Qwen3.6-35B-A3B-heretic-dialog-lora API & Inference Endpoint

Qwen3.6-35B-A3B Heretic dialog LoRA

Response-stage adapter from checkpoint-17 of the 2026-10-12 HOT LLM training run. This adapter was saved before the later routing and extraction training stage.

Base model revision: dbfd9eb0cdc7c33fc970b06429f6e043b6851190
LoRA rank: 16; alpha: 32
Response training: 518 examples, one epoch, 17 optimizer updates
Export: 256-expert MoE layout conversion and trained context-token input embedding LoRA
Intended serving alias: dialog
Validation: the Moscow bars/no-results prompt produced a concise Russian response and stopped normally; this is a focused check, not a full response benchmark.

For vLLM, load the tokenizer from this adapter and use the provided chat template. The LoRA module must specify is_3d_lora_weight=true. The included qwen_embedding_lora.py enables the required Qwen input-embedding LoRA compatibility in vLLM 0.26.0. Training data and optimizer states are not included.

Qwen3.6-35B-A3B Heretic dialog LoRA

Response-stage adapter from checkpoint-17 of the 2026-10-12 HOT LLM training run. This adapter was saved before the later routing and extraction training stage.

Base model revision: dbfd9eb0cdc7c33fc970b06429f6e043b6851190
LoRA rank: 16; alpha: 32
Response training: 518 examples, one epoch, 17 optimizer updates
Export: 256-expert MoE layout conversion and trained context-token input embedding LoRA
Intended serving alias: dialog
Validation: the Moscow bars/no-results prompt produced a concise Russian response and stopped normally; this is a focused check, not a full response benchmark.

Qwen3.6-35B-A3B-heretic-dialog-lora

README

Qwen3.6-35B-A3B Heretic dialog LoRA

Explore FriendliAI today

README

Qwen3.6-35B-A3B Heretic dialog LoRA