Ant Group Releases Ling-3.0-flash-VL Vision Model
Ant Group's InclusionAI lab has launched Ling-3.0-flash-VL, an open-source 124-billion-parameter model that brings multimodal vision capabilities to its sparse reasoning architecture.
InclusionAI, a lab under Ant Group, has released Ling-3.0-flash-VL under a permissive MIT license. This mixture-of-experts (MoE) reasoning model features 124 billion total parameters but activates only 5.5 billion active parameters per token, routing inputs through eight of its 512 specialized experts and one shared expert. It supports a 256K-token context window and accepts text, images, and short video clips to produce text-based reasoning traces. The attention mechanism combines 35 Kimi Delta Attention layers with seven Gated Multi-head Latent Attention layers in a 5:1 ratio to optimize memory usage during long-context processing.
On the Artificial Analysis Intelligence Index v4.3, the model scored 25, placing it on the intelligence-versus-active-parameters Pareto frontier compared to a median score of 8 for similarly sized open-weight models. However, its performance varies across tasks. In AA-Omniscience, it demonstrated a conservative profile with 14% accuracy and a low 22% hallucination rate, whereas Inkling Small achieved 33% accuracy but suffered a 63% hallucination rate. Ling-3.0-flash-VL struggled on agentic benchmarks, scoring 16% on AutomationBench-AA and 0% on Terminal-Bench v4.0. It is also highly verbose, averaging roughly 50,000 output tokens per Intelligence Index task—about 1.7 times more than Inkling Small's 30,000 tokens.
For practitioners, deploying the model requires significant hardware. The BF16 weights require 255GB of memory, while the FP8 weights require 128GB, with FP4 and INT4 quantizations planned. InclusionAI supports serving via SGLang and a custom vLLM fork, recommending a temperature of 0.6, top_p of 0.95, and top_k of 20. In hosting tests, the model achieved 142.9 output tokens per second with a 2.24-second time to first token. Free hosted access is temporarily available on Novita AI with limits of 262K input and 32K output tokens. While official pricing for the vision model is not yet public, its text-only predecessor cost $0.075 per million input tokens and $0.22 per million output tokens, previously running an index evaluation for $73 using 240 million output tokens.
This is our own summary of reporting by AlphaSignal
