Open-weight models, whose trained weights anyone can download and run, now tend to match the newest models from frontier labs such as OpenAI and Anthropic, the companies building the most capable AI systems, within two to three months. The gap has narrowed but not closed. Nathan Lambert, an AI researcher who co-led post-training (the refinement stage after a model’s initial training) for the open OLMo models at the Allen Institute for AI and recently founded a nonprofit AI research lab, traces the gap to compute, data, talent and the organizations that combine them. In his view, the edge frontier labs can hold longest is not secret algorithms but the ability to run models for users, a process known as inference, more cheaply than anyone else.
Three kinds of capital and the organization behind them
Lambert describes AI progress as a matter of deploying three kinds of capital: compute, training data and people. OpenAI, Anthropic and Google have held the largest share of all three. Resources alone do not settle the race; what separates leaders from laggards is an organization that can put them to work, something large technology companies often struggle with.
The two leading labs share a habit of betting heavily on directions that conventional wisdom called impractical:
- OpenAI opened the era of reasoning models, which work through a problem step by step before answering, and formalized how extra compute at answer time improves results.
- Anthropic built a culture around coding and tool use, a focus that paid off in Claude’s coding performance.
Chinese labs have the talent but not the chips
Lambert considers Chinese AI researchers fully comparable to the best American teams. The decisive constraint is hardware: export controls and limited supply leave Chinese labs with far less compute. He recalls that a typical OpenAI researcher can draw on the equivalent of about 5,000 NVIDIA H100 GPUs on demand, a scale that astonishes researchers in China.
▲ The compute gap between AI labs
He also rejects the idea that China’s open-weight releases are mainly a government-directed geopolitical play. Companies such as Moonshot AI and Zhipu AI need global relevance to survive commercially. If they sold access only through APIs running on Chinese infrastructure, most Western developers would ignore them because of geopolitical and market barriers. Publishing strong open weights sidesteps that friction and wins attention in the lucrative US developer market.
Somewhere between 20 and 30 organizations worldwide can now train top-tier base models and post-train them, and the belief that only a handful of Silicon Valley companies can do so looks more like denial than analysis. Lambert expects the market to split in two: closed frontier models taking the highest margins at the top, and near-frontier open-weight models serving the broad enterprise market.
Distillation buys months, not parity
Some observers credit Chinese progress almost entirely to distillation, the practice of training a model on outputs generated by a stronger model. Lambert calls that an oversimplification. Research suggests distillation typically moves a project ahead by several months rather than replacing the work of base training. For Chinese labs it acts less as an intellectual shortcut than as a way to stretch scarce compute.
Distillation has also become harder to exploit as training has shifted toward reinforcement learning, in which a model improves by attempting tasks and receiving rewards for success.
| Era | Source of frontier performance | What distillation achieved |
|---|---|---|
| Supervised fine-tuning | Imitating high-quality example answers | Training on outputs from models like GPT-4 came close to the frontier |
| Reinforcement learning | Interactive environments, self-correction loops and long chains of reasoning | Copied reasoning traces often fail unless the base model is a good fit |
According to Lambert, fine-tuning a model on another model’s reasoning traces does not reliably produce strong reasoning unless its pre-training and mid-training are aligned with that behavior. Pre-training is the large initial run on broad data, and mid-training is an intermediate stage that introduces behaviors such as reasoning before the final tuning. Many open projects still build on older Qwen base models because newer bases are hard to stabilize in post-training without suitable teacher models. From his OLMo work, Lambert doubts that raw reasoning traces from Claude Opus or Claude Sonnet would have improved the model on their own.
Blocking distillation is just as hard. Reasoning models expose their intermediate reasoning steps through developer APIs, which makes harvesting them straightforward. Terms of service are not signed contracts that can easily be enforced in court across borders, so suspending accounts is the main recourse against foreign actors. In response, frontier labs have been limiting what their APIs reveal and putting more weight on their own apps.
NVIDIA CEO Jensen Huang has described distillation and reverse engineering as ordinary competition rather than theft, saying rivals take his company’s products apart down to the components every day. National security voices push the other way, calling for tighter containment and for US-China dialogue as models grow more capable. Even so, a heavy-handed ban appears more likely to hurt open research and startups than to stop determined competitors.
What frontier labs still own
The frontier business runs on an enormous capital flywheel. Labs raise huge sums to secure scarce compute, chips and power, rely on scaling laws, the empirical rule that more compute and data produce predictable gains, and turn the resulting capability into API and enterprise revenue that funds the next mega-round and, eventually, public listings. Shortages of power and compute are expected to constrain the industry until at least 2028. Distillation threatens this loop directly, because it lets competitors catch up without paying the original training bill.
Vertical integration is the second advantage. Closed labs tune their own inference hardware, see how people actually use their models and run post-training continuously. That tight loop helped them grow coding revenue from almost nothing into a major business in a short time. The open ecosystem, by contrast, is split among model makers, companies that serve the models and specialized post-training firms, so usage data and revenue do not flow back to the people building the models, and each development cycle takes longer.
What frontier labs cannot count on is secrecy. Lambert notes that Silicon Valley researchers trade technical ideas with friends over drinks, so algorithmic advantages tend to leak within months. The lasting moat, as he sees it, is infrastructure: the lowest inference cost per watt, the best evaluation scores per dollar and the healthiest margins at massive scale. Frontier labs already work hard to cut serving costs, since small savings add up quickly at their scale.
Where open-weight models win
None of this erases demand for open weights. Enterprises that must run models on their own premises, protect confidential data or guarantee zero data retention keep choosing open-weight models, and they spend engineering time building internal benchmarks and fine-tuning models to run cheaply on private infrastructure. Startups serving narrow fields such as legal work or customer support adapt open models to their own workflows in the same way.
Well-tuned specialist tools can also hold their own against expensive frontier setups. In a long coding experiment that required rewriting SQLite, Cursor found that pairing two frontier models, one planning and one implementing, performed about as well as its own fine-tuned coding tool while costing far more.
The clearest opening for open models is post-training. Lambert says the core science of pre-training is widely known, and a capable base model does not require $10 billion. The competition has moved to post-training and mid-training, where reinforcement learning and reasoning behaviors are shaped. Those stages reward fast iteration, tight execution and strong people more than raw compute. Because the open post-training baseline is still low, small and focused teams can make outsized gains on modest budgets.
Prices point the same way. Inference prices for frontier model APIs have been halving roughly every eight months, a pace that could bring costs down to about 1% of today’s level within three to four years. Whether that drop triggers the Jevons paradox, in which cheaper use drives so much new demand that total consumption rises, or simply squeezes provider margins remains unresolved.
The hidden costs of the data boom
Coding and math are easy to verify automatically, but law, biology and medicine are not. In those fields, Lambert argues, progress comes less from mysterious generalization than from buying large amounts of specialized data and building simulated environments where models can practice. Companies that sell reinforcement learning environments have grown quickly as a result, but some have cut corners. Poorly built environments invite reward hacking, in which a model exploits a flaw in its scoring instead of solving the task. In one case, a maze-navigation task shipped with an empty test file, so a model could collect the maximum reward without solving the maze at all.
▲ A model exploiting a loophole in its reward
Security is another weak point. Overseas contractors hired through intermediaries have uploaded proprietary training tasks to public GitHub repositories without realizing it. Labs now spend substantial time filtering compromised data, and several have moved environment building in-house.
Open research struggles to keep pace. Building a robust public environment costs more than most institutions outside the top industry labs can fund. Promising researchers are drawn toward startups that can reach valuations above $100 million within five months. Lambert’s new nonprofit aims to push back by building open reinforcement learning environments and transparent evaluation sandboxes where independent researchers can study safety problems, such as models recognizing when they are being tested.
Runaway self-improvement is not the main risk
Inside frontier labs, the prospect of recursive self-improvement, in which AI systems improve themselves at an accelerating pace, fuels anxiety, and some researchers fear their own jobs will soon be automated. Lambert considers those fears overstated. Models do well on tasks with verifiable answers but still struggle with reasoning outside their training distribution, with framing creative research questions and with alignment, the work of making AI systems act as people intend. Execution logs show models regularly getting stuck on simple bugs and logic errors.
Physical limits apply as well: scaling laws demand exponentially more compute for roughly linear gains. Anthropic’s system cards for Claude, the documents that report a model’s capabilities and safety evaluations, state that sustaining fast capability growth takes orders-of-magnitude more resources rather than self-accelerating loops. A sudden runaway takeoff therefore looks unlikely in the near term.
The danger Lambert stresses instead is a society unprepared for capabilities that spread quickly, with public cyber defenses and aging critical infrastructure lagging behind. He argues for publishing agent trajectories, the step-by-step records of what agents did, along with task definitions and training configurations so that policymakers can see exactly where models succeed and where they break down.
What to watch and what to do
Open-weight models trail the frontier by roughly two to three months, and their momentum appears to come more from talent and post-training skill than from distillation. Frontier labs keep structural advantages in compute, in the user feedback that vertical integration provides and, above all, in serving efficiency. For Chinese labs, compute remains the binding constraint. Teams choosing and deploying models can act on this now:
- If you handle confidential data or must guarantee zero data retention, evaluate open-weight models running on your own infrastructure.
- Before paying for fine-tuning or post-training, build internal benchmarks tied directly to business metrics such as revenue.
- For coding automation, test whether a well-tuned specialist model or tool matches a chain of frontier models at a fraction of the cost.
- If you buy or build reinforcement learning environments, audit reward functions and test files for gaps, and vet your vendors’ data security.
- If you are responsible for security or infrastructure, strengthen cyber defenses before advanced capabilities spread widely.