Summary
Fahd Mirza puts the newly released Spark X2.5 4B through a full local deployment and capability evaluation on a 48GB VRAM GPU. The model is a 4-billion-parameter open-weight release under the Apache 2.0 license, trained on approximately 20 trillion tokens, supporting over 200 languages and native agentic harness compatibility via Hermes and Qwen 3 tool-call formats. Its headline architectural claim is a hybrid attention design using three sliding-window attention layers for every one full global attention layer, which the model card argues allows a native one-million-token context window at a fraction of the usual compute cost.
Mirza serves the model using a Docker-pinned SGLang image exposing an OpenAI-compatible API on port 30,000, capping context at 65k tokens to fit available memory — with actual VRAM consumption measured at just under 42GB once the KV cache is active. Two primary evaluations follow. First, an autonomous bug-fix challenge on a real four-tier stack (PostgreSQL, FastAPI, Nginx, Docker Compose): the model is given a goal via a Hermes agent harness, reasons for 16–17 minutes, and correctly identifies and patches an emergency dispatch triage ordering bug that was sending crews to the smallest incidents first. Second, a pure-HTML rotating donor kebab rendered from scratch without libraries — testing procedural generation capability at small scale.
The verdict is nuanced: the 4B model matches larger models on correctness for both tasks but takes three to four times as long. Mirza frames this as a meaningful result for teams constrained by hardware budget who can tolerate higher latency.
📺 Source: Fahd Mirza · Published September 06, 2026
🏷️ Format: Review
