US open-weight models close the gap on code while China's compute gap stays wide
Four US model families passed 90% on a C++-to-Rust benchmark, efficiency methods cut training and serving costs, and Huawei's biggest campus still holds under a tenth of the largest US facility's capacity.
- Code Metal says four US model families (Muse Open, Beam, Nemotron 3 Ultra, Inkling) exceeded 90% judged function-body correctness on its internal C++-to-Rust benchmark, with Muse Open reaching 100% at medium and xhigh reasoning settings. Self-reported m:…6891
- Epoch AI estimates Huawei's Horinger data center holds less than one-tenth the capacity of the world's largest AI data center, Colossus 2, at about 81,000 H100-equivalents against at least 1.8 million expected for Colossus 2 by 2028. Self-reported m:…9922
- In a preprint, StepX-OPD trains SD3.5-Medium with 37% fewer GPU-hours than sequential diffusion distillation and step distillation. Self-reported m:…7411
- In a preprint, the Muir linear-attention architecture trains 2.7x faster than Transformer at 32K context length and extrapolates from 4K training length to 10M-token streams without perplexity drift. Self-reported m:…6090
Four US families pass 90% on code translation
Code Metal says US-developed open-weight models are increasingly competitive on code translation, narrowing the performance gap with Chinese-developed peers. On its internal benchmark, which translates a C++ library covering 14 algorithms into Rust, four US model families, Muse Open, Beam, Nemotron 3 Ultra and Inkling, exceeded 90% judged function-body correctness. Muse Open reached 100% judged correct at medium and xhigh reasoning settings and matched the strongest reference models on judged function-body correctness. The result matters most for teams with strict security constraints, because a growing pool of high-performing US models gives them more flexibility to deploy and evaluate AI on controlled, local infrastructure. m:…6891
The gap has not closed evenly. OpenAI's gpt-oss-120b reached only 42% correctness at its best tested setting, while DeepSeek V4.1 Flash reached 98% judged correct in 61 seconds with thinking off, against a median 99 normalized seconds per task for Muse Open. Serving speed also splits the field: the DeepSeek endpoint decoded at 255 tokens per second per session against 106 for Muse Open, which generated about 8,700 completion tokens per task. Code Metal's own note is that more reasoning time does not always yield better results and can significantly increase token usage, so the choice of model is also a choice about serving cost. m:…6891
Huawei's Horinger campus and the compute gap
Epoch AI's analysis of Huawei's Horinger data center in Inner Mongolia tells the other half of the story. The campus holds less than one-tenth the capacity of the world's largest AI data center, Colossus 2, with an estimated 81,000 H100-equivalents of compute capacity and 286 MW of IT power capacity. Horinger runs about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips, with a rack occupancy rate estimated at 76%. The constraint is efficiency as much as scale: Huawei's best chips are less energy-efficient than Nvidia's export-controlled chips, and Epoch AI reports Nvidia B200 and B300 chips are 2.5 times more energy-efficient than the Ascend 910C. m:…9922
Horinger is expected to grow its AI compute by over 70% by early 2028, and Huawei plans to deploy Ascend 950 chips in the new Horinger building, with its sister campus in Wuhu projected to reach 151,000 H100-equivalents by December 2027. But Epoch AI's analysis says this growth relies on adding power rather than closing the efficiency gap, and Colossus 2 is expected to reach at least 1.8 million H100-equivalents by 2028. This suggests the compute gap is a hardware-efficiency gap first and a building gap second, and it will not close just because more campuses come online. m:…9922
Efficiency is how the gap closes
Two preprints in the same batch show the efficiency techniques that let smaller players compete. StepX-OPD merges capability consolidation and few-step sampling into a single training stage for text-to-image diffusion models, using frozen specialists to supervise students on states visited during their own rollouts. It trains SD3.5-Medium with 37% fewer GPU-hours than sequential DiffusionOPD followed by step distillation, and it outperforms the forty-step Flow-OPD model. The point is that training a deployable model gets cheaper when the capability and compression stages stop being separate jobs. m:…7411
Muir, a linear attention architecture that allocates gate capacity by sensitivity, attacks the cost of running models instead. In a preprint, Muir lowers perplexity by 1.8 to 2.0% over Gated DeltaNet-2 and 2.5 to 3.6% over Transformer baselines, trains 2.7x faster than Transformer at 32K context length, and reaches its 100B-token validation loss 1.2x faster than Gated DeltaNet-2. It also extrapolates from a 4K training length to 10M-token streams without perplexity drift. Taken together, the benchmark result, the data-center analysis and the two efficiency papers describe one race run on two tracks: model quality, where the US open-weight field has caught up on code, and compute, where it has not. m:…6090
What is still unresolved
- Full-library runtime performance including compilation, repair and verification, which Code Metal says remains to be measured.
- Whether faster serving translates into higher full-library migration success rates.
- How Huawei's Ascend 950 deployment changes Horinger's efficiency picture.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.