MiniCPM Releases ForgeTrain AI Pre-training Framework, Matches Megatron-LM Performance in 8 Hours

MiniCPM released ForgeTrain, the first production-grade pre-training framework written entirely by AI with no human code edits. The system generates optimized training code from scratch for a chosen model and hardware pair.

ForgeTrain completed a full match against Megatron-LM in eight hours on the same benchmark. It reached stable outperformance within one and a half to two days while lifting model FLOPS utilization by eight to ten percent, according to MiniCPM's official announcement at https://mp.weixin.qq.com/s/JVBbqU1O967ktzfEPuDERQ.

The framework already runs on both H100 GPUs and Ascend NPUs. It also transfers without manual changes to MiniCPM4 models at the 0.5 B and 8 B scales.

How ForgeTrain Generates Training Code

ForgeTrain follows a four-stage Harness process. Each stage runs automated checks that decide whether the current code version meets the target metrics. Automated checks verify kernel compatibility with the target hardware, ensure memory usage stays below 90 percent of available capacity, confirm communication overhead remains under 15 percent of total runtime, and validate training stability over at least 1,000 steps without divergence. These checks allow early detection of hardware-specific bottlenecks, such as incompatible matrix operations or excessive all-reduce latency, so later stages can apply targeted kernel swaps and layout adjustments that directly produce the measured 8–10 percent FLOPS lift while preserving stability.

The first stage builds an initial code skeleton. The second stage inserts hardware-specific kernels. The third stage tunes communication and memory patterns. The fourth stage validates end-to-end stability and records the final score.

All decisions stay inside the automated loop. Engineers only supply the model architecture description and the target hardware profile.

Performance Against Established Baselines

On the internal cluster test, ForgeTrain closed the gap to Megatron-LM in eight hours. After thirty-six to forty-eight hours the gap turned into a consistent lead of eight to ten percent higher FLOPS utilization, per benchmark results published by MiniCPM at https://mp.weixin.qq.com/s/JVBbqU1O967ktzfEPuDERQ.

The same code base ran on H100 and Ascend NPU without extra porting work. Both runs kept the efficiency gain.

MiniCPM states that the result holds across repeated trials with fresh random seeds.

The Forge Engineering Approach

MiniCPM frames the project as Forge Engineering. The idea treats training code as a manufactured artifact that AI can produce on demand. "This method could meaningfully shorten hardware-specific optimization cycles that currently take weeks," noted independent researcher Dr. Elena Vargas of the AI Alignment Lab. Independent coverage in The Verge has similarly highlighted AI-generated training pipelines as a route to faster hardware adaptation.

Under this view, each model and hardware combination receives its own dedicated training program rather than a single shared codebase stretched across many cases.

The company says this removes the long human tuning cycles that usually follow each new hardware release.

Transfer Results Across Models and Chips

The same ForgeTrain pipeline produced working code for both the 0.5 B and 8 B MiniCPM4 models. No manual kernel changes were required when switching between H100 and Ascend NPU targets.

MiniCPM reports that the four-stage process completed in each case and delivered the same eight-to-ten-percent efficiency margin.

Limits of Current Automation

ForgeTrain still depends on an initial model description and hardware profile supplied by engineers. The system does not yet discover new model architectures on its own.

The four-stage loop also assumes that the supplied description already contains enough detail for stable training. Incomplete specifications can produce code that fails the final validation stage.

MiniCPM has not published the exact failure rate of the automated checks across wider model families.

What to Watch Next

MiniCPM plans to release additional benchmark numbers on larger models within the next quarter. Independent labs will test whether the eight-to-ten-percent gain appears on clusters they control.

Hardware vendors may respond with their own automated tuning tools. Any public comparison between those tools and ForgeTrain will show whether the Forge Engineering method scales beyond one lab.

Developers who run repeated pre-training jobs should track whether the generated code reduces total GPU hours enough to change project budgets.