4B / 8B dense language backbones; 4,437,815,808 / 8,767,123,696 total checkpoint parameters including vision encoder and mergers
262,144 tokens native (shared text/image/video context); up to 1M advertised with explicit YaRN extension, not the default
Apache 2.0 (checkpoint declarations and explicit family weight grant; standalone LICENSE absent from the four weight repositories)
Four separate BF16 checkpoints pair dense Qwen3-derived 4B and 8B decoders with Instruct or reasoning-focused Thinking post-training. Both decoders have 36 layers and grouped-query attention with 32 query heads and 8 KV heads; the 4B ties its input/output embeddings, whereas the 8B does not. The 4B uses a 24-layer, width-1024 SigLIP2-Large-derived vision encoder with DeepStack features from layers 5/11/17; the 8B uses a 27-layer, width-1152 SigLIP2-SO-400M-derived encoder with features from layers 8/16/24. A main vision-language merger plus three DeepStack mergers integrate multilevel visual features; interleaved MRoPE and explicit textual video timestamps support spatial-temporal modeling. Inputs include text, images and sampled video frames, with generated text, document parsing/OCR, 2D/3D grounding, image-to-code and GUI-agent use cases; native audio input/output is not supported, and GUI execution requires an external agent runtime. The retrospective November 26 technical report describes merger-only alignment followed by full-parameter multimodal training at 8,192, 32,768 and 262,144 tokens; post-training uses standard or long-CoT SFT, strong-to-weak distillation for lightweight students using teacher responses and on-policy student/teacher KL alignment, then reasoning/general RL. Its family-level token budgets are not separately audited 4B/8B training totals. Reported size-specific Thinking scores are AIME-25 74.5/80.3 and LiveCodeBench v6 51.3/58.6 for 4B/8B, with an 81,920-token output allowance on these evaluations; corresponding Instruct scores are 46.6/45.9 and 37.9/39.3. Thinking is not uniformly better: 8B BFCL-v3 is 63.0 versus Instruct 66.3. The publisher advertises expanded multilingual OCR and optional 1M context, but its 32-of-39-language OCR threshold and 1M long-video needle-retrieval headline do not establish equivalent small-model accuracy or lossless full-resolution video comprehension. Official FP8 versions are public related packages, excluded from this BF16 release grouping because their exact public-download date is unresolved. The October 15 date is provisional: the official download-linked release statement was committed October 14 UTC, and contemporary October 15 local 4B execution corroborates availability by then, not the exact first unlock. The October 4 30B-A3B and October 21 2B/32B releases are separate events.