Qwen3.6-35B-A3B-REAM-160-ru-agent

Build

• Base: Qwen/Qwen3.6-35B-A3B
• REAM commit: 84a3030716a0059589e9d10e2ea049e32b76cfa6
• Transformers: 5.12.1
• Target main MoE experts: 160
• Calibration mix: code+math+agent+ru
• Mix ratio: 0.20,0.10,0.35,0.35
• Batch/sequence/seed: 3072, 512, 42
• REAM merge: grouping=ream, saliency=reap, merging=logits+weights, group_size=32

MTP

The main language MoE layers were REAM-merged to 160 experts. Qwen3.6 MTP tensors are packed differently from the upstream REAM MTP helper assumptions, so packed MTP tensors are reduced with a saved REAM group map rather than upstream --mtp_safe_tensors.

Structural audit:

• Main language MoE layers: 160 experts
• MTP MoE: 160 experts
• Final model.safetensors.index.json includes mtp.* keys
• vLLM qwen3_next_mtp drafter load must pass before the checkpoint is accepted

Calibration Focus

The recipe favors:

• agentic/tool-use behavior via zake7749/Qwen-3.6-plus-agent-tool-calling-trajectory
• Cyrillic/Russian chat retention via ZeroAgency/ru-big-russian-dataset
• code and math anchors from the original REAM calibration paths

Agent calibration had 2477 unique accepted samples and was deterministically repeated to fill the 3072 batch. RU calibration had 3072 unique accepted samples.

Verification

Artifacts in this repo include:

• stats/structural_audit.json
• stats/smoke_outputs.json
• stats/agent_calibration_stats.json
• stats/ru_calibration_stats.json
• logs/main_merge.log
• manifests/hf_final_audit.json

Smoke prompts covered Russian Cyrillic chat, coding-agent debugging, and XML-style tool calling.