Views
No views yet
Blackfrost-AI/LING-3.0-FLASH-ABLITERATED — the abliterated (uncensored) variant of LING 3.0 Flash, converted with stock llama.cpp.| Property | Value |
|---|---|
| Architecture | bailingmoe3 (BailingMoeV3ForCausalLM) |
| Parameters | 124B total / ~5.1B active (MoE, 512 experts, 8 active) |
| Layers | 42 (layer group size 6) |
| Context | 262,144 (hardware-dependent) |
| License | MIT |
| Quantization | Q4_K_M, 4.83 BPW — no imatrix |
| File size | 77.0 GB (77,010,145,120 bytes) |
bailingmoe3 support (commit 6d0549831 or newer — upstream since Aug 2026).1llama-server -m LING-3.0-FLASH-ABLITERATED-Q4_K_M.gguf \
2 --host 0.0.0.0 --port 8080 \
3 -ngl 99 # offload all layers to GPU(s)1curl http://localhost:8080/v1/chat/completions \
2 -H "Content-Type: application/json" \
3 -d '{
4 "messages": [{"role": "user", "content": "Hello!"}],
5 "max_tokens": 512
6 }'1llama-cli -m LING-3.0-FLASH-ABLITERATED-Q4_K_M.gguf -ngl 99 \
2 -p "Your prompt here" -n 512ollama create), and any llama.cpp-compatible client.reasoning_content in the OpenAI-compatible API). Leave enough max_tokens headroom for thinking + answer.blk.*.attn_k_b.weight, ncols=128 not divisible by 256) fell back to q5_0 due to the Q4_K_M block-size constraint — negligible impact.sha256: f47f38cfdac87837220fa34a3ba026b83498d9aa19996b18ba7f312b11be9fa66d0549831 (fact/QA, math word problem, code generation, creative writing).