jaymanaryan/qwen3-0.6b-reasoning-indic-merge

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The jaymanaryan/qwen3-0.6b-reasoning-indic-merge is a 0.8 billion parameter language model based on the Qwen3-0.6B architecture, created by jaymanaryan. This model is a direct merge of two specialized clusters: one for reasoning and another for Indic languages, achieved using arcee_fusion. It is optimized for reasoning, coding, and Indic/Hinglish language tasks, demonstrating healthy perplexity and coherent outputs across these domains. The model does not include creative writing or medical/general capabilities, focusing instead on its specialized strengths.

Loading preview...

Model Overview

The jaymanaryan/qwen3-0.6b-reasoning-indic-merge is a 0.8 billion parameter language model built upon the Qwen3-0.6B architecture. It is a specialized merge, created by jaymanaryan using arcee_fusion, combining two distinct clusters: a reasoning_stem cluster and an indic cluster. This direct 2-model merge was chosen due to incompatibility issues with a planned third creative-writing ingredient.

Key Capabilities & Characteristics

  • Specialized Focus: Optimized for reasoning, coding, and Indic/Hinglish language tasks. It explicitly does not include creative-writing or medical/general-domain capabilities.
  • Merge Method: Utilizes arcee_fusion for a direct, single-step combination of the two source models.
  • Performance Validation: Achieves a perplexity (PPL) of approximately 28.1 on a fixed neutral-text sample, which is in line with its source clusters. It produced 0/8 degenerate outputs across a battery of reasoning/math, coding, Indic/Hinglish, and general instruction prompts under both greedy and greedy+repetition_penalty decoding.
  • Context Length: Supports a context length of 32768 tokens.
  • Reasoning Traces: Outputs include exposed <think>...</think> reasoning traces by default, inherited from the reasoning_stem ingredient.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

  • Logical Reasoning: Tasks involving problem-solving and structured thought processes.
  • Coding Assistance: Generating or understanding code snippets.
  • Indic Language Processing: Handling content in Indic languages, including Hinglish.
  • Instruction Following: Responding to general instructions within its specialized domains.

Finetunes
1 models