InclusionAI's new small Ling model can easily be filed as another parameter-efficiency release. Its real target is more specific: becoming a runtime base for local agents.
On August 11, Ant Group's Ling team open-sourced Ling-3.0-tiny. Chinese coverage from Jiemian reported the basic release facts: 7.9B total parameters, only 1.3B activated parameters per token, and BF16, FP8 and INT4 weights. The Hugging Face model card, ModelScope page and SGLang documentation provide the harder deployment details.
“We are introducing Ling-3.0-tiny, a lightweight hybrid reasoning MoE model”
Why a small model talks about agents
Ling-3.0-tiny uses a 3:1 alternating stack of Kimi Delta Attention and Multi-Head Latent Attention. Its sparse MoE feed-forward layer has 128 routed experts; each token activates 8 routed experts and 1 shared expert.
The engineering point is plain: long context, low memory and tool-using agent tasks must fit together. Small chat models can be cheap, but long coding sessions, tool calls and multi-step decisions often expose planning and context limits. Ling-3.0-tiny is framed around agentic workflows, not single-turn chat.
Local deployment is the release
The model card says the model has been validated on NVIDIA DGX Spark, Apple Silicon MacBook and Mac mini. With FP8, it reaches about 100-105 tokens per second on DGX Spark and 86-90 tokens per second on an M4 Pro MacBook, with roughly 8.34 GiB peak memory at an 8K context length.
Those figures matter because they point to machines developers already own. A newer MacBook or compact workstation can now run a reasoning-capable local agent model. Data can stay on the device, and cost shifts from cloud token billing to hardware and electricity.
The limits are just as important. SGLang's low-latency recipe for 256K YaRN context and NEXTN speculative decoding still mentions a 141GB-class GPU or a single-GPU Blackwell node. The MacBook numbers are FP8, local-hardware and 8K-context figures; they should not be treated as a 256K-task promise.
Benchmarks set the boundary
The model card reports an Artificial Analysis Intelligence Index v4.1.1 score of 25 and an Agentic Index score of 16. In that testing, output speed exceeds 160 tokens per second and a 500-token response takes about 18 seconds end to end, including reasoning time.
That does not make Ling-3.0-tiny a flagship model. Its argument is efficiency per active parameter. Ling-3.0-flash scores higher on the same platform, but it is also much larger. The tiny model is closer to a local execution kernel: run reliably, call tools, and stay resident.
The next checks are practical: whether SGLang and vLLM support stabilizes, whether INT4 performs well on ordinary consumer devices, and whether developers wire it into IDEs, browser automation and internal enterprise tools. The model card shows how to start it; long-running tasks will decide how useful it is.
Sources: Hugging Face model card, ModelScope, Jiemian, Artificial Analysis, CocoLoop, SGLang documentation; verification covers 7.9B total parameters, 1.3B active parameters, BF16/FP8/INT4 weights, KDA/MLA and 128-expert architecture, local DGX Spark/M4 Pro speed, 8K memory scope, Artificial Analysis indexes and SGLang 256K deployment scope.