OpenAI’s open-weight GPT-OSS models land on Workers AI

OpenAI has released its first open-weight models, and Cloudflare is a Day 0 launch partner bringing them to the Workers AI platform. The models are available immediately under the model identifiers @cf/openai/gpt-oss-120b and @cf/openai/gpt-oss-20b. This addition reinforces Workers AI’s focus on open models, giving developers options that favor transparency, customizability, and flexible deployment while offering enterprises a path to full data control.

Both model variants are Mixture-of-Experts (MoE) architectures: a 120-billion-parameter version and a 20-billion-parameter version. MoE models route queries to relevant experts rather than activating all parameters, which improves efficiency. Notably, these models operate natively at FP4 quantization, significantly reducing GPU memory footprint compared to a similarly sized FP16 dense model. The combination of quantization and MoE design results in faster and more resource-efficient inference than traditional dense models of comparable scale.

The models are text-only but include reasoning capabilities, tool calling, and built-in support for Code Interpreter and Web Search (Web Search support is coming soon). Cloudflare has implemented Code Interpreter atop Cloudflare Containers, enabling stateful code execution—a novel approach detailed below.

API compatibility and endpoints

Workers AI supports both OpenAI’s recommended Responses API format and the legacy Chat Completions API format (the latter coming soon). The endpoint behavior depends on how the model is invoked:

  • Workers Binding: accepts and returns Responses API format — env.AI.run("@cf/openai/gpt-oss-120b")
  • REST API /run endpoint: accepts and returns Responses API format — https://api.cloudflare.com/client/v4/accounts/<account_id>/ai/run/@cf/openai/gpt-oss-120b
  • REST API /responses endpoint: accepts and returns Responses API format — https://api.cloudflare.com/client/v4/accounts/<account_id>/ai/v1/responses
  • OpenAI-compatible endpoint: returns Chat Completions format (coming soon) — https://api.cloudflare.com/client/v4/accounts/<account_id>/ai/v1/chat/completions
curl https://api.cloudflare.com/client/v4/accounts/<account_id>/ai/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $CLOUDFLARE_API_KEY" \
  -d '{
    "model": "@cf/openai/gpt-oss-120b",
    "reasoning": {"effort": "medium"},
    "input": [
      {
        "role": "user",
        "content": "What are the benefits of open-source models?"
      }
    ]
  }'

Stateful Code Interpreter with Sandboxes

LLMs tend to struggle with logic-heavy tasks like mathematics and coding. Rather than attempting to reason through such problems, they commonly issue a tool call to execute AI-generated code. OpenAI’s new models are explicitly trained for stateful Python code execution and ship a built-in Code Interpreter feature aimed at this use case.

Cloudflare’s broader developer platform—spanning compute, storage, and inference—makes it uniquely positioned to support more than one-off code execution. Code Interpreter runs on Cloudflare Sandboxes, a secure execution environment built on the Containers product. When Code Interpreter is invoked, the platform spins up a Sandbox container scoped to the user’s session, kept alive for 20 minutes so code can be edited and reused across subsequent model queries. Sandboxes for Code Interpreter are pre-warmed to minimize startup latency.

A working example demonstrating how to use gpt-oss on Workers AI with Sandboxes and the OpenAI SDK is slated for publication in the Developer Docs. Efficient serving at launch is supported by contributions from vLLM and HuggingFace. Full model details and integration guidance are available in the Developer Docs.

BLOG-2912 3