← Back to blog
AiAbout 5 min read

A Chinese GPU Maker Published Its Training Numbers and Left Out the Baseline

Published Nov 15, 2026
A Chinese GPU Maker Published Its Training Numbers and Left Out the Baseline

On 6 November, at a telecom industry conference in China, Moore Threads published a set of training results for its MTTS5000 platform. The headline claims are specific enough to be checkable in principle: a 236-billion-parameter mixture-of-experts model trained from scratch, a corpus of more than 25 trillion tokens, MMLU scores level with mainstream international compute cards, and effective training time above 90%.

The platform runs on PH100 chips, which the company says were in the first batch to pass China's national information security evaluation, together with the MUSA software stack.

There is a lot to unpack, and one large gap. Every number is self-reported, and the comparison target is described only as "international mainstream training GPUs", with no model named. That phrase could mean anything from a current flagship to a three-year-old part. Until the baseline has a name and a configuration, the multiples are marketing.

That said, some of the claims are more informative than a throughput ratio, and one in particular deserves attention.

The number that actually matters

Effective training time above 90% is the least glamorous figure in the announcement and the most revealing. It measures the fraction of wall-clock time the cluster spent doing useful work rather than waiting: on a failed node, a checkpoint stall, a network hiccup, a job restart.

A row of identical steel bars on a dark surface, one polished to a mirror finish and lifted by a thin brass shim.

Peak FLOPS is a datasheet number. Effective training time is an operations number. Large clusters at scale are notorious for drifting well below their theoretical utilisation, and the reasons are rarely the chip. They are scheduler behaviour, interconnect stability, failure recovery, and how much of the stack was written by the vendor versus borrowed.

A domestic stack that sustains 90% on a from-scratch 236B MoE run is claiming something harder than a faster matrix multiply. It is claiming the boring parts work. Whether it is true is a different question, but it is the right thing to claim.

The failure modes that push a cluster below 90% are well documented and none of them are glamorous: a straggler GPU that slows the whole synchronous step, a checkpoint write that blocks training for minutes, a network fabric that degrades under all-reduce traffic, a scheduler that leaves cards idle between jobs. Fixing them is software work measured in engineer-years, and it is the part of a platform that a datasheet cannot show.

Why from-scratch matters

Most domestic accelerator announcements describe fine-tuning or inference, because those are easier and the results are quicker. Training a 236B MoE from random initialisation on 25 trillion tokens is a different class of claim: months of continuous operation, checkpoint management at scale, and a loss curve that either converges or does not.

MMLU parity is a weak proxy for that. It is a saturated benchmark, and a model can land near the same score through very different training paths. What it does establish is that the run did not collapse. A broken distributed training job does not produce a model that scores in the normal range.

The other two items in the announcement are more interesting than MMLU. The company says the same stack trained Peking University's EvoPhys-World5D world model natively, and that the model has held first place on Stanford's WorldScore in the world-generation dimension for 37 consecutive days. It also says it supported the training of RoboBrain2.5, an embodied model from the Beijing Academy of Artificial Intelligence, end to end.

Those are third-party workloads with named owners, which makes them more falsifiable than a house benchmark. If the models exist and the leaderboard position is real, the cluster did something.

The autonomous driving claim

The one multiple with a concrete workload attached is in BEV perception for autonomous driving, where the company reports single-machine and dual-machine training throughput at 1.56 to 1.67 times an international mainstream training GPU.

Perception models are a fair test. They are mid-sized, data-hungry, and sensitive to memory bandwidth and interconnect, which is exactly where a newer architecture can win or lose. A 1.6x multiple on that workload is plausible for a well-tuned stack against an older part, and implausible against the current top of the line without a lot of context about batch size and precision.

No context was given.

What would make this credible

Three disclosures would settle most of it.

Name the baseline part, its memory configuration, and the interconnect topology. A throughput comparison without those is not a comparison.

Publish the loss curve and the token-level throughput for the 236B run, rather than the final MMLU alone. Training efficiency shows up in the curve long before it shows up in a benchmark.

And state what fraction of the training code is MUSA-native versus ported from CUDA. Ports are fine. Ports that are described as native are a different claim, and it is the claim enterprises actually care about when they plan a migration.

None of this is unusual to ask. It is what any buyer of a training cluster asks before signing, and the answers are normally covered under NDA rather than in a keynote.

There is a reason vendors publish the numbers they publish. Enterprise buyers in China cannot easily benchmark competing accelerators themselves, because the comparison requires both platforms in one room, and the export controls make that room hard to assemble. In that vacuum, a vendor deck becomes the reference document, and a claim repeated in enough conference talks starts to function as a fact.

That gap is the difference between a procurement decision and a press release.

The strategic read

China's domestic accelerator story has moved through three phases: can it run inference, can it fine-tune, can it train. This announcement is an attempt to close the third. The interesting detail is that the claim is now framed around a full training run of a large MoE rather than a benchmark kernel.

If the numbers hold up under third-party replication, the practical consequence runs past Nvidia losing a sale: a Chinese lab can plan a frontier-scale training run without needing an export licence, which changes the planning horizon more than it changes the price.

If they do not hold up, the announcement still marks something: the point where the pitch stopped being about equivalence in inference and started being about equivalence in training. That is a harder claim to walk back from, and the next set of numbers will be read against it.

Related articles