The Scientist

The Convergence That Changes Your Architecture

Meituan put out LongCat-2.0 this week, and the self-hosting calculus moved from interesting hedge to economically obvious default. It is a 1.6-trillion-parameter sparse MoE, 48B active per token, under MIT license, and it edges GPT-5.5 on SWE-bench Pro at 59.5 to 58.6. Z.ai shipped GLM-5.2 the same week: a 744B MoE, 40B active, a real 1M-token context window, 28.5T training tokens, MIT weights. And Wafer showed AMD MI355X serving at 2626 tok/s/node, roughly 2x cheaper than NVIDIA Blackwell, using framework configuration rather than custom kernels.

The forcing function landed the same week. Alibaba banned Claude Code overnight, ordering all Claude models off work machines. Anthropic's Fable 5 was down 19 days on export-control issues. These stopped being hypotheticals.

The Practical Recipe

The AMD result is worth trusting because it is configuration, not engineering:

• MXFP4 quantization via Quark
• sglang as serving framework, which beat vLLM and ATOM in their testing
• FP8 KV cache for memory efficiency
• ROCm speculative-decoding fixes
• MoE kernel tuning for fp4 shapes

To benchmark before you touch hardware, Vercel's AI Gateway now exposes GLM 5.2 Fast as an endpoint.

The MoE Memory Trap

Capacity planning is where the sparse MoE numbers mislead. LongCat-2.0 activates 48B params per token, so inference compute tracks a 50B model. The thing that number doesn't tell you is that you still have to hold all 1.6T weights in memory. Your tok/s bill reads mid-size and your VRAM and storage bill reads enormous. Size the full parameter set before committing.

Cross-Source Validation

The SWE-bench Pro gap of 59.5 to 58.6 is sub-one-point and vendor self-reported, with no confidence intervals. That is a leaderboard result, not a production verdict. Do not put 'beats GPT-5.5' in a design doc without re-running on your own harness. The delta is noise. What is not noise is that an MIT-licensed model now sits in the same performance band as a proprietary frontier model, and that is what makes self-hosting a real option.

Local models now answer 71.3% of ChatGPT-style queries correctly, up from 23.2% in 2023 per Stanford, and Qwen 3.6 27B runs at 32 tok/s on an M5 MacBook. Frontier capability for the hard tasks via self-hosted MoE, viable local inference for the routine majority.

The default architecture of piping everything to a frontier cloud API is now leaving 50-60% of cost on the table, plus geopolitical risk you don't control.

What Landed

Google Research released TabFM, a foundation model for tabular data that wants to do for structured prediction what GPT did for text: handle new datasets zero-shot, with no per-dataset training. The architecture uses alternating row and column attention to capture feature interactions natively, plus row compression and an ICL (in-context learning) Transformer. It was pre-trained on hundreds of millions of synthetic datasets generated from structural causal models and evaluated across 51 TabArena datasets.

If it generalizes to your data, the payoff is real. The usual loop of feature-engineer, train a GBM, tune hyperparameters, retrain on drift collapses into one inference call with your data as context.

Why You Shouldn't Believe It Yet

Two sources flag the same gaps, independently.

The non-commercial license is a hard production blocker. The 10-class classification cap rules out many real tasks. And there is no direct comparison against a properly tuned gradient-boosted tree on the same datasets. The thing this doesn't tell you is whether it beats the baseline you already run, which is the only comparison that decides a migration.

The Right Experiment

Treat this as a one-day spike, not a migration plan:

1. Grab 3-5 of your real tabular datasets spanning classification and regression
2. Run TabFM zero-shot on the same holdout your GBM baselines use
3. Measure both accuracy (AUC/RMSE) and inference latency, because zero-shot moves the cost from training to inference and you pay it on every call
4. If it lands within ~2% on multiple datasets, you have a prototyping baseline that skips training entirely
5. If not, you spent a day and confirmed GBMs still own structured data

The architectural idea, row plus column attention over tabular features, is worth understanding whether or not this specific model ships. It points at a future where per-dataset retraining becomes optional. TabFM may not be the version that gets you there.

TabFM is the first credible attempt to make 'train a model on your tabular data' feel as antiquated as 'write a parser for your text data' — but the non-commercial license means it stays in your research notebook, not your feature store.

The Spread Is Faster Than You Think

A clean event-study design shows that following each major US export control since 2022, China-linked developers added 0.143 additional forks per repo-week versus 0.012 for US developers — an 11x gap. Qwen and DeepSeek now diffuse globally almost as fast as the best US frontier models. This isn't an emerging trend — it's established fact with rigorous measurement.

The compliance problem: Chinese-origin models are already appearing undisclosed in US patents. Chinese domestic science underlying US patent citations jumped from 1% in 2000 to 26% in 2025. If your team is using open-weight Chinese models without provenance tracking, and any output feeds into IP filings or regulated decisions, you have unmanaged legal exposure.

Both Directions Cut

Geopolitical risk isn't one-directional. Alibaba banned Claude Code and removed all Anthropic models from work machines — demonstrating that US tools can be yanked from Chinese orgs overnight. Simultaneously, Anthropic's Fable 5 was offline 19 days on export-control grounds, showing US tools can be yanked from non-US users by regulation. Both events happened this month.

The new models reinforcing this pattern:

The chip claims (50K domestic cards, zero Nvidia) are unverified and strategically motivated. But for your purposes, what matters is the runtime may be tuned for non-CUDA hardware — benchmark throughput on your own serving stack.

The Governance Gap

Most ML teams track model performance in their registry but not model provenance. You need to know:

• Which models in your serving and eval stack originate from Chinese labs?
• Are any outputs flowing into patent filings, regulated decisions, or government contracts?
• Does your legal team know which weights are running?

This isn't about avoiding Chinese models — several are genuinely best-in-class and MIT-licensed. It's about having an audit trail before regulators or legal asks. The provenance field costs nothing to add; the lack of one could cost everything in a compliance review.

The Chinese open-weight models you dismissed last year now diffuse as fast as frontier US models — benchmark them on your own data, but track their provenance like your compliance depends on it, because it does.