WaveCut/Qwen3.6-35B-A3B-REAM-160-ru-agent-GGUF-MTPLess

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

MTPLess GGUF variants of `WaveCut/Qwen3.6-35B-A3B-REAM-160-ru-agent`.

These files are derived from the MTP-preserved GGUF release by rewriting each GGUF without Qwen3.6 MTP/NextN tensors. Existing quantized trunk tensors are preserved; the repack step does not dequantize or re-quantize them.

Use this repo when model size and VRAM headroom matter more than MTP/speculative decoding. The MTP-preserved GGUFs remain available in `WaveCut/Qwen3.6-35B-A3B-REAM-160-ru-agent-GGUF`.

Expected Files

Files are uploaded only after an upstream llama-cli smoke test passes:

• —Standard: Q1_0-MTPLess, IQ1_S-MTPLess, IQ1_M-MTPLess, Q2_K-MTPLess, Q2_K_S-MTPLess, IQ2_XXS-MTPLess, IQ2_XS-MTPLess, IQ2_S-MTPLess, IQ2_M-MTPLess, Q3_K_S-MTPLess, Q3_K_M-MTPLess, Q3_K_L-MTPLess, IQ3_XXS-MTPLess, IQ3_XS-MTPLess, IQ3_S-MTPLess, IQ3_M-MTPLess, Q4_0-MTPLess, Q4_1-MTPLess, Q4_K_S-MTPLess, Q4_K_M-MTPLess, IQ4_NL-MTPLess, IQ4_XS-MTPLess, MXFP4_MOE-MTPLess.
• —Mixed precision: WC-Q3_K_XL-MTPLess, WC-Q4_K_XL-MTPLess.

Per-file size and SHA256 are recorded in manifests/gguf-*.json.

Repack Method

The repack script removes:

• —blk.40.*
• —mtp.*
• —blk.*.nextn.*
• —qwen35moe.nextn_predict_layers
• —qwen35moe.block_count=41, rewritten to 40

It also adds the mtpless tag to general.tags when tags are present.

The resulting files are ordinary GGUF model files for normal next-token generation. They do not contain an embedded MTP draft head.

Validation

Every published file must pass:

• —llama-gguf-dump manifest generation;
• —llama-cli Russian Cyrillic smoke;
• —llama-cli XML tool-call smoke;
• —llama-cli coding-agent smoke.

See RECIPE.md, manifests/, stats/, and logs/ for the full audit trail.