Ling 3.0 Tiny is a lightweight open-source mixture-of-experts model from ANT Group's Ling team, built for developers who want to run capable AI agents locally on a laptop or workstation instead of a cloud GPU cluster. Released on Hugging Face and ModelScope on August 11, 2026, it targets small, reasoning-capable models that stay resident and call tools on device.
Core Features
- 7.9B total parameters with only 1.3B activated per token, shipped in BF16, FP8 and INT4 weight variants for flexible deployment.
- A 3:1 alternating stack of Kimi Delta Attention and Multi-Head Latent Attention, plus a sparse MoE feed-forward with 128 routed experts (8 routed plus 1 shared per token).
- Fast-response and multi-step reasoning modes via an enable_thinking toggle, validated on NVIDIA DGX Spark, Apple Silicon MacBooks and Mac mini.
- Runs at 86 to 90 tokens per second on an M4 Pro MacBook with about 8.34 GiB peak memory at 8K context, and 100 to 105 tokens/s on DGX Spark under FP8.
Use Cases
- Local coding assistants and IDE agents that need long context without cloud token billing.
- Privacy-sensitive enterprise tools that must keep data on the device.
- Browser automation and internal workflows where a small model stays loaded as an execution kernel.
Pricing
Ling 3.0 Tiny is free and open-weight, with no paid plan: the weights are released on Hugging Face and ModelScope, so cost shifts to your own hardware and electricity. Hosted inference is also available free on Vercel AI Gateway and OpenRouter for quick testing.
Our Take
Best for solo developers and small teams who want a genuinely local, tool-using agent model without renting GPUs. The trade-off is context length: the headline speed and 8.34 GiB memory figures are measured at 8K context, while the full 256K context still wants a 141GB-class GPU. Against peers like Meta's MuseGlimmer and NVIDIA's Nemotron Lightning, Ling 3.0 Tiny wins on everyday practicality rather than raw benchmark scores. For more open models, browse our AI Engine/Model directory.
