leandronunes/qwen3.8-4b-a1-omnimerge-m8

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

leandronunes/qwen3.8-4b-a1-omnimerge-m8 is a 4.5 billion parameter multimodal language model based on Qwen3.5-4B, created by leandronunes. This model is a merge of Qwen3.8-4B-Distill and Agents-A1-4B using the OmniMerge v2 method, designed to combine logical deduction and agentic capabilities. It features a hybrid architecture with Gated DeltaNet layers and a vision tower, supporting a 256K token context window. This model is optimized for tasks requiring both strong reasoning and causal judgment.

Loading preview...

Model Overview

leandronunes/qwen3.8-4b-a1-omnimerge-m8 is a 4.5 billion parameter multimodal language model built upon the Qwen3.5-4B base. It was created by leandronunes through a sophisticated merge process using OmniMerge v2 (incorporating OBIM-lite, DAREx, and EMR) to combine the strengths of two distinct models:

  • empero-ai/Qwen3.8-4B-Distill: Contributes strong logical deduction capabilities (78% on BBH Logical Deduction).
  • InternScience/Agents-A1-4B: Enhances agentic and causal judgment abilities (38% on BBH Causal Judgement).

Key Features & Technical Notes

  • Hybrid Architecture: Integrates 3 linear_attention (Gated DeltaNet) layers followed by 1 full_attention, repeated 8 times.
  • Multimodal (VLM): Includes a dedicated 24-block vision tower in addition to the language model.
  • Extended Context Window: Supports a substantial context length of 262,144 tokens (256K).
  • "Thinking Mode": Offers an optional extended reasoning mode, which can be disabled via enable_thinking=False in the apply_chat_template function.
  • Tokenizer: Utilizes the tokenizer from Qwen3.8-4B-Distill with corrected added_tokens_decoder.

Use Cases

This model is particularly well-suited for applications requiring a blend of:

  • Logical Deduction: Tasks that benefit from strong reasoning and inferential capabilities.
  • Agentic & Causal Judgment: Scenarios where understanding cause-and-effect and agent-like behavior is crucial.
  • Multimodal Understanding: Applications that can leverage its integrated vision capabilities.

Finetunes
1 models