AI4AI Tech Talk: How ModelBest ForgeTrain Caught Up in 8 Hours and Surpassed NVIDIA Megatron-LM in 2 Days
On the evening of June 27, 2026, ModelBest, together with the OpenBMB open-source community and AGI BAR, hosted the "AI4AI Fermentation Night" offline event.
Li Yuxuan, head of AI Infra at ModelBest, gave an in-depth talk titled "AI Building AI: Paradigm, Measurement, Key — A Single-Point Proof Toward Level 4 with ForgeTrain." He systematically laid out the underlying logic, theoretical framework, and frontier practice of AI's self-iterating R&D, and detailed the technical thinking behind ForgeTrain, the world's first production-grade large model pre-training framework written entirely by AI with no human intervention.
Afterwards, dozens of developers, investors, industry practitioners, and researchers from academia and industry held a lively discussion on frontier topics such as the debate between customized and general-purpose frameworks, NVIDIA's CUDA ecosystem moat, and the ceiling of recursive bootstrapping.
Below are highlights from the talk and the Q&A:
"AI building AI": breaking through the boundaries of compute and intelligence
Li noted that since the beginning of this year, multiple signals have shown that AI capabilities are undergoing a qualitative change:
Computer science giant Donald Knuth has re-examined the potential of generative AI; OpenAI's Harness Engineering wrote a million lines of code with zero human involvement; DeepMind let AI autonomously research scientific questions. From individual events to the accelerating iteration of the entire industry, a consensus is forming: further gains in model capability can no longer rely solely on piling up data and compute.
Li Yuxuan, head of AI Infra at ModelBest, giving a talk at AGI BAR Beijing
From an economic perspective, large language models have come this far driven by Transformers and Scaling Laws, but high-quality internet data and compute supply are both approaching their limits, and marginal returns are falling sharply. Once models gain preliminary self-iteration capability, "AI building AI" becomes the only efficient path to break this bottleneck. Especially given China's constraints on high-end chips, multiplying compute efficiency through algorithmic and engineering means is no longer just a technical pursuit but an industrial strategy.
A broader view compares this process to the Industrial Revolution: it moved from "machines amplifying muscle" to "machines building machines," and the intelligence revolution is now moving from "AI replacing brainpower" to "AI building AI." The two are deeply isomorphic — except this time, what is being replaced and built is intelligence itself.
A ruler for "AI building AI": the L1 to L5 capability ladder
The industry's current problem is that, although the direction is clear, there is no systematic theoretical framework to measure and guide it. Drawing on OpenAI's levels of general intelligence, Li proposed five levels of "AI building AI":
L1, prompt suggestion: AI can offer suggestions on AI research, while humans perform all operations;
L2, assisted R&D: AI can assist with specific steps in the R&D chain, such as writing code, debugging, and data cleaning;
L3, closed-loop delivery: AI delivers a model end to end within an established technical paradigm, autonomously completing architecture adjustments, loss function modifications, hyperparameter optimization, evaluation, and data cleaning;
L4, recursive improvement: AI begins to have research judgment, can propose new architectures, optimizers, or training paradigms, and "forges on the spot" dedicated infrastructure to achieve recursive bootstrapping;
L5, co-evolution: AI sets its own agenda and co-evolves with the environment, becoming a question-asker rather than merely a problem-solver.
Li's assessment: in general intelligence, a preliminary, small-scale L4 has already been achieved; but in the vertical capability of "AI building AI," L2 is now solid, and only a very few teams such as Anthropic can reach L3, with results still unstable. What ModelBest is aiming at is exactly how to leapfrog L3 and strike directly at L4.
ForgeTrain in practice: the power of customization that caught up in 8 hours and surpassed in 2 days
On the practical side, Li systematically introduced the technical approach and empirical results behind ForgeTrain, ModelBest's self-developed, fully AI-written, human-free production-grade pre-training framework.
He said that realizing "AI building AI" requires an iterable evaluation and training system. ModelBest's solution borrows the logic by which general models improve through massive math and code training — first building a complex engineering evaluation environment (Harness) so that AI grows its capabilities while continuously solving real engineering problems.
Is it possible to leapfrog straight to AI building AI? The key from the previous stage, "Harness," is a general solution that climbs level by level, and it has been validated in four directions: kernels, paper reproduction, training recipes, and data selection. In the engineering dimension, however, dramatic changes on both the supply and demand sides allow us to skip levels and go straight for L4 — on the demand side, new algorithms and models have a surging appetite for "dedicated optimized infrastructure"; on the supply side, the cost of AI coding has approached zero, so customization is no longer a luxury. Where the two meet, ForgeTrain was born.
ForgeTrain's core idea: let AI "forge" a dedicated training framework from scratch for a specific model and hardware platform, rather than having humans write a general-purpose framework. According to benchmark data, ForgeTrain matched the performance of the industry flagship framework Megatron-LM within 8 hours and achieved a stable lead within 1.5 to 2 days (MFU improved by roughly 8%–10%), and the capability transfers to different models (such as MiniCPM4-0.5B/8B) and different hardware platforms (H100 and Huawei Ascend NPU).
Li further explained that this result rests on a four-stage Harness optimization process:
Anchor stage: lock down binary invariants, achieve bit-for-bit consistency with the reference implementation (max_abs_diff = 0), and build a trustworthy baseline;
Bit-for-Bit stage: while preserving binary consistency, quickly generate a code version that supports basic features such as checkpointing;
Surpass stage: lift the binary constraint, allow the Agent to autonomously enumerate kernel paths, graph capture, and scheduling combinations, and pursue longer training stability and optimization of multiple metrics;
Per-Op stage: deeply customize kernels such as GEMM and FlashAttention one by one, combined with scientific fusion, to maximize single-kernel MFU.
The whole process advances monotonically; correctness properties locked in earlier stages are not rolled back later, and every decision is made automatically without human intervention. This approach of "using AI to break through Infra" is not just a technical breakthrough but represents a new engineering philosophy — Forge Engineering.
Forge Engineering: a new engineering paradigm for the AI era
Li summarized the engineering thinking behind ForgeTrain as "Forge Engineering."
He argued that the traditional approach tends to maintain a general-purpose framework to fit many kinds of hardware, but at a time when AI models and chips iterate far faster than foundational software, this model is becoming increasingly heavy and inefficient. Forge Engineering advocates "cheap production, efficient customization":
generate a tailor-made training software stack on the fly based on the specific characteristics of the model and hardware, turning the moat NVIDIA built over the years through manpower and ecosystem into a technical problem that AI can automatically decouple and reconstruct.
He believes this idea applies not only to training frameworks but will sweep across the entire AI Infra industry and most of software engineering. In the future, everyone will be able to train their own models; models will become deeply personalized intelligent assistants, and software will emerge at scale in the form of "model customization." This is the beginning of dismantling the old ecosystem — and of giving individuals and small teams superpowers.
Open questions: frontier challenges in algorithms, engineering, data, and alignment
At the end of the talk, Li also discussed several open technical challenges.
In algorithms, how to get AI out of local improvements and have it autonomously propose paradigm-level innovations remains a fundamental problem that has not yet been effectively solved;
In engineering, whether super-large-scale parallel infrastructure can be built directly with access to only a small amount of hardware, or even none, will determine the efficiency ceiling of self-evolving R&D;
In data, autonomously constructing data synthesis methods that can continuously drive model intelligence, and filtering high-quality data from massive datasets that helps models acquire better abstraction capabilities, are current research priorities;
In alignment, how to achieve co-evolution between AI and the environment and ensure that AI development always serves human interests remains a long-term proposition for the entire industry.
Q&A highlights: Q1: Will coding agents wash away CUDA's ecosystem moat?
Li: CUDA's moat consists of three things: code accumulation, ecosystem inertia, and numerical trust. What agents mainly dismantle is "migration cost" — porting a framework to a new chip used to take months; now it can be compressed to days.
But "ecosystem trust" will not disappear on its own; users need to verify that agent-forged code is actually correct. That is why we treat Harness and the corresponding models as long-term assets, establishing new trust anchors. The moat will not be filled in overnight, but it is shifting from "trust in code" to "trust in intelligence" — the dimension of competition has already changed.
Q2: What are the core paths and methods for AI to achieve self-evolution?
Li: For now, the most important path is still ample compute, and the efficient conversion of compute into high-quality data. At the method level, reinforcement learning is the main engine: choose the right scenarios and the right RL algorithms so that models keep improving through self-play or environmental feedback.
Q3: How can customized solutions convince domestic hardware vendors to accept them?
Li: Domestic vendors currently tend to maintain large, all-encompassing general AI frameworks. Change takes time, and it also takes further leaps in model capability to prove the enormous appeal of customization gains. As model capabilities keep growing, the demand for compute efficiency will push vendors to embrace new approaches. ModelBest has already seen many collaboration opportunities and will gradually drive industry adoption.
Q4: How do you choose a good research direction?
Li: The future imagination space of the large model industry ultimately comes from realizing AGI. If you want your research direction to stay ahead, choose the steepest segment on the path to AGI.
Looking back at the truly phenomenal breakthroughs, every one of them was the next hop along the AGI mainline: around 2025, closed-source OpenAI O1 and open-source DeepSeek R1 achieved a Level 2 Reasoner with MoE architecture + RL; around 2026, closed-source Claude Opus and open-source GLM 5.2 achieved a Level 3 Agent by scaling Code Agent data.
AI's next hop is now a consensus: how to achieve a Level 4 Innovator. Research should focus on directly realizing, or helping AI complete, the capability leap toward Innovator. As for which work is relevant, you might refer to the directions Anthropic focuses on — Anthropic is the company doing this best today, and it has had a clear technical direction since its founding. Beyond that, using your own judgment to choose the direction you believe in is also a very good option.
The wave of AI building AI has just begun. ModelBest will continue to work at the frontier of intelligence evolution, letting models write the code for the next generation of models themselves.