Open models are pulling in two directions at once. By usage they are spreading faster than ever, mostly on Chinese weights. Measured against the frontier, they are falling slightly further behind, and the hardware that would let people run them at home is getting more expensive.
Key points
- The best open-weight models trail the closed frontier by about four months, a slightly wider gap than before, yet they carry most of the open traffic and a large share of scientific use.
- US export controls were meant to slow China. Instead they are now monetized and leaky, while China builds a separate chip stack that it anchors with open releases.
- Small open models are now good enough for real on-device work, but an AI-driven memory shortage is raising the price of the hardware they need.
- Record-low token prices make cloud inference the easy default, which undercuts local autonomy just as it becomes technically feasible.
- We think each of these tensions will likely resolve within 12 to 18 months, and the signals that settle them are already visible.
The tensions
These build on the patterns in our signal scan.
Are open models catching up or falling behind?
Falling behind. Epoch AI finds that since January 2026 the strongest open-weight models, all Chinese, have trailed closed frontier systems by about four months on average. Over 2023 to 2025 the lag was three months1. Commentators argue the real gap is larger, because open models are tuned to public benchmarks and closed labs hold back their best internal systems1.
Catching up, if you measure use. Chinese open-weight models carried about 61% of OpenRouter tokens by May 2026, and about 70% of new derivative models were built on Qwen2. In science, open weights appear in 44% of single-model-family papers, up from 12.8% in 20233. Four months behind turns out to be good enough for most of the work people actually do.
What would settle it: evidence of whether frontier capability matters commercially. Suppose closed labs start releasing capabilities that open models cannot match within two quarters, such as reliable autonomous cyber work4, and paying customers follow them. Then the gap is what counts. If usage share keeps rising while the gap holds, the frontier matters less than the price of being "good enough."
Do export controls slow Chinese open models or entrench them?
They constrain. Chinese labs still want Nvidia hardware. Firms reportedly ordered more than 2 million H200s once BIS moved to case-by-case licensing, and that licensing comes with a 25% tariff, a 50% volume cap and KYC checks5. The cap sets a real ceiling on training scale.
They entrench. The same rule turns controls into a transaction, and a US official says DeepSeek trained on top Nvidia chips despite the ban5. At the same time DeepSeek is moving onto Huawei Ascend, with a reported 160,000-chip order, and Huawei plans to double 910C output6. Open releases then make the resulting models the global default layer, which is the national strategy the USCC describes2. The controls push China to decouple while doing little to stop diffusion.
What would settle it: whether a frontier-competitive Chinese model trained mainly on Ascend ships in 2027. If one does, we would put the chance that controls have stopped constraining open releases at about 70%. If Ascend clusters stall on yield or software, the controls still bind.
Is local AI getting easier or harder?
Easier. Models between 3B and 10B parameters, including Phi-4-mini, Gemma and Qwen3, now handle chat, summarization and tool use on laptops and phones. Ollama's MLX backend roughly doubled decode speed on Apple Silicon7.
Harder. AI data centers may take up to 70% of global memory output in 2026. OEMs received only 50 to 66% of the memory they ordered, and Micron retired its consumer brand8. Meanwhile token prices hit a record low of $0.97 per million, and low-end frontier tiers cost about $0.109. The cloud keeps getting cheaper as the machine on your desk gets dearer.
What would settle it: consumer DRAM contract prices and the RAM that new laptops ship with through 2027. If 16GB stays the mainstream floor and prices ease, local inference becomes a mass option. If OEMs cut memory to protect margins, local AI stays a hobbyist and privacy niche, and the grid strain from data centers1011 keeps growing on the cloud side instead.
Why these questions matter
Anyone building on open models should treat the model lineage, the chip supply behind it and the hardware it runs on as three separate bets. Our advice: track Epoch's gap readings, Ascend-trained releases and DRAM contract prices each quarter. Keep deployments portable across model families, so that a sanction, a license change or a price shift costs you a migration rather than your product.
Sources
- Best open-weight models trail the closed frontier by about four months, and the gap has widened slightly · grey-lit ↩
- Chinese open-weight models dominate real-world token use and derivative models · institutional ↩
- Open-weight use in science reaches 44%, driven mainly by Qwen · peer-reviewed ↩
- Mythos-class AI finds more than 10,000 high/critical vulnerabilities across every major OS and browser · disclosure · also helpnetsecurity.com ↩
- US loosens H200 exports to China but adds tariffs, caps and know-your-customer checks · institutional ↩
- DeepSeek moves frontier-adjacent training and infrastructure onto Huawei Ascend · grey-lit ↩
- Small open models (3-10B) become capable of useful on-device agent work · expert ↩
- AI-driven memory shortage raises the cost of local-inference hardware · institutional ↩
- Token prices fall to record lows as efficiency gains and price wars compound · disclosure ↩
- NERC issues a rare Level 3 alert after 1,000+ MW data center loads drop off the grid in seconds · journalism · also powermag.com ↩
- DOE models a 100x rise in US blackout risk by 2030 as load outruns firm capacity · institutional ↩