Dense 2B / 32B; full BF16 checkpoints: 2,127,532,032 / 33,357,390,064 parameters, including vision encoder and four vision-language mergers
262,144 tokens native; optional extension to 1M with YaRN/configuration overrides
Apache 2.0 (explicit model-family weight grant and per-checkpoint hub declarations; standalone LICENSE absent from these HF packages)
Four separately released BF16 checkpoints: 2B-Instruct, 2B-Thinking, 32B-Instruct and 32B-Thinking. October 21 is the publisher-stated release date, retained provisionally: contemporaneous 2B local-installation evidence corroborates availability by that date, but exact first public access for all four is unresolved. Related FP8 packages are excluded from this same-day grouping pending historical timing and packaging review; the original 32B-Instruct-FP8 upload duplicated Thinking weights and was replaced on October 22. Dense Qwen3 text decoders accept text, images and sampled video frames and generate text; no native audio. Architecture combines interleaved MRoPE, textual video timestamps and DeepStack visual features injected into the first three decoder layers. The 2B decoder has 28 layers, width 2048, 16 query/8 KV heads and tied embeddings; its 24-layer, width-1024 vision encoder uses DeepStack indexes 5/11/17. The 32B decoder has 64 layers, width 5120, 64 query/8 KV heads and untied embeddings; its 27-layer, width-1152 vision encoder uses indexes 8/16/24. Native 256K context is shared across text, visual tokens and generation; advertised 1M requires positional-extension overrides, not the stock configuration. Publisher-described applications include OCR/document parsing, visual grounding, screenshot-to-code and GUI/tool-assisted agents; actual screen capture and actions require an external runtime. The retrospective technical report describes merger-only alignment (67B tokens), two all-parameter pretraining stages (~1T tokens each), then 100B-token 256K adaptation; exact per-checkpoint budgets are not separately disclosed. Post-training describes staged 32K-to-256K SFT with separate standard and long-CoT formats, strong-to-weak student distillation using teacher responses then on-policy KL logit alignment, and reasoning/general RL with SAPO. Small-model distillation is explicitly supported for 2B; the report does not establish an identical student recipe or exact optimization budget for every 32B checkpoint or the 235B teacher. Reported AIME-25 / LiveCodeBench v6 scores are 22.2/20.3 for 2B-Instruct, 39.0/29.3 for 2B-Thinking, 66.2/43.8 for 32B-Instruct and 83.7/65.6 for 32B-Thinking. Thinking uses an 81,920-token output budget on these reasoning evaluations, not an 81,920-token context limit. Reasoning is not universally better: 32B-Thinking scores 60.5 on GPT-4.1-judged Arena-Hard v2 versus Instruct 64.7; its LiveCodeBench advantage over text-only Qwen3-32B coexists with lower CFEval/OJBench scores. Reported hours-long video needle retrieval at 1M tests 235B-A22B-Instruct, not these sizes; no equivalent 2B/32B guarantee or universal minimal-VRAM claim is established.