Ling-3.0-tiny-GGUF

Ling-3.0-tiny GGUF

GGUF conversions of inclusionAI/Ling-3.0-tiny, converted directly from the released BF16 safetensors.

🦙🚨 llama.cpp 🦙🚨

Consistent agentic use (tool calling, reasoning split) currently requires two unmerged llama.cpp PRs:

• Dedicated Ling parser: #28682 (✅ Merged as of 9/19)
• Invalid UTF-8 handling in the PEG parser: #29161 (✅ Merged as of 9/20)

Without both, tool calls inside an unclosed think block are dropped and some turns fail with a 500. Will update this note as they merge.

The model does occasionally terminate its response, mid-think, without any sort of closing.
This is inherent in the weights, even at full precision.

To run with llama-server:

llama-server -hf bloomer010/Ling-3.0-tiny-GGUF:Q4_K_M

Quant Sizing

For tiny models, precision is especially crucial.

Generally... <BR> Larger files = More precision.<BR> Smaller files = More compression = More slop and misbehavin'.

Use UD-Q8_K_XL for near-full precision performance.

¹ MXFP4_MOE runs its native path on MXFP4-capable GPUs (Blackwell RTX 50-series, GB10/DGX Spark). Elsewhere it falls back to a slower dequant path — prefer a K-quant on older hardware.

Importance Matrix

The IQ-quant rungs (IQ1_S through IQ4_XS) were generated with a model-specific importance matrix:

• Wikitext-2 raw training text
• 100 chunks
• 512 tokens per chunk
• 51,200 calibration tokens total
• 332 matrix entries

XL Quantization Recipes

UD-Q8_K_XL uses Q8_0 for the main expert gate and up tensors. Token embeddings, expert down projections, attention and Q-LoRA projections, and KDA projections remain BF16.

UD-Q6_K_XL uses Q6_K for the main expert gate and up tensors. Token embeddings, output weights, expert down projections, attention and Q-LoRA projections, and KDA projections use Q8_0. It was generated with the importance matrix described above.

Architecture

• 7.9B total parameters and 1.3B active parameters per token
• 24 layers: 18 KDA layers and 6 MLA layers
• 128 routed experts, 8 active per token, plus 1 shared expert
• Q-LoRA rank 256 and KV-LoRA rank 512
• 131,072-token context in the released configuration
• No bundled MTP block for this model (num_nextn_predict_layers: 0)

Validation

• BF16 conversion completed with 526 tensors, including all 18 Q-LoRA tensors
• CPU and CUDA architecture tests passed
• BF16, Q8_0, Q6_K, Q4_K_M, and MXFP4_MOE loaded and generated tokens with CUDA
• Q1_0, IQ2_M, Q3_K_M, Q5_K_S, and Q5_K_M passed CPU-only prompt processing and token generation tests
• UD-Q6_K_XL and UD-Q8_K_XL passed CPU-only prompt processing and token generation tests
• IQ1_S, IQ1_M, IQ2_S, IQ2_XS, IQ2_XXS, IQ3_XXS, IQ3_S, IQ4_XS, Q2_K, Q3_K_S, Q4_K_S, Q4_0, and Q5_0 passed load and generation tests
• CUDA testing used an RTX 4070 and RTX 3060

Build

git clone https://github.com/ggml-org/llama.cpp.git   # bailingmoe3 merged 2026-08-17
# pre-merge builds:
# git clone --branch bailingmoe3-support https://github.com/aetherbird/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j --target llama-cli llama-server

Usage

./build/bin/llama-server \
  -m Ling-3.0-tiny-Q4_K_M.gguf \
  -c 131072 \
  -ngl auto \
  --flash-attn auto \
  --temp 1.0 --top-p 0.95 --top-k 20 \
  --jinja

Thinking is enabled by default; disable per request with "chat_template_kwargs": {"enable_thinking": false}. Recommended sampling parameters from the source model card are temperature=1.0, top_p=0.95, and top_k=20.