Open models in perpetual catch-up

Nathan Lambert wrote this for interconnects in February 2026, prompted by Z.ai's GLM 5, the latest open-weight release to set off talk that open models were about to match the closed frontier. Every 4-6 months some release triggers the same conversation, and over the previous year every model in it came from China, where it used to be one of Meta's Llamas. Lambert builds olmo at Ai2 and is one of open models' most prominent advocates, and he says the narrative is overblown: open models are not accelerating toward the best closed models, and a gap of about six months is holding steady. His open-source-ai-reading-list files the post as the explanation of why open models will always trail.

Why the chart looks closer than it is

The chart everyone cites is the Artificial Analysis Intelligence Index for open versus closed models. Lambert's objection is that one score compresses too many error bars and weaknesses to show the frontier, and the frontier has never been harder to measure publicly. Good benchmarks are now expensive and require knowing what the newest models do well. Known flaws, such as SWE-Bench being nearly three quarters Django or Terminal Bench 2 being crowdsourced and noisy, disappear into an average. The US frontier labs have repeatedly shown a better sense of which capabilities matter, and public benchmarks are easier to overfit. Qwen's flagship 3.5 model drew another round of benchmaxing complaints. This is why Lambert argues against averaging across the evaluation suite when presenting Olmo results. His bet is that the index misrepresents the true frontier, not that open models are closer than ever. The exception is coding, where public GitHub data and verifiable rewards leave open models plenty of room to keep up.

What does surprise him is that the gap is not larger. On paper Anthropic, OpenAI and Google have far more compute, purchased data and user data, and many experts expected a widening margin. He wonders whether the log-linear relationship between compute and performance in scaling laws explains it. The balance he describes is between the value of the smartest model, which he still pays for, and the cost collapse that follows once a permissively licensed open model can do a task. The best closed models keep unlocking more valuable tasks, and open models keep following.

Trends

The second half covers what Lambert calls the seven other big trends in open models. The section headings did not survive the clipping, so the grouping below is reconstructed from the text.

Adoption of open models is brutally concentrated. 2025 brought a "Cambrian explosion" of open models with strong benchmark scores, from far more builders than the four substantive closed API providers, and only the most successful ever get used. Models are stickier than coverage suggests, since many are set up once and never replaced, but the chance a new model is tried at all falls every month. Lambert was surprised to find Qwen dominated on adoption among the largest models. Small and mid-sized builders will have to move to a niche or a different business plan.

The niche he points to is specialized enterprise models, particularly for tool use and repetitive agent subtasks: cheaper, faster, private and owned, perhaps one model with a PEFT adapter per skill. Several enterprises building agents have asked him for exactly this. Open models are jagged, so one family will not cover it, and frontier-quality post-training recipes for adapting a model to a domain are scarce. By this standard today's math and biology models are not specialized enough. He develops the idea at length a month later in what-comes-next-with-open-models.

Open models are poorly understood, because they take a long time to be figured out and to spread. Lambert wants research organizations, even DC think tanks, dedicated to how open models work technically and geopolitically, and describes his work on Interconnects and atom-project-american-truly-open-models as digging up messy raw data on usage. OpenRouter usage data, shown as a chart the clipping lost, mirrors the trends in Hugging Face downloads, including on US versus China.

Sovereign AI will grow as every wealthy nation treats AI as a source of influence and a security need, and open models are likely the only practical way to build it.

The status quo is by far the most likely outcome, with the best open models 6-9 months behind. Lambert attributes a large portion of the catch-up to open builders distilling from the strongest closed APIs, but thinks this matters less as post-training shifts to RL, where a model learns from experience rather than from the smartest available teacher. Open models could only pull ahead through fundamental innovation, such as merging and sharing expert models or a 100x cut in training cost, which he declines to predict.

China has many labs building on each other's work. That deliberate sharing, often with close ties to academic institutions, lets standards converge faster than Silicon Valley's informal trading of secrets among friends, and Lambert writes that the US has no answer if it takes off. chinas-structural-advantage-in-open-source-ai covers the same structure.

Finally, the biggest influence of open models will be slow. One channel is research: when Llama was the best open model everyone did science on it, and now the default is Qwen, the anchor of the Chinese ecosystem. The other is global diffusion, since outside the wealthiest countries the default ways to use AI will be free apps like ChatGPT or open-weight models. Lambert admits that trying to understand the frontier through open models was partly a mistake, one that missed the slower story of how technology reshapes the world's largest companies.

Where the numbers disagree

The gap figure moves across Lambert's own writing. This post says about six months in its opening and 6-9 months in its forecast. A month later, what-comes-next-with-open-models says open models "have always been 6-18 months behind" and expects the gap to grow. His September 2026 reading list puts it at roughly 4-6 months, citing SemiAnalysis's are-open-models-catching-up. The spread depends on which benchmarks are used and what "behind" means, which is this post's own point; open-closed-model-gap and how-far-behind-are-open-models collect the measurements.

The role of distillation shifts too. Here it accounts for "a large portion" of the catch-up. how-much-does-distillation-matter-for-chinese-llms, from the same month, and the-distillation-panic in May argue that distillation helps Chinese labs without explaining away their progress, and the reading list's distillation notes add that reasoning-trace extraction may have mattered more for DeepSeek R1 than Lambert credited in 2025.

Mark Zuckerberg's 2024 claim in open-source-ai-is-the-path-forward that open models were advancing faster than closed ones is the view this post tests against two more years of data, and rejects.