Ant Group's inclusionAI Open-Sources Ling-3.0-flash-VL, a 124B Sparse MoE Vision-Language Model
On September 8-9, Ant Group's inclusionAI open-sourced Ling-3.0-flash-VL on Hugging Face and ModelScope: a sparse MoE vision-language model with 124B total parameters and only 5.5B activated per token, natively supporting image and video input under the MIT license. BF16 and FP8 weights are available now, with FP4 and INT4 quantized versions to follow. As Ant's first natively multimodal model, it balances capability and inference cost and is worth watching.
Confirmed
- The model natively handles image and video input, follows an "understand → reason → act → verify" pipeline, and supports tool calling
- It can interpret images, videos, documents and UI, reasoning, searching and verifying along visual cues; weights can be downloaded and plugged into custom multimodal workflows
- Full model configuration and chat template (including an audio field) are live on Hugging Face
- Per @智东西, the model is Ant's first natively multimodal LLM, activating 5.5B parameters per inference, and outscored GPT-5.4 on visual programming benchmarks
Unconfirmed
- Context length is disputed: @jacek2023 and @AdinaYakup report 1M context, while @智东西 reports 256K; the official technical docs should be authoritative
Why it matters
- As Ant's first natively multimodal model, it uses sparse MoE to balance large capacity with low inference cost; the MIT license and multiple quantized weight tiers lower the adoption barrier, adding new competitive pressure to the open-source multimodal ecosystem
2026-09-08 ~ 2026-09-10 · 9 related posts
Primary sources
- inclusionAI open-sources Ling-3.0-flash-VL: 124B MoE with 5.5B active and 1M context — jacek2023 · 2026-09-08
- [source] inclusionAI open-sources vision model Ling-3.0-flash-VL with BF16 and FP8 weights — FellMentKE · 2026-09-09
- Ant's Ling-3.0-flash-VL Open-Sourced: A Vision Model That Uses Tools — FellMentKE · 2026-09-09
- [source] Ant's open-source Ling-3.0-flash-VL edges out GPT-5.4 on Image-to-WebDevArena — 智东西 · 2026-09-09
- [source] inclusionAI open-sources Ling-3.0-flash-VL: 124B total/5.5B active, 1M context, MIT licensed — AdinaYakup · 2026-09-09
- Ling-3.0-flash-VL weights and config now live on Hugging Face — AdinaYakup · 2026-09-09
- Ling-3.0-flash-VL open-sourced: BF16 and FP8 weights available now — FellMentKE · 2026-09-10
- Ling-3.0-flash-VL open weights released in BF16 and FP8 — FellMentKE · 2026-09-10
1 near-duplicate retellings: Aiden_Tech_Ai