zime-ai/dwight-4b

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Dwight 4B by Zime AI is a 4-billion-parameter language model based on Qwen3-4B-Instruct-2507, specifically designed for analyzing sales call transcripts. It excels at two primary tasks: making one-token decisions with calibrated confidence on specific questions about a call, and generating structured JSON insights about call content. With a 32,768-token context window, it can process up to 90-minute sales calls efficiently, offering rapid analysis and cost-effectiveness for sales teams.

Loading preview...

Dwight 4B: Specialized AI for Sales Call Analysis

Dwight 4B, developed by Zime AI, is a 4-billion-parameter model built on Qwen3-4B-Instruct-2507, engineered to provide precise and rapid analysis of sales call transcripts. It is optimized for two core functions: making quick, confident decisions on specific questions and generating structured insights in JSON format.

Key Capabilities

  • Rapid Decision Making: Answers typed questions about sales calls (yes/no, multiple choice, 0-10 score) with a single token and calibrated confidence, typically in about 0.1 seconds on an H100 GPU. Subsequent questions on the same call are even faster.
  • Structured Insights: Generates detailed JSON outputs for call insights, including headings, specific details, grades, scores, timestamps, and verbatim excerpts from the transcript.
  • High Context Length: Processes up to 32,768 tokens, equivalent to approximately a 90-minute sales call, making it suitable for comprehensive analysis.
  • Cost-Effective at Scale: Designed for efficient self-hosting on a single GPU, offering significant cost savings compared to larger API models for high-volume analysis.
  • Accuracy: Demonstrates strong accuracy in classifying call events like "Next meeting set" and "Pricing sentiment" compared to other models, with high agreement with frontier models at scale.

Good For

  • Sales Teams: Automating the analysis of sales call transcripts to extract key information and insights.
  • Call Classification: Quickly determining specific outcomes or characteristics of sales interactions with high confidence.
  • Structured Data Extraction: Generating machine-readable JSON summaries of call content for further processing or reporting.
  • Cost-Sensitive Operations: Deploying an efficient, specialized model for large-scale sales call analysis without incurring high API costs.
  • English B2B Software Sales Calls: Optimized for this specific domain, though it may be less effective on other call types or languages.

Finetunes
2 models