The issue for teams running H100 clusters in production without access to Chinese models is not a shortage of open LLMs: it's a shortage of open LLMs in the 120B+ class with vision capabilities. The discussion starts with a practitioner who is blunt: for personal projects and school, they use GLM, Qwen, and DeepSeek, but their organization forbids any Chinese lab on local clusters. If it were up to them, they would deploy GLM 5.3 Flash immediately. They can't, and that changes the entire calculation.

The gap is not perception. Among Western candidates, Thinking Machines Inkling Small is described as the front runner, but according to the comparison shared in the discussion it still trails GLM 5.3 Flash by 16 points on the AA benchmark. Cohere Command A+ checks most boxes but its context window stops at 128k tokens. Then there are Poolside Laguna S 2.1 and Nvidia Nemotron 3 Super: they fit the parameter class but lack vision capabilities. That is not a wide field; the company policy narrows the options to a few models.

This has structural implications, not just technical ones. Chinese labs have raised the bar for open source in the high-end vision segment, pushing quality and cost competition. Excluding those models by policy removes competitive pressure from the Western vendor pool and imposes an implicit cost: more hardware, more fine-tuning, or more development time may be needed to reach similar performance on multimodal workloads. On an H100 cluster, a 16-point gap does not stay on paper; it shows up in pipelines requiring additional models, manual checks, or narrower scope, especially when vision is a requirement rather than an option.

There is also a knowledge asymmetry. The same practitioner knows GLM, Qwen, and DeepSeek from personal projects and understands what is being lost; management enforces a perimeter that practitioners see as a brake. This is not merely a cultural conflict. It signals that sourcing policies have not yet learned to separate technical evaluation, data sovereignty, and the availability of local alternatives. For teams evaluating on-premise deployment, the trade-off is between control and the effective capability of the model portfolio. AI-RADAR explores these issues in its /llm-onpremise section, but the choice still depends on organization-specific constraints.

The question closing the thread — are we missing any other strong contenders in the 120B category? — is more revealing than the answers that are absent. If there were an obvious alternative, this discussion would probably not be happening.