SAIL-VL2 Technical Report

TL;DR

SAIL-VL2 is ByteDance Douyin Content’s open vision-language foundation model family at the 2B and 8B parameter scales, positioned as the successor to SAIL-VL and as a Qwen3-backbone alternative to Qwen3-VL. Three contributions: a large-scale data curation pipeline with scoring-and-filtering across captioning / OCR / QA / video, a progressive training framework that starts from a pretrained SAIL-ViT encoder and runs multimodal pretraining followed by a “thinking-fusion” SFT+RL hybrid, and a move from dense LLMs into efficient sparse Mixture-of-Experts at the architecture level. Reports SOTA on MMMU and MathVista among open models in its class, and first place on the OpenCompass leaderboard for officially released open-source models under 4B parameters (SAIL-VL2-2B). Companion HF release includes SAIL-VL2-8B and -2B weights.

Key claims

The model is open-suite at 2B and 8B parameter scales, with both dense and efficient sparse MoE architectural variants covered by the report [Abstract].
A large-scale data curation pipeline with scoring and filtering strategies enhances both quality and distribution across captioning, OCR, QA, and video data, improving training efficiency [Abstract].
The training framework is progressive: it begins with a pre-trained vision encoder (SAIL-ViT), advances through multimodal pre-training, and culminates in a thinking-fusion SFT-RL hybrid post-training stage [Abstract].
Architectural extension beyond dense LLMs to efficient sparse Mixture-of-Experts designs is one of the three explicit contributions [Abstract].
Competitive performance is reported across 106 datasets, with SOTA results on the challenging reasoning benchmarks MMMU and MathVista [Abstract].
SAIL-VL2-2B ranks first on the OpenCompass leaderboard among officially released open-source models under 4B parameters at the time of release [Abstract].

Method

SAIL-VL2 keeps the standard vision-encoder + LLM decoder shape, with SAIL-ViT as the pretrained visual encoder and (per the Slack note that motivated this filing) a Qwen3 LLM backbone — making it the first VLM family to pair an external visual encoder with the actual Qwen3 base, as opposed to Qwen3-VL’s in-house integrated stack. The training recipe runs in three named stages: (1) data curation, in which scoring-and-filtering strategies are applied across captioning, OCR, QA, and video corpora to lift quality and distributional coverage; (2) multimodal pre-training, kicked off from the SAIL-ViT initialization; and (3) a thinking-fusion SFT-RL hybrid post-training paradigm that combines supervised fine-tuning with reinforcement learning on reasoning traces (this is the “thinking” component referenced in the Abstract). The 8B variant is shipped as a sparse Mixture-of-Experts; the report frames the MoE extension as a deliberate architectural contribution rather than a packaging choice, on the grounds that efficient sparse routing is what makes the perception+reasoning capability targets reachable at the 8B activated-parameter scale.

Results

OpenCompass: SAIL-VL2-2B ranks first among officially released open-source models under 4B parameters [Abstract]. The reasoning benchmarks called out by name in the Abstract are MMMU and MathVista, both reported as SOTA in the open-source class at the model’s scale. Aggregate evaluation spans 106 datasets covering image and video benchmarks, ranging from fine-grained perception to complex reasoning [Abstract]. The HF paper page lists five derivative models citing this paper at filing time, including SAIL-VL2-2B and SAIL-VL2-8B checkpoints under BytedanceDouyinContent/.

Why it’s interesting

The most direct comparison is to Ovis2.5 Technical Report — both papers are 2B/9B-class open VLM technical reports anchored on a perception + reasoning agenda, both ship a “thinking mode” component, and both report OpenCompass and MathVista as headline benchmarks. The contrast is on the architecture side: Ovis2.5 keeps a dense LLM, replaces the encoder with NaViT for native-resolution support, and routes the reasoning stage through a DPO + GRPO pipeline with vision modules frozen. SAIL-VL2 instead leans on a sparse MoE LLM at 8B activated, keeps a more conventional vision encoder (SAIL-ViT), and frames its reasoning stage as a single thinking-fusion SFT-RL hybrid rather than a two-step DPO→GRPO. Reading both together gives the wiki a clean dense-encoder-novelty vs sparse-LLM-novelty pair at matched scale.

Pablo’s filing note (“first alternative to Qwen3-VL using the actual Qwen3 backbone”) is the lens that makes this paper relevant beyond the Ovis comparison: Qwen3-VL is widely cited but is not the Qwen3 LLM with a vision encoder bolted on — it has its own training stack. SAIL-VL2 occupies the practical niche of “Qwen3-LLM-as-backbone VLM you can fine-tune from”, which is the form most downstream Luma multimodal experiments will actually want. The MoE direction also lands cleanly in MoE Routing Design alongside ERNIE 4.5’s modality-isolated routing — both papers argue that scaling multimodal capability through MoE is now the default for new releases at this scale, not an experimental detour.