Skip to content

EthenEthenEthen

LoginPlatformTry Ethen

Open Source Model Profile · 0xA50C1A1

Qwen3-4B-Instruct-2507-SOM-MPOA

Qwen3-4B-Instruct-2507-SOM-MPOA is a 4.02B-parameter Qwen3 text-generation fine-tune from 0xA50C1A1. Its model card documents Heretic-based decensoring of unsloth/Qwen3-4B-Instruct-2507.

Publisher
0xA50C1A1
Task
text-generation
Model type
qwen3
License
apache-2.0
Library
transformers
Publication status
Approved for indexing

Model overview

Qwen3-4B-Instruct-2507-SOM-MPOA is published by 0xA50C1A1 as a text-generation model. The captured configuration identifies Qwen3ForCausalLM, and Safetensors metadata reports 4,022,468,096 parameters. According to the model card, it is a decensored Heretic v1.2.0 derivative of unsloth/Qwen3-4B-Instruct-2507.

Recorded capabilities

Decensored Qwen3-4B variant

According to the model card, this is a decensored version of unsloth/Qwen3-4B-Instruct-2507 produced with Heretic v1.2.0, with reported refusal and KL figures.

Instruct-model capabilities

According to the model card, the base Qwen3-4B-Instruct-2507 description covers instruction following, reasoning, text comprehension, mathematics, science, coding, and tool use.

Documented usage patterns

According to the model card, usage covers Transformers loading, chat-template preparation, sampling settings, math formatting, multiple-choice JSON, and agent tool examples.

Use cases in the source record

  • Conversational text-generation experiments using the documented Qwen3 instruct behavior described in the model card.
  • Sampling, math-format, multiple-choice, and tool-use workflows that follow the card's Transformers, chat-template, and agent examples.

Limitations and unknowns

  • No evaluation results beyond the publisher's refusal and KL figures were extracted.
  • Benchmark, hardware, and inference-performance detail is deferred by the card to external blog, GitHub, and documentation sources.
  • Provider state is historical snapshot data, not independently refreshed current availability.
  • Decensored behavior, capability, and lineage claims come from the publisher model card and were not independently verified.

Source and provenance

Source: 0xA50C1A1/Qwen3-4B-Instruct-2507-SOM-MPOA

Captured: Unknown. Processed: 2026-09-07T19:35:33.334592+00:00.

This is a decensored version of unsloth/Qwen3-4B-Instruct-2507 , made using Heretic v1.2.0 Abliteration parameters Parameter Value direction_index 18.74 attn.o_proj.max_weights.0 0: 1.39 attn.o_proj.max_weights.1 1: 1.50 attn.o_proj.max_weights.2 2: 1.03 attn.o_proj.max_weights.3 3: 0.88 attn.o_proj.max_weights.4 4: 0.97 attn.o_proj.max_weights.5 5: 1.18 attn.o_proj.max_weights.6 6: 0.97 attn.o_proj.max_weights.7 7: 1.49 attn.o_proj.max_weight_position 23.00 attn.o_proj.min_weights.0 0: 0.26 attn.o_proj.min_weights.1 1: 0.32 attn.o_proj.min_weights.2 2: 0.75 attn.o_proj.min_weights.3 3: 0.80 attn.o_proj.min_weights.4 4: 0.13 attn.o_…

F001F002F003F004F005F006F007F009F010F011F012F013F015F016F017F018F019