WaylonJBrown/Maq-1.0-4B-SWE

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Maq-1.0-4B-SWE by WaylonJBrown is a 7.9 billion parameter instruction-tuned model, post-trained from Google's Gemma 4 E4B-IT base. It is specifically optimized for software engineering tasks, coding-agent interactions, and competitive programming challenges. This model excels at generating code, handling multi-turn agentic coding traces, and solving code-reasoning prompts, making it suitable for local and on-device development experiments.

Loading preview...

Overview

Maq-1.0-4B-SWE is a 7.9 billion parameter model developed by WaylonJBrown, built upon the google/gemma-4-E4B-it base. It has been instruction-tuned using LoRA SFT on a curated mix of software engineering, coding-agent, and competitive programming datasets, including NVIDIA Nemotron and Fable/OpenCode agent traces. The LoRA adapters were manually merged into full BF16 weights, resulting in a specialized model for code-centric applications.

Key Capabilities

  • Software Engineering: Optimized for issue-to-plan-to-patch conversations and general SWE-style chat.
  • Agentic Coding: Proficient in tool-using and multi-turn agentic coding traces, including bash commands and file edits.
  • Competitive Programming: Designed to handle competitive programming problems and code-reasoning prompts.
  • Local Deployment: Suitable for on-device and local experimentation due to its size and specialized focus.

Training Details

The model underwent a short post-training phase of 400 steps on a sampled subset of 4,096 rows from a larger dataset mix, using an Intel Arc iGPU. The training utilized a maximum sequence length of 1024 tokens. While no external benchmark suite results are reported, the training loss converged, indicating effective specialization. The model is text-only, with multimodal towers dropped during SFT.

Intended Use Cases

  • Developing coding assistants.
  • Simulating agentic coding workflows.
  • Solving competitive programming challenges.
  • Experimenting with Gemma 4 E4B on local hardware.

It is not intended as a general-purpose chatbot, a safety-tuned model, or a direct replacement for larger 70B+ SWE agents.

Finetunes
1 models