Ling-3.0-tiny NVFP4 GGUF
GGUF quantized version of inclusionAI/Ling-3.0-tiny, a lightweight hybrid reasoning MoE model with 7.9B total parameters and 1.3B active parameters per token, supporting English and Chinese, with a 256K context window, tool-calling, and a built-in thinking/reasoning mode.
Model Overview
Ling-3.0-tiny is a hybrid-linear MoE model built on a native hybrid linear attention architecture, featuring a 3:1 alternating stacking of Kimi Delta Attention (KDA) and Multi-Head Latent Attention (MLA), combined with a sparse MoE FFN comprising 128 routed experts. With 7.9B total parameters and 1.3B activated parameters per token, it balances long-context modeling capability and computational efficiency for local and resource-constrained deployment.
The model uses a Bailing V3-style conversation format with role-based delimiters (<role>SYSTEM</role>, <role>HUMAN</role>, <role>ASSISTANT</role>, <role>OBSERVATION</role>) separated by <|role_end|>. It supports special tokens for thinking/reasoning (controlled via detailed thinking on / detailed thinking off in the system prompt, with reasoning output between think/answer tokens) and tool calling (function blocks with parameter tags).
Model Architecture
| Property | Value |
|---|---|
| Architecture | Hybrid-linear MoE |
| Total Parameters | 7.9B |
| Activated Parameters | 1.3B |
| Attention Stacking | 3 KDA : 1 MLA |
| Routed Experts | 128 |
| Activated Experts | 8 + 1 shared |
| Context Length | 262,144 (256K) |
| Original Precision | bfloat16 |
| Supported Languages | en, zh |
Available GGUF Files
| File | Quantization | Use Case |
|---|---|---|
ling-3.0-tiny-nvfp4.gguf | NVFP4 | Optimized for NVIDIA FP4 hardware acceleration |
Usage
llama.cpp CLI
./llama-cli \
-m ling-3.0-tiny-nvfp4.gguf \
-p "Explain quantum computing in simple terms." \
--temp 1.0 --top-p 0.95 --top-k 20
llama-server (OpenAI-compatible API)
./llama-server \
-m ling-3.0-tiny-nvfp4.gguf \
--host 0.0.0.0 --port 8080
Thinking Mode
Thinking mode is enabled by default. The model outputs its reasoning between think and answer tokens before providing the final answer. To control thinking behavior, include detailed thinking on or detailed thinking off in the system prompt. When using an OpenAI-compatible API, you can also pass "chat_template_kwargs": {"enable_thinking": false} to disable thinking per request.
Tool Calling
The model supports function/tool calling. Tool definitions are injected into the system prompt within <tools></tools> XML tags, and the model responds with function blocks containing parameter elements. Tool results are returned via the observation role channel.
Generation Parameters
Recommended parameters from the original model:
| Parameter | Value |
|---|---|
| Temperature | 1.0 |
| Top-P | 0.95 |
| Top-K | 20 |
Quantization
This GGUF file was created from the BF16 source model using llama-quantize from the llama.cpp project.
Acknowledgements
- Original model: inclusionAI/Ling-3.0-tiny
- Quantization tool: llama.cpp