Qwen3-VL-2B and Qwen3-VL-32B dense — Instruct + Thinking, FP8 variants (Alibaba Qwen)

TL;DR

Alibaba Qwen extends the Qwen3-VL dense family with 2B and 32B endpoints, completing the dense lineup (2B / 4B / 8B / 32B) alongside the existing 30B-A3B and 235B-A22B MoE tiers. Each new size ships in both Instruct and Thinking post-training variants, plus FP8 quantized checkpoints. The headline marketing claim is that the dense 32B outperforms GPT-5 mini and Claude 4 Sonnet across STEM / VQA / OCR / video understanding / agent tasks and matches the 235B MoE — even beating it on OSWorld.

Key claims

• Qwen3-VL-32B (dense) is claimed to outperform GPT-5 mini and Claude 4 Sonnet across STEM, VQA, OCR, video understanding, and agent tasks [tweet body].
• Qwen3-VL-32B (dense) is claimed to match Qwen3-VL models up to 235B parameters, and to beat them on OSWorld specifically [tweet body].
• Both Instruct and Thinking variants are released at each new dense size, matching the family-wide Instruct/Thinking pairing pattern [tweet body].
• FP8-quantized checkpoints are released alongside BF16 for ultra-efficient deployment [tweet body].
• Released on 2025-10-21 (per the tweet timestamp), about a week after the 4B/8B dense announcement on 2025-10-15 [tweet timestamp].

Method

Tweet-only artifact; no architectural detail in the source. The Qwen3-VL family’s published upgrades over Qwen2.5-VL are Interleaved-MRoPE (positional embedding allocated across time/width/height), DeepStack (multi-level ViT feature fusion), and Text–Timestamp Alignment for event localization, with 256K native context (expandable to 1M), 32-language OCR, and 2D/3D grounding. The 2B and 32B checkpoints inherit this stack; the tweet emphasizes “ultimate performance per GPU memory” as the design target.

Results

No benchmark numbers in the tweet beyond the qualitative claims above. OSWorld is called out specifically as a benchmark where dense-32B beats the 235B MoE; STEM / VQA / OCR / video / agent are aggregated into the GPT-5 mini and Claude 4 Sonnet comparisons.

Why it’s interesting

This completes the dense ladder that started with Qwen3-VL compact dense — 4B and 8B (Instruct + Thinking) (4B/8B): the 2B endpoint is the smallest “full Qwen3-VL capability surface” model and the 32B is the largest dense (non-MoE) option, giving downstream teams a coherent 2B → 4B → 8B → 32B dense sweep for ablation studies and deployment matching. The 32B-dense beats 235B-MoE-on-OSWorld claim is the more interesting datapoint: if true, it argues against the “MoE always dominates at parity” assumption embedded in much of the open-release landscape (Qwen3-VL — Multimodal LLM series (Dense + MoE, Instruct + Thinking), Qwen3-VL-30B-A3B-Instruct) — though as a marketing tweet without numbers it should be treated as a hypothesis to verify rather than a finding. Fits the Open foundation-model releases pattern (coordinated Instruct+Thinking+FP8 multi-variant drop, day-0 HF/ModelScope/API availability) the wiki has been tracking.

See also

• Qwen3-VL compact dense — 4B and 8B (Instruct + Thinking) — the immediately prior dense announcement (4B and 8B), same Instruct+Thinking matrix; this tweet adds the 2B and 32B endpoints.
• Qwen3-VL — Multimodal LLM series (Dense + MoE, Instruct + Thinking) — the umbrella Qwen3-VL repo covering all dense + MoE variants.
• Qwen3-VL-30B-A3B-Instruct — the 30B-A3B MoE Instruct variant the 32B-dense claims to match/beat.
• Qwen3-VL-30B-A3B-Thinking — the 30B-A3B MoE Thinking variant; same pairing pattern.
• Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking — Qwen3-VL-Embedding/Reranker tech report; downstream specialization of this base series.
• Open foundation-model releases — broader pattern of coordinated multi-variant open releases.