sajalregmi4/arkios-1b-chat

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

sajalregmi4/arkios-1b-chat is a 1.04 billion parameter bilingual English–Nepali chat model, instruction-tuned from arkios-1b-base. It features a unique tool-use contract where tool calls are conditional on a declared tool manifest, rather than fixed built-in tools. This model is designed for applications requiring controlled tool interaction and improved English-Nepali translation capabilities.

Loading preview...

Overview

sajalregmi4/arkios-1b-chat is a 1.04 billion parameter bilingual English–Nepali chat model, instruction-tuned from the arkios-1b-base. It distinguishes itself with a sophisticated tool-use contract: the model will only call tools if they are explicitly declared in the system prompt via a manifest. If no tools are declared, it will never attempt a tool call, ensuring predictable behavior.

Key Capabilities

  • Conditional Tool Use: Tools are called only when declared in the system prompt, preventing hallucinated tool calls. This contract is robust for preventing calls without permission.
  • Bilingual Chat: Offers improved English-Nepali translation capabilities compared to its base model, especially for English to Nepali translation (58% accuracy in zero-shot instruction).
  • Arithmetic Tool Reliability: When a calculator tool is declared, the model reliably uses it for arithmetic tasks.

Limitations & Considerations

  • Unit Conversion Weakness: The model is unreliable at calling convert_units even when declared, often answering in prose instead. Its internal unit knowledge is also unreliable.
  • Word Problem Arithmetic: Without a declared calculator tool, the model may attempt prose arithmetic for word problems and make errors, rather than refusing.
  • Nepali Language Nuances: Tool contract reliability is slightly lower in Nepali than in English, and translation quality, while improved, still requires verification for production use.
  • General Small Model Limitations: Inherits limitations such as occasional factual errors in prose answers and a 4096-token context length.