---

license: apache-2.0

language:

• en

base_model: inclusionAI/Ling-3.0-flash

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

tags:

• llama.cpp
• gguf
• metal
• apple-silicon
• bailingmoe3
• kda
• mla
• moe
• quantized

---

Ling-3.0-flash-GGUF

Canonical bailingmoe3-arch GGUF of inclusionAI/Ling-3.0-flash

— a 124B-parameter hybrid-linear MoE (~5B active; 512 experts/8 used; 42 layers =

35 KDA recurrent + 7 gated-MLA attention; one NextN/MTP head) — converted and

quantized on Apple Silicon, with full provenance for independent reproduction.

| File | Quant | Size | Arch |

|------|-------|------|------|

| Ling-3.0-flash-bailingmoe3-IQ4_XS.gguf | IQ4_XS (~4.3 bpw) | 64 GB | bailingmoe3 (canonical) |

This file loads without shims or overrides in the runtime linked below

(ssm_a is exp'd at conversion, kda.gate_lower_bound is in the file).

Why this file exists

The HuggingFace checkpoint is bf16 (~240 GB) and uses model_type: bailing_hybrid,

which upstream llama.cpp does not support. This repo publishes a canonical

bailingmoe3-arch GGUF so anyone can run the model on Apple Silicon without

re-deriving the (non-trivial) conversion.

Repositories — everything needed to reproduce

| What | Where |

|------|-------|

| Runtime (this GGUF's home) | ljubomirj/Ling-3.0-flash-llama.cpp — standalone llama.cpp fork with BailingMoE3 support (provenance + research summary in README) |

| Living branch the repo was split from | ljubomirj/llama.cpp branch atomic-bailingmoe3 — same code, keeps advancing. Built on top of the original AtomicBot-ai/atomic-llama-cpp-turboquant (Atomic llama.cpp, feat/bailingmoe3) — huge thanks to the Atomic team for the bailingmoe3 architecture support this is derived from |

| Upstream llama.cpp | ggml-org/llama.cpp |

| Upstream BailingMoE3 support (PR #26608) | ggml-org/llama.cpp#26608 — aetherbird's converter (conversion/bailingmoe3.py) used for this file |

| Original checkpoint | inclusionAI/Ling-3.0-flash (bf16, ~240 GB, 24 shards) |

| Early non-canonical GGUFs (legacy bailing-hybrid arch) | aj9o9/Ling-3.0-flash-GGUF — same model, older arch string; loads via the legacy shim in the runtime |

Reproduce this file (from scratch)

Ingredients: the bf16 checkpoint, the PR-26608 converter, llama.cpp's quantizer.

# 1. Download the bf16 checkpoint (~240 GB)
huggingface-cli download inclusionAI/Ling-3.0-flash --local-dir $NV/inclusionAI-Ling-3.0-flash

# 2. Convert to f16 GGUF using the PR-26608 (aetherbird) converter
#    (conversion/bailingmoe3.py in the PR; registers BailingMoeV3ForCausalLM
#     -> model_arch bailingmoe3, exp's A_log, writes kda.safe_gate + lower_bound)
cd <aetherbird-llama.cpp>   # PR #26608
python3 -m convert_hf_to_gguf $NV/inclusionAI-Ling-3.0-flash \
  --outtype f16 --outfile $NV/Ling-3.0-flash-f16.gguf --split-max-size 40G

# 3. Quantize to IQ4_XS
llama-quantize $NV/Ling-3.0-flash-f16-00001-of-00007.gguf \
  $NV/Ling-3.0-flash-bailingmoe3-IQ4_XS.gguf IQ4_XS

Disk discipline: peak ~540 GiB (bf16 240 + f16 255 + IQ4_XS 64); delete the bf16

checkpoint after step 2 to stay within ~700 GiB.

Run it

git clone git@github.com:ljubomirj/Ling-3.0-flash-llama.cpp.git
cd Ling-3.0-flash-llama.cpp
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release \
  -DGGML_METAL=ON -DGGML_METAL_EMBED_LIBRARY=ON -DGGML_ACCELERATE=ON
cmake --build build --config Release --target llama-cli llama-server -j8

./build/bin/llama-cli -m Ling-3.0-flash-bailingmoe3-IQ4_XS.gguf \
  -ngl 99 -p "The capital of France is" -n 16

Measured performance (Apple M2 Max 96 GB, Metal)

llama-bench -m Ling-3.0-flash-bailingmoe3-IQ4_XS.gguf -ngl 99 -r 2:

| Test | t/s |

|------|-----|

| pp2048 @ d0 | 301.5 |

| pp2048 @ d29664 | 154.4 |

| tg128 @ d0 | 35.0 |

| tg128 @ d29664 | 24.3 |

Decode is memory-bandwidth-bound — M2 Max ~400 GB/s ceiling. Roofline check:

~2.96 GB active bytes per token at 8/512 experts (only the selected experts are

read; 97% of the file is MoE weights) → 400 GB/s ÷ 2.96 GB ≈ 135 t/s ceiling,

and we measure 35 t/s — only ~27% of the roofline. There is clearly a lot of

scope for improvement here. Flash attention is enabled by default and mandatory

at depth (fa=0 collapses to ~5 t/s @64K).

Verification / golden logits

The conversion was validated against an independent HF-bf16 golden reference

(golden-hf-bf16.json: 16 steps, full logits, vocab 157,184), captured via a

torch shim of fla-core's KDA kernels (triton is unavailable on macOS). The

first-token distributions match within IQ4_XS noise (p(198): golden 0.296 vs

this file 0.370); later divergence is near-tie argmax flips compounding through

the recurrent state — no graph bug. Details in the runtime repo's README.

License / provenance notes

• Model: inclusionAI/Ling-3.0-flash

(check its license/card for usage terms).

• Conversion tooling: PR-26608 (Apache-2.0, llama.cpp).
• Runtime: ljubomirj/Ling-3.0-flash-llama.cpp,

derived from AtomicBot-ai/atomic-llama-cpp-turboquant

feat/bailingmoe3 — the Atomic team's bailingmoe3 architecture support is

the foundation this work builds on; our thanks. Atomic llama.cpp itself is a

fork of ggml-org/llama.cpp.