takes 90 seconds to complete a complex 50-step task is perfectly acceptable for most enterprise use cases. The bottleneck in most agentic workflows is not model inference speed but rather the latency of external tool calls — web requests, database queries, code execution.
For teams using frameworks like LangChain, LlamaIndex, or AutoGen, MiroThinker integrates through standard APIs. The model ships with tool-calling support built in, compatible with the OpenAI function-calling format, which means existing agentic codebases require minimal modification to switch from a proprietary backend to MiroThinker.
Cloud deployment is also an option for teams not wanting to manage GPU infrastructure. Several cloud providers now offer model hosting services where you can run your own weights on managed GPU infrastructure, paying for compute time rather than per-token API costs. For medium-volume applications, this can be a cost-effective middle ground between fully managed API services and fully self-hosted infrastructure.
The release of MiroThinker 72B reshapes the strategic calculus for anyone building AI-powered products.
For startups, the most immediate implication is freedom. Building on proprietary APIs creates dependencies — on pricing, on terms of service, on the API provider's continued existence and interest in your use case. Open-source frontier models eliminate that dependency. A startup that bases its core product on MiroThinker owns its model stack in a way that API-dependent companies do not.
This matters increasingly as AI becomes a core competitive differentiator. If your product's core intelligence runs on the same model as your competitor's product, differentiation has to come entirely from your data, your workflow design, and your user experience. With an open-source model, you can fine-tune on proprietary data, modify the inference stack, adjust the system prompt in ways that are not possible through a public API, and build moats that go deeper than "we use GPT-5."
For developers, MiroThinker opens up a new class of applications that were previously economically infeasible. Long-running research agents, automated due diligence workflows, complex data analysis pipelines — all of these become viable at meaningful scale when the per-inference cost approaches zero. The design space for AI products expands significantly when you remove the per-token cost constraint.
There are also implications for the consulting and professional services world. Law firms, accounting firms, research organizations, and others handling sensitive client information have been cautious about sending data to proprietary API providers. Running MiroThinker on-premise eliminates the data residency concern entirely. Frontier-level agentic AI can now run inside a company's own network, subject to their own data governance policies.
The counterbalancing consideration is operational complexity. Running and maintaining a 72B model requires ML engineering capacity that many early-stage startups do not have. The managed API services offer a trade-off — higher cost per inference, but zero operational overhead — that remains attractive for many teams. MiroThinker is not a replacement for API-based development for all use cases. It is a new option that changes the calculus for a specific set of applications: high-volume, sensitive data, or deep customization requirements.
Is MiroThinker 72B truly equivalent to GPT-5 on all tasks?
No. GAIA benchmark performance is one important dimension, but it does not capture the full capability profile of a model. MiroThinker 72B is frontier-competitive on agentic task completion — complex, multi-step reasoning with tool use. Proprietary frontier models may still hold advantages in areas like broad factual knowledge, nuanced language generation, multimodal understanding, or safety and alignment properties that are harder to measure in benchmarks. The 81.9% GAIA score means frontier-tier on this specific, practically important capability axis, not a general claim of equivalence across all dimensions.
What hardware do I realistically need to run MiroThinker 72B in production?
For full-precision inference, you need approximately 140-150GB of GPU VRAM — achievable with 2x H100 80GB, 4x A100 80GB, or similar configurations. With 4-bit quantization, you can run it on 40-50GB, which fits on a single A100 or H100 server. For production workloads with reasonable throughput, plan for at minimum a 4xA100 or 2xH100 setup. Cloud GPU rental is a practical option for teams not ready to invest in owned hardware.
How does MiroThinker compare to other open-source frontier models like Llama or Mistral?
The GAIA benchmark is the key differentiator. While Llama and Mistral models have strong performance on traditional benchmarks, MiroThinker was specifically engineered for long-horizon agentic workflows with its 256K context window and 600 tool call support. For conversational applications or RAG pipelines with short context, recent Llama or Mistral models may be equally capable. For complex agentic tasks — the GAIA category — MiroThinker currently leads the open-source field.
Can I fine-tune MiroThinker on my own data?
Yes, and this is one of the key advantages of the open-source release. The model weights are publicly available, and fine-tuning with techniques like LoRA or QLoRA is practical on relatively modest GPU hardware. Teams with domain-specific data — legal documents, medical records, financial reports — can adapt MiroThinker to their specific use case in ways that are not possible with proprietary API-based models. The base model's strong agentic capabilities provide a high-quality starting point for domain adaptation.
Will proprietary labs respond by accelerating their own model releases?
Almost certainly. The release of a competitive open-source model at GPT-5 tier performance creates commercial pressure on proprietary providers. One likely response is accelerated capability releases to stay ahead of open-source parity. Another is increased focus on dimensions where open-source struggles to compete — such as multimodal capabilities, real-time data access through integrated search, or safety properties that require proprietary data and evaluation infrastructure. The competitive dynamic between open-source and proprietary AI development is healthy for the field and has historically accelerated progress on both sides.
MiroThinker 72B is not a model that merely approximates frontier performance. It achieves it, on a benchmark specifically designed to resist approximation. The 81.9% GAIA score places it in territory that, a year ago, was considered the exclusive domain of multi-billion dollar proprietary labs with access to supercomputer-scale training infrastructure.
What comes next matters as much as the achievement itself. Will the open-source ecosystem build on MiroThinker's architecture to push GAIA scores above 85%? Will fine-tuned variants on domain-specific data outperform the base model for specialized enterprise applications? Will the inference efficiency improvements seen in projects like Flash-MoE eventually make 72B parameter frontier models practical on consumer hardware?
The answers to all three questions are probably yes, and probably faster than the industry expects. The gap between open-source and proprietary frontier AI closed faster than anyone predicted to get to this point. There is no obvious reason it stops closing now.
An open-source project called Flash-MoE demonstrates that a 397-billion-parameter language model can run inference on consumer laptop hardware — a breakthrough that could democratize access to frontier-scale AI.
Mistral AI unveils Forge, an enterprise platform for building custom AI models, alongside Mistral Small 4 under Apache 2.0 — as CEO Arthur Mensch targets $1 billion in annual recurring revenue.
NVIDIA releases Nemotron 3 Super, an open-source model combining Mamba and Transformer architectures with mixture-of-experts routing — purpose-built for multi-agent AI systems with a 1-million-token context window.