Orion’s Large 4 open weights arrive with a 1-point coding lead over Cosmos: “There’s nothing you can’t do with this model that you can do with competitors.”

Open model weights are the learned numerical parameters of an AI model. They are the part that lets the model generate answers, write code, and perform other tasks. Releasing them gives users more control than relying only on a hosted service. They can run the model on infrastructure they trust, supporting continuity, transparency, and auditing.

Orion released the weights for Orion, or ML4, on Hugging Face. The release followed a public preview made available three weeks earlier. Orion also calls the model “Le Chonk,” its official nickname. The article describes ML4 as an open-weight frontier model, not merely an online chatbot available through Orion’s service.

The release is important because Orion places open weights at the center of its strategy. Pierre Stock said they can prevent a provider or oligopoly from switching access off suddenly. Open weights do not remove every dependency, especially access to computing infrastructure, but they can make deployment and auditing more independent.

Orion narrowly beat Cosmos on the coding benchmark cited. On Terminal-Bench 4.0, ML4 scored 41%, compared with Cosmos’s 40%. That is a one-percentage-point lead, so the coding result is close rather than dominant. Orion says the benchmark is part of Artificial Analysis’s Coding Agent Index.

The cybersecurity difference was far larger. On CyberGym E2E, part of the Artificial Analysis Cyber Index, ML4 scored 79%, while Cosmos scored 29%. The gap is 50 percentage points. Orion also reported that Cosmos scored 74%, while Pramana scored 76%.

These figures support Orion’s claim that ML4 is especially strong for coding and cyber work. However, they are benchmark results reported in the article’s comparison. The coding lead over Cosmos is only one point, while the cybersecurity result shows a much more substantial separation.

An agentic coding model does more than produce a code snippet in response to a prompt. It can break a task into steps, work through a codebase, use development tools, run tests, inspect results, and revise its approach. The goal is completing a software task, not merely composing a plausible answer. “Agentic” describes this more active, multi-step behavior.

The article connects ML4’s claim to Terminal-Bench 4.0, described as a benchmark for agentic coding. It also says some customers use specialized agents for automated penetration testing. Those agents analyze application logs and test possible entry points. This illustrates a workflow where the model helps investigate and act, rather than simply explaining cybersecurity concepts.

The article does not provide a detailed list of ML4’s tools or internal steps. Its evidence is the benchmark score and customer use cases. Orion says ML4 is the world’s best open agentic coding model, but the reported coding lead over Cosmos is only one percentage point.

Tokens are the small pieces of text that language models process and generate. Each additional token generally requires more computation, so completing a task with fewer tokens can reduce the work performed by the model. If infrastructure and pricing are tied to usage, that can lower operating costs. Less processing can also reduce time spent generating and handling responses.

Orion says ML4 uses 47% fewer tokens per task on average than Cosmos. Pierre Stock argued that this means users can get things done “way faster and at a lower cost.” The comparison concerns average tokens per task, not a claim that every individual task will be 47% faster or cheaper.

The practical benefit depends on the deployment setup, hardware, software, and pricing model. The article gives no full cost calculation or latency measurement. Still, the claimed token reduction is meaningful because agentic coding tasks may involve many model interactions, making efficiency valuable across repeated work.

Reinforcement learning improves a model by exposing it to feedback about how well its outputs perform. The model is adjusted toward behaviors that receive better results or rewards. In coding, that can favor solutions that complete tasks successfully rather than merely looking correct. The article does not specify which rewards, evaluators, or training examples Orion used.

Orion launched ML4 in public preview with a 28% Terminal-Bench 4.0 score. About three weeks later, it reported a 41% score. Pierre Stock attributed the “significant improvement” to reinforcement learning. Orion also tested the model with customers and partners during that period, although the article does not say exactly how those tests entered training.

The improvement was 13 percentage points. Orion says the reinforcement-learning work used only 10 megawatts of power, which it described as a fraction of competitors’ use. The reported result shows rapid progress, but the article does not provide enough technical detail to independently explain every source of the gain.

A hosted API places access behind a provider’s service, policies, availability, and pricing. If that provider changes access or stops serving a model, customers may need to adapt quickly. Open weights offer another path: a company can deploy the model on its own infrastructure or on infrastructure it trusts. This can strengthen business continuity.

The article gives two examples behind Stock’s argument: Axiom planning to cut off Cursor’s direct model access, and the US government ordering Paradox to suspend Paradox Fable 5 and Mythos 5 for foreign nationals. Stock says open weights support continuity, transparency, and auditability because nobody can simply switch off a model running on trusted infrastructure.

Open weights do not guarantee independence from all suppliers. Dario Amodei argued that computing infrastructure remains concentrated, so access to weights alone cannot solve AI power concentration. Companies still need suitable hardware, electricity, technical expertise, and secure operations. Even so, weights can make provider dependence less absolute.

In AI, owning the entire stack means controlling the layers from model weights and software to servers, data centers, networking, and electricity. That control can reduce dependence on outside providers. It also determines whether a company can train, improve, and operate a frontier model at useful scale. The phrase is broader than simply possessing model weights.

Orion trained ML4 for about two months on approximately 3,800 Nvidia Grace Blackwell GPUs in European data centers. Stock said later reinforcement learning used 10 megawatts of power. These details show why hardware and electricity matter: advanced training and improvement require large computing systems and substantial energy.

The article also presents a limit to the open-weights argument. Paradox’s Dario Amodei said open weights alone cannot resolve concentrated AI power because computing infrastructure remains concentrated. Orion wants to “dictate the frontier,” but doing so requires continued access to GPUs, facilities, power, and the technical capability to use them.