Pelican1.0-VL-235B-A22B-FC

Pelican1.0-VL-235B-A22B-FC is the largest member of the Pelican-VL 1.0 family, published by the Beijing Innovation Center of Humanoid Robotics (X-Humanoid) on 7 February 2026 — the state-backed centre behind the Tiangong humanoids. Like the rest of the family it is an embodied brain model: it looks at a scene and reasons about space, order of operations and grasp points, rather than holding a conversation. Where the earlier 72B build was dense, this one is a mixture of experts. It holds 235.67 billion parameters in total across 94 layers and 128 experts, of which eight are routed per token, so roughly 22 billion parameters do the work on any given step — that is what the A22B in the name records. The architecture moves from Qwen2.5-VL to the Qwen3-VL mixture-of-experts design, and the context window grows fourfold to 262,144 tokens. The FC suffix marks the addition the vendor considers the point of the release: function calling, so the model can drive external tools rather than only describe what should happen. Two qualifications belong on the record. First, the vendor labels this checkpoint a preview optimised for tool use, not a finished release, and publishes no separate benchmark table for it — the numbers in the Pelican-VL report cover the 7B and 72B builds. Second, the model card that carries it still describes the family as spanning 7B to 72B, which was accurate in November 2025 and has not been updated since this build appeared. The licence is Apache 2.0, the same as the rest of the family, and the weights can be downloaded from Hugging Face and ModelScope. Running them is another matter: at this size a self-hosted deployment needs a multi-GPU server, and the centre publishes no hosted API or pricing.

▸Videos

No videos yet.

▸Reviews