Model library / inclusionAI
Ling-3.0-tiny is a 7.9 billion-parameter hybrid MoE language model with 1.3 billion activated parameters per token.
Parameters in total
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
Ling-3.0-tiny is a lightweight hybrid-reasoning model released by inclusionAI. It combines alternating KDA and MLA layers with a sparse MoE feed-forward network of 128 experts, activating eight routed experts and one shared expert per token. The publisher positions it for fast responses, multi-step reasoning, and agentic tasks on local or edge hardware.
The model contains 7.9 billion total parameters but only 1.3 billion are active per token, reducing compute load. It supports a 131,072-token context window and is distributed under the MIT licence. Weights are provided in bfloat16 format, with optional FP8 and INT4 variants for diverse deployment environments.
Specification
| Repository | inclusionAI/Ling-3.0-tiny |
| Publisher | inclusionAI |
| Published | 10 August 2026 |
| Architecture | BailingMoeV3ForCausalLM |
| Model type | bailing_hybrid |
| Parameters | 7,893,392,800 |
| Layers | 24 |
| Hidden size | 1,536 |
| Attention heads | 16, 16 key/value heads |
| Context length | 131,072 tokens |
| Vocabulary | 157,184 tokens |
| Weight format | bfloat16 |
| Licence | mit |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 14.7 GiB | Full precision as released |
| fp8 | 7.4 GiB | 8-bit, near-lossless on most models |
| int4 | 3.7 GiB | 4-bit, smallest footprint |
The cache costs 192 KiB per token at 16-bit, so the full 131,072-token context of one request needs about 24.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses compressed key-value attention, so the real cache is a fraction of the figure above.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 1 | 4.4 GiB | about 24,015 tokens |
| RTX 5090 32GB | 1 | 11.6 GiB | about 63,336 tokens |
| L40S 48GB | 1 | 26.0 GiB | the full 131,072 tokens |
| A100 80GB | 1 | 54.8 GiB | the full 131,072 tokens |
| H100 80GB | 1 | 54.8 GiB | the full 131,072 tokens |
| RTX PRO 6000 96GB | 1 | 69.2 GiB | the full 131,072 tokens |
| H200 141GB | 1 | 109.7 GiB | the full 131,072 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve inclusionAI/Ling-3.0-tiny \ --max-model-len 20480 \ --gpu-memory-utilization 0.90
Questions
The weights take about 15 GiB at the released precision. That fits on one RTX 4090 24GB, which leaves roughly 4 GiB for the key-value cache.
The repository declares mit. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 131,072 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
LLaDA2.2-mini is a 16.3 billion-parameter MoE diffusion language model with 128 k token context, supporting Levenshtein editing.
Ling-3.0-flash-Fin is a 127.5 billion-parameter finance-enhanced language model for long-context financial research and analysis.
Ling-3.0-flash is a 127.5 B parameter hybrid-linear language model from inclusionAI for reasoning and long-context applications.
ZGCM-1 is a 7.4 billion-parameter dense language model designed for mathematical reasoning and tool-assisted search.
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.