bloomer010/Ling-3.0-flash-GGUF

Ling-3.0-flash GGUF

GGUF conversions of inclusionAI/Ling-3.0-flash (124B total / 5.1B active, hybrid KDA + gated MLA, 512-expert MoE), converted directly from the released BF16 safetensors.

These are the reference conversions for the bailingmoe3 architecture submitted to llama.cpp in PR #26608. Every file bundles the MTP (NextN) block and Ling 3.0's trained per-layer SwiGLU clamp metadata, and no separate drafter file, nor fork required once merged.

Compatibility

bailingmoe3 support is pending merge upstream in PR #26608. Once merged, any llama.cpp build from that point onward loads these files directly:

llama-server -hf bloomer010/Ling-3.0-flash-GGUF:Q4_K_S

Until then, use the fork the PR was developed on: https://github.com/aetherbird/llama.cpp/tree/bailingmoe3-support

Pick a file

Generally... Larger files = more precision. Smaller files = More compression = More slop and misbehavin'.

Weights and context share your memory, so be sure leave headroom.

your memory	file	size
192 GB+	UD-Q8_K_XL	177 GB
128 GB	Q8_0	136 GB
96 GB	UD-Q6_K_XL	116 GB
80 GB (A100/H100)	Q5_K_M	92 GB
64 GB	Q4_K_M	78 GB
56 GB	Q4_K_S / MXFP4_MOE¹	74 / 70 GB
48 GB	Q3_K_M	63 GB
32 GB	UD-Q2_K_XL / IQ2_M	43 / 42 GB
24 GB	IQ1_M (with expert offload, see below)	30 GB

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 Q4_K_S on older hardware.

With less VRAM than the file size, keep the experts on CPU and the rest on GPU, e.g.:

llama-server -hf bloomer010/Ling-3.0-flash-GGUF:IQ1_M \
  -ngl 99 -ot "ffn_.*_exps\.weight=CPU" -c 32768

Usage

Recommended sampling from the source model card: temperature 0.6, top_p 0.95, top_k 20. Thinking mode is on by default; disable per request with "chat_template_kwargs": {"enable_thinking": false}.

./build/bin/llama-server \
  -m Ling-3.0-flash-Q4_K_S.gguf \
  -c 131072 \
  -ngl auto \
  --flash-attn auto \
  --temp 0.6 --top-p 0.95 --top-k 20 \
  --jinja

MTP speculative decoding

Every quant bundles the MTP/NextN block. Enable it with --spec-type draft-mtp:

./build/bin/llama-server \
  -m Ling-3.0-flash-Q8_0.gguf \
  -c 131072 \
  -ngl auto \
  --flash-attn auto \
  --temp 0.6 --top-p 0.95 --top-k 20 \
  --jinja \
  --spec-type draft-mtp

During ordinary inference, llama.cpp skips the MTP tensors and may report them as unused. With --spec-type draft-mtp, the same GGUF is opened as an MTP draft model and block 42 is loaded and executed. No separate drafter file is required.

MoE placement can be adjusted for available VRAM with -ncmoe N. Draft-model placement can be controlled separately with -ncmoed N and -ngld N.

Supports up to 256K context.

Conversion and Quantization

Taken directly from the released inclusionAI/Ling-3.0-flash BF16 safetensors.

Conversion-specific tensor transformations include:

• A_log stored as exp(A_log)
• MLA kv_b_proj split into separate K and V tensors, with the K tensor transposed
• KDA convolution weights reshaped for llama.cpp
• Per-expert tensors stacked into GGUF expert tensors
• KDA and MLA g_proj tensors mapped separately

Norms, routing tensors, expert routing bias, KDA state scalars, dt_bias, and convolution weights remain F32.

Importance Matrix

Importance matrix generated from the Q8_0 model:

• wiki.train.raw
• 100 chunks
• 512 tokens per chunk
• 51,200 calibration tokens total
• 573 matrix entries

Quants

MXFP4_MOE:

• Quantized using llama.cpp's MXFP4_MOE quantization type (4.25 bpw)

Q8_0:

• 8.51 BPW
• 126.3 GiB
• Includes MTP block

UD-Q2_K_XL:

• Model-specific Unsloth-style mixed tensor recipe
• Main expert gate/up tensors: IQ2_XS
• Main expert down tensors: IQ3_XXS
• Final target layer experts: IQ3_XXS and IQ4_XS
• Attention, shared experts, and KDA projections retained at higher precision
• MTP experts: Q3_K and Q4_K

IQ1_S:

• Expected size: approximately 24.9 GiB
• Preserves MTP functionality

Notes

The GGUF contains 43 blocks:

• 42 target-model layers
• 35 KDA layers
• 7 gated MLA layers at zero-based indices 5, 11, 17, 23, 29, 35, and 41
• One MTP/NextN block at index 42

The first two target layers use dense FFNs. The remaining target layers use 512 routed experts with top-8 selection plus one shared expert. Routing uses sigmoid scoring, expert bias, eight expert groups, and four selected groups.

The KDA safe gate is implemented as:

lower_bound * sigmoid(exp(A_log) * (f_proj(x) + dt_bias))

The lower bound is -5.0. The GGUF stores the positive exp(A_log) value, while the sign is supplied by the negative lower bound.

Validation Completed

• BF16 architecture load and tensor round-trip
• CPU and CUDA execution on a reduced-size BailingMoE3 fixture
• Target next-token parity against the released Hugging Face implementation before the missing trained clamps were identified
• Nonzero SwiGLU clamp execution and GGUF round-trip on the reduced-size BailingMoE3 fixture
• First three recursive MTP proposals matched the Hugging Face implementation
• Full MXFP4_MOE target and MTP graph smoke test
• Q8_0 conversion completed successfully with all 938 tensors

Build

# until PR #26608 merges:
git clone --branch bailingmoe3-support https://github.com/aetherbird/llama.cpp.git
# after merge: git clone https://github.com/ggml-org/llama.cpp.git

cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j --target llama-cli llama-server

Upstream PR: https://github.com/ggml-org/llama.cpp/pull/26608