Fireworks AI Powers Cursor's Composer 2 With Distributed RL Across Global Clusters

Fireworks AI has published details on how it supports Cursor's Composer 2, a coding model built on the Kimi 2.5 base. The model is trained on long-horizon software tasks, including tool use, debugging, terminal execution, and multi-file edits. According to Fireworks, Composer 2 achieves a CursorBench score of 61.3, a Terminal-Bench score of 61.7, and a SWE-bench Multilingual score of 73.7. The company also reports that inference cost is 6 to 10 times lower than comparable frontier coding models.

Distributed RL and Infrastructure

The reinforcement learning runs are distributed across three to four global clusters. Fireworks supplies the inference and rollout layer that enables Cursor's distributed RL. Model synchronization between clusters uses compressed weight updates rather than full parameter transfers, achieving roughly 98 percent or more reduction in transfer size. Synchronization latency between clusters is on the order of minutes. Stable rollout fleets support large mixture-of-experts models, and low-latency inference is available during both training and evaluation loops.

Production inference is reused during training to speed up RL sampling. This approach reduces the need for Cursor to build a dedicated inference stack, according to Fireworks. The infrastructure quality is tightly coupled to model quality in large-scale RL, and the system impact includes faster iteration cycles, reduced hyperscaler dependence, and higher reliability.

RL as a Product Mechanism

Cursor uses RL as a product mechanism to shape real-world behavior. Federico Cassano, Research Lead at Cursor, says mid-training teaches code generation, while RL teaches correct code execution in real systems. Each RL rollout mimics a full development session, requiring fast inference and faithful environments. RL improves multi-step reasoning, terminal and tool success rates, best-of-K performance, and workflow stability.

Composer 2 targets a single environment: software engineering inside Cursor. This focus allows the RL training to be tightly aligned with the actual deployment context. The combination of distributed training, compressed synchronization, and reused inference is presented as a way to scale RL for coding models without the usual infrastructure overhead.

For inference engineers, the key takeaway is that RL training at this scale depends on the same infrastructure that serves production traffic. Reusing production inference for sampling cuts cost and latency, while compressed weight updates keep clusters in sync without saturating network links. The reported 6 to 10 times lower inference cost suggests that the infrastructure choices have direct economic impact for model builders.

All scores and cost figures are Fireworks AI's own claims, not independent measurements. The technical details, such as the 98 percent transfer reduction and minute-level synchronization, are also from the company's description. Readers should treat these as vendor-reported numbers until verified by third parties.

For teams building their own RL pipelines, the model here is clear: invest in inference infrastructure that can double as a rollout engine, and design model synchronization to minimize data transfer. The Cursor example shows how a single-environment focus can make RL training more efficient, but the same principles could apply to other narrow domains where the deployment context is well defined.