intellectlabs/Kepler-8B-Instruct-v2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Kepler-8B-Instruct-v2 by intellectlabs is a 7.6 billion parameter instruction-following model, built by merging Qwen2.5-7B-Instruct and DeepSeek-R1-Distill-Qwen-7B. This model combines Qwen2.5's reliable instruction-following with DeepSeek's reasoning capabilities. It is specifically packaged in GGUF format, optimized for local deployment across various hardware with llama.cpp and related tools, offering multiple quantization levels for flexible resource usage.

Loading preview...

Kepler-8B-Instruct-v2: A Merged Instruction-Following Model

Kepler-8B-Instruct-v2 is an 8-billion-parameter instruction-following assistant developed by intellectlabs. It was created by merging Qwen/Qwen2.5-7B-Instruct (85%) with deepseek-ai/DeepSeek-R1-Distill-Qwen-7B (15%) using MergeKit. This blend aims to leverage Qwen2.5-Instruct's robust instruction-following foundation while integrating DeepSeek's reasoning-distilled behavior.

Key Capabilities & Features

  • Hybrid Intelligence: Combines strong instruction-following with enhanced reasoning capabilities from its merged base models.
  • GGUF Format: Packaged in the universal GGUF format, ensuring broad compatibility with llama.cpp, LM Studio, Ollama, koboldcpp, GPT4All, Jan, and text-generation-webui.
  • Flexible Quantization: Offers seven precision levels, from ultra-light Q2_K (3.1 GB) to full-precision F16 (16 GB), allowing users to balance quality and resource efficiency. Q4_K_M is recommended for most users.
  • Native ChatML Support: Utilizes the ChatML prompt format, with automatic application by most compatible tools and a built-in default identity as Kepler by intellectlabs.
  • Local Deployment Optimized: Designed for easy deployment on diverse hardware, from gaming laptops to Raspberry Pi-class boards, without complex Python or CUDA setups.

Limitations

As a merged model, its behavior is a blend of its parent models and may inherit their characteristics or biases. Lower-bit quantizations trade quality for size, potentially leading to more inconsistency. The model has not yet been evaluated on standard benchmarks, so caution is advised for factual or high-stakes applications.

Finetunes
1 models