sahilchachra/Tessera-4B-Preview

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 27, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Tessera-4B-Preview by sahilchachra is a 4.5 billion parameter instruction-tuned model based on Qwen/Qwen3.5-4B, specialized through a 4-stage curriculum fine-tune focusing on reasoning, planning, and tool use. It retains general abilities while significantly improving internal proxy scores for reasoning and agentic tasks. This model is designed for research and non-commercial applications requiring enhanced problem-solving and function calling capabilities in a small, locally deployable package.

Loading preview...

Tessera-4B-Preview: A Curriculum-Tuned Qwen3.5-4B Model

Tessera-4B-Preview is a 4.5 billion parameter language model developed by sahilchachra, built upon the Qwen/Qwen3.5-4B (instruct) base. Its core innovation lies in a 4-stage curriculum fine-tuning approach (Reason, Plan, Use Tools, Align), where each stage builds upon the previous one and replays earlier data to prevent catastrophic forgetting. This method, implemented using QLoRA, was achieved with a modest computational budget of approximately $40 on a single NVIDIA B200 GPU.

Key Capabilities & Features

  • Enhanced Reasoning & Planning: The curriculum significantly improves internal proxy scores for reasoning (from 0.375 to 0.785), indicating stronger problem-solving and agentic capabilities.
  • Tool Use Specialization: The model is specifically fine-tuned for tool calling, utilizing an XML-style format for function selection.
  • Thinking Channel: Inherits the <think>…</think> channel from its Qwen base, allowing it to emit internal thought processes for complex problems.
  • Efficient Training: Achieved through QLoRA (4-bit NF4) and paged AdamW, demonstrating cost-effective specialization.
  • Quantized Variants: Available in various MLX and GGUF quantizations (MXFP4, MXFP8, INT4, INT8, OptiQ 5-bpw, Q4_K_M, Q5_K_M, Q8_0) for local and edge inference.

Good For

  • Research and Non-Commercial Use: Due to the inclusion of CC-BY-NC licensed data (xLAM) and its Qwen3.5 derivative nature.
  • Reasoning and Agentic Tasks: Excels in scenarios requiring structured thought processes and multi-step problem-solving.
  • Function Calling: Optimized for selecting and using external tools based on user prompts.
  • Local Deployment: Its 4.5B parameter size and available quantizations make it suitable for running on consumer hardware like Apple Silicon or via llama.cpp.

Finetunes
9 models