N Neurarch Try it See it Changes Bot Pricing API Open app

Changes / Feeds / Bahushruth/Qwen3.6-35B-A3B-abliterated-v4

Bahushruth/Qwen3.6-35B-A3B-abliterated-v4

2 changes to a tracked key since 2026-06-11, newest first. Head as last read: 1c28f5014d7e9c20170af40e5aa028679fb30bfb.

Declares Qwen/Qwen3.6-35B-A3B as its base but has a different set of tensors: 693 align (121 of them changed), 0 are new and 352 of the base's are gone. Compared tensor by tensor from sha256 fingerprints: 1c28f5014d7e against Qwen/Qwen3.6-35B-A3B at 995ad96eacd9. A hash is equal or not; it says nothing about what an edit does.

DateAdvisoryWhat changed, and what it does
2026-07-03x_NXMC-2026-0087 breakingarchitectures ["Qwen3_5MoeForConditionalGeneration"] to ["Qwen3_5MoeForCausalLM"]; a different class is instantiated from the same id (2026-07-03).
1424051343e8 -> 1c28f5014d7e Open in app
2026-07-03x_NXMC-2026-0086 breakingarchitectures ["Qwen3_5MoeForCausalLM"] to ["Qwen3_5MoeForConditionalGeneration"]; a different class is instantiated from the same id (2026-07-03).
8f5458e5254a -> 1424051343e8 Open in app

Does vLLM send the reference prompt?

vLLM hands this template message content as text parts, and it renders the same prompt as the reference on the three fixed conversations. Template at 1c28f5014d7e, read 2026-11-17.

npx neurarch template-check Bahushruth/Qwen3.6-35B-A3B-abliterated-v4 --engine vllm

Rendered with @huggingface/jinja from the template at that commit, the way vLLM hands it messages with --chat-template-content-format auto (the default): system, user and assistant content as text parts when the template reads parts, tool results as strings. That handling was measured end to end on 2026-11-13 through vLLM 0.33.0's own request parsing (parse_chat_messages) and transformers' apply_chat_template; it agreed with this page on all 18 models it could load; repositories that need custom code (InternVL, sarvam) could not be loaded there and are unconfirmed. SGLang, llama.cpp and other engines are not covered. same holds for a system+user turn, a multi-turn chat and a tool call, not every input.

Pin it, and let a bot move the pin

from_pretrained("Bahushruth/Qwen3.6-35B-A3B-abliterated-v4", revision="1c28f5014d7e9c20170af40e5aa028679fb30bfb")

With the Renovate preset, a pull request moves this pin when the head moves, and links back here.

config changes badge

[![config changes](https://img.shields.io/endpoint?url=https%3A%2F%2Fwww.neurarch.com%2Fchanges-feeds%2Fbadge%2FBahushruth%2FQwen3.6-35B-A3B-abliterated-v4.json)](https://www.neurarch.com/changes-feeds/models/Bahushruth/Qwen3.6-35B-A3B-abliterated-v4.html)