Irfanuruchi/Nanbeige4.2-3B-CE-v1.0-OpenVINO-FP16

Nanbeige4.2-3B-CE v1.0 — OpenVINO FP16

OpenVINO FP16 stateful export of Irfanuruchi/Nanbeige4.2-3B-CE-v1.0 v1.0.

Provenance

• —Canonical source revision: v1.0
• —Canonical source commit: 171d26c743c6f1f5e3fd49f88c0454448a28c39e
• —Architecture: NanbeigeForCausalLM
• —Physical transformer layers: 22
• —Recurrent loops: 2
• —Effective KV-cache entries validated during export: 44
• —skip_loop_final_norm: false

OpenVINO validation

This FP16 stateful artifact was validated on Intel integrated graphics using OpenVINO 2026.3.1 and OpenVINO GenAI 2026.3.1.

Validated release properties:

• —Stateful GenAI interface: 3 public inputs (input_ids, attention_mask, beam_idx) and 1 public output (logits)
• —OpenVINO tokenizer and detokenizer are included
• —Hugging Face and OpenVINO token IDs were verified identical for the release prompt when both tokenize the already-formatted chat prompt with no additional special tokens
• —Real generation completed on Intel Graphics (iGPU)
• —The runtime path was validated with pre-tokenized inputs to avoid duplicate special-token insertion

Important quality note

Runtime validation is not a factual-quality benchmark. This release intentionally preserves the known limitations of the frozen Nanbeige4.2-3B-CE v1.0 checkpoint. Do not interpret an OpenVINO generation smoke test as a new semantic-quality pass.

Recommended prompt path

Apply the canonical Hugging Face chat template with enable_thinking=False, tokenize the resulting formatted prompt without adding another special-token layer, then pass the tokenized input to OpenVINO GenAI.

License

Apache License 2.0. See LICENSE.