logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Link to Whittle-Next-27B-A3B (3B active) with an honest trade-off note
eval evidence for the arith LoRA
eval evidence for the arith LoRA
arith consolidation LoRA (shared expert, gate/routers frozen): probe-5 14/25 vs base 11/25, unseen transfer (25%-of-480 5/5, division 5/5), reasoning-abort 3/24 = shipped rate
correct the counted comparison: v2.2 rep4 is also 0.141, not 0.237 - the new gate does not improve repetition; add like-for-like distinct counts
card: v2.2.1 fixes the reasoning abort; GGUF tiers not yet rebuilt
archive the v2.2 gate (root weights + these 64 tensors = v2.2)
v2.2.1: replace shared_expert_gate with the CoT-trained gate (reasoning abort 88% -> 12%)
Warn: v2.2 can terminate inside its chain of thought when reasoning is enabled; serve with --reasoning off
Card: v2.2 release notes - stop gate, measured results incl. the counting regression, v2.1 pin SHA
v2.2: corrected shared_expert_gate (0.33M params, --eos-ce objective)
Fix text_config.model_type: qwen3_5_moe -> qwen3_5_moe_text
loop_test: send both thinking-off flags (llama.cpp kwarg + ollama think), verified end to end
Upload loop_test.py with huggingface_hub
Upload model.safetensors.index.json with huggingface_hub
Upload model-00015-of-00015.safetensors with huggingface_hub
Upload model-00014-of-00015.safetensors with huggingface_hub
Upload model-00013-of-00015.safetensors with huggingface_hub
Upload model-00012-of-00015.safetensors with huggingface_hub
Upload model-00011-of-00015.safetensors with huggingface_hub
Upload model-00010-of-00015.safetensors with huggingface_hub
Upload model-00009-of-00015.safetensors with huggingface_hub
Upload model-00008-of-00015.safetensors with huggingface_hub
Upload model-00007-of-00015.safetensors with huggingface_hub
Upload model-00006-of-00015.safetensors with huggingface_hub
Upload model-00005-of-00015.safetensors with huggingface_hub
Upload model-00004-of-00015.safetensors with huggingface_hub
Upload model-00003-of-00015.safetensors with huggingface_hub
Upload model-00002-of-00015.safetensors with huggingface_hub
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Copy files from models/logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Move remaining GGUFs to the -GGUF repo; weights repo is now safetensors + adapters only
Move v2.1 GGUFs to the dedicated -GGUF repo (linked via base_model)
Update README.md
Upload README.md with huggingface_hub
Upload README.md with huggingface_hub