Ling-3.0-flash REAP176 (46B total / 5.1B active)
[176 of 512 routed experts kept per layer - 65.6% of experts pruned; 176 = 22 expert groups of 8, the group-size-divisible step nearest the 174 target] from inclusionAI/Ling-3.0-flash.
🚨 This is essentially a lobotomized model and it does not work as expected. 🚨
It remains public in case anyone is interested in using it for experimentation or testing. This was the most heavily pruned REAP of the bunch that was performed against Ling 3.0 Flash.
The following REAPs are less degraded and likely worth testing (if your hardware allows):
- https://huggingface.co/bloomer010/Ling-3.0-flash-REAP384-97B-A5B
- https://huggingface.co/bloomer010/Ling-3.0-flash-REAP320-81B-A5B
- https://huggingface.co/bloomer010/Ling-3.0-flash-REAP288-73B-A5B
Method: one-shot REAP (Router-weighted Expert Activation Pruning) - experts scored by router-gate-value × output-L2-norm over calibration data, lowest-scoring deleted. No fine-tuning, no recovery training.
Calibration: 1M tokens, 50/25/25 ultrachat / wikitext / code
This is a very heavily pruned/reaped model and is still mostly untested.
BF16 safetensors. Loads with trust_remote_code=True (custom bailing_hybrid / BailingMoeV3 code).
Deepest published cut of the sweep - expect noticeably more drift than the 288/320/384 siblings
when used out-of-domain. Research artifact; quantized builds live in the sibling -GGUF repo.
2026-08-21: chat_template.jinja updated to support reasoning_effort (low = no thinking, high = thinking, default unchanged).