開放模型與評測
Ling-3.0-flash-VL Adds Low-Precision Weights, but Its Independent Evaluation Is Not Directly Comparable to the Official Score of 42
InclusionAI’s 124B multimodal MoE now offers FP8, FP4, and INT4 weights as well as a working SGLang recipe, bringing image, video, and tool calling capabilities into a single model. The latest independent test by Artificial Analysis produced a score of 25, but it used a different evaluation version from the model card’s claimed score of 42, so deployers should not treat the two as being on the same scale.
Ling-3.0-flash-VL’s release became more complete over the weekend. In addition to the MIT-licensed BF16 weights, InclusionAI has published FP8, FP4, and INT4 versions, while the model card has been updated with a validated SGLang launch recipe. The model has 124B parameters in total, but its sparse MoE activates only about 5.5B per token. On the vision side, a ViT extracts image and video features, which are then projected into language representations through a two-layer MLP. The 42-layer backbone interleaves KDA and Gated MLA at a 5:1 ratio, while VideoRoPE encodes both spatial positions and temporal order. The stated maximum context length is 256K tokens.
In practice, deployment still does not have the hardware cost of a “5.5B model.” The official BF16 recipe recommends four GPUs in roughly the 141GB class, while 80GB H100/H800 GPUs require scaling to eight cards. SGLang uses YaRN to extend the original 131K positional range to 262,144 and automatically selects the Ling3 reasoning and tool-call parsers. The vLLM path currently requires installing InclusionAI’s dedicated fork and enabling `trust-remote-code`; upgrades, supply-chain review, and compatibility with upstream vLLM should therefore all be included in deployment checks.
The evaluation results show a discrepancy worth tracking. The official model card claims a score of 42 on Artificial Analysis Intelligence Index v4.1.1, whereas Artificial Analysis currently lists an independently measured score of 25 using v4.3. It also reports throughput of about 140.7 tokens per second, time to first token of 1.78 seconds, and as many as 160 million evaluation output tokens. The two scores use different versions, question sets, and execution conditions, so they cannot be interpreted directly as evidence that the model has regressed or that the official figure has been disproven. The speed measurement also comes from the InclusionAI API rather than a self-hosted quantized checkpoint. Engineering teams should next wait for a retest using the same evaluation version and compare the quality and actual memory usage of BF16, FP8, FP4, and INT4 across video understanding, GUI operation, tool-call formatting, and long-context workloads.