# What comes next with open models

Nathan Lambert wrote this for [[interconnects]] in March 2026. It is a business piece. After DeepSeek R1, many companies began treating open models as a route to influence in AI, but Lambert argues very few have a real monetary reason to build them, that the gap to the closed frontier is more likely to grow than shrink, and that effort in open models is aimed at the wrong target. On his [[open-source-ai-reading-list]] he describes it as the role open models will play as a complement to strong closed ones, used to build custom agentic workflows in enterprises.

## Why the gap will grow

Lambert puts open models at 6 to 18 months behind the best closed models, and says he and other analysts are surprised the gap has stayed that small given the difference in budgets. He expects it to widen. The top labs keep shipping strong models, and many of their gains do not show on public benchmarks. Distillation used to mean training a student on a teacher's full completions; with coding agents, what matters is the RL environments and the prompts that place agents in them, and those are much easier to hide (see [[how-much-does-distillation-matter-for-chinese-llms]]). As models move into long-horizon work in fields gated by expensive institutions, such as law and healthcare, the training data is not on the public web. Coding can stay close, since GitHub scrapes and clever environments go a long way. Open models winning outright would need a hidden breakthrough or a hard wall in progress, and Lambert rates both very unlikely.

His range is wider than the roughly six months he gave a month earlier in [[open-models-in-perpetual-catch-up]], and than the four to six months on his September 2026 reading list. [[open-closed-model-gap]] collects the measurements.

## Three classes of model

The core of the piece divides the future into three kinds of model. True closed frontier models will run the strongest knowledge-work and coding agents. Open frontier models, the best large open-weight models, will fall short on some tasks but work very well on many others, including parts of coding; the class around GPT-OSS 120B, Nvidia Nemotron 3 Super and MiniMax M2.5 is the balance of price and performance that can run locally, as private assistants served at the cost of electricity. The third class, small open models as distributed intelligence, is the one Lambert expects to be the most successful, and the one almost nobody is building.

## Systems, not weights

AI products are weights plus tools plus a harness, and Lambert asks when anyone was last impressed by raw token output alone. That favors closed labs, which can integrate the chips, inference software, weights, tools and interface. His examples are [[claude-code]] with Opus 4.6 and Codex with GPT-5.4. An open model has to work on every inference setup and with many harnesses. Some are starting to target one interface such as [[opencode]], but tying a model to one product sits awkwardly with being open. The same shift could cut the other way: if weights alone do less, labs may be more willing to release them.

He reaches back to two older texts. Jonathan Rosenberg's 2009 Google memo "The Meaning of Open" says that in an open system the winner is whoever understands it best and innovates fastest, not whoever locks customers in. Bill Gurley's 2011 post on Android and Chrome calls them expensive moats funded by the search castle, and points out that Rosenberg exempted search and ads from openness. AI has no equivalent yet. A model such as GLM 5 served with vLLM is only part of a system, and deploying it still takes private GPUs and local business data. If a fully open AI system existed, Lambert suggests, it would win by default, as open systems did in earlier generations of technology.

There are early signs of Gurley's pattern. Moonshot and [[z-ai]] sell coding plans with high demand even though their weights are open, because most users want a cheap interface with inference rather than to run the model themselves. Lambert is still not optimistic that companies will become more open. Nvidia has one clear reason to release, selling GPUs to people who build on open models, and he sees no one else with an obvious reason. Without one, the builders at the open frontier will have less to spend and will consolidate to the best few.

## Small, boring, specialized

The recommendation is concrete. Enterprises want cheap, reliable models for repetitive tasks, and any task a frontier agent performs tens or hundreds of times could be handed to a small model that is 10x faster and 100x cheaper. Lambert pictures one small model with a set of LoRA adapters for internal skills, deployed as a tool that the frontier orchestrator calls, with side benefits such as a local model reading private files and summarizing them for Claude. He wants to build "open models that Claude Code is desperate to use as a tool". Such models should be "brain-numbingly boring and specific", excellent at one to three tasks, and measured by savings in compute and time rather than general benchmarks.

The evidence so far is thin but real. Qwen 3.5 4B arguably beats the original ChatGPT, research recipes exist for fine-tuning an open model on one codebase to match much larger models, and Moondream builds competitive small multimodal models on a tiny budget. Supply is the problem. Qwen, the best customizable small family, is uncertain after key staff left Alibaba, and Gemma, Phi and [[olmo]] are major steps down. The vault has a working example: [[rl-finetune-beats-frontier]], where a $500 fine-tune of a 9B model beat every frontier configuration on one narrow task at 68x lower cost. [[local-ai-is-not-opus]] reaches the same place from the other side, that a tuned local model earns its keep on specific work without being a general substitute.

## Where the essay lands

Lambert thinks many open model builders overrate their competitiveness, and that the consolidation coming for closed AI companies will reach them sooner. The open model world needs to become an ecosystem of many approaches, which he says China is closer to than the US, though still short on variety. Open models should solve problems the frontier labs don't, find places where they are effectively free alternatives, and do it before building strong models gets too expensive for today's open labs.

Three months later, [[open-and-closed-models-are-on-different-exponentials]] turns this into a prediction about value capture. Christian Catalini's [[some-simple-economics-of-open-versus-closed-ai]] arrives at a similar split from economic theory: commodity tokens served by open models, value-added tokens by whoever holds the data.

Bill Gurley's 2026 essay [[from-open-source-software-to-open-source-strategy]] disagrees with the premise. Lambert says open weights without an open system around them are not enough, and that Nvidia is the only company with a clean reason to release. Gurley treats open weights as the familiar open source playbook and expects coalitions of followers to fund them, as they funded Android and Kubernetes.