Skip to page content
SAVRN
Search Plan a project

Open-weight model

umba-5.6.600

by Regis Tomkiel Tomkiel/umba-5.6.600

umba-5.6.600 is an open-weight model from Regis Tomkiel, released under Apache License 2.0. Its published files total 3.2 KB.

umba-5.6.600 is a state-of-the-art agentic software engineering model fine-tuned on top of Kwaipilot/KAT-Coder-V2.5-Dev (Sparse MoE 35B total / 3B active parameters). Quantized using LocalAI APEX (Adaptive Precision for EXpert Models)

Parameters—
Context—
Weights3.2 KB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Model Card

By Regis Tomkiel, published under apache-2.0, revision e96d1ba97c32.

umba-5.6.600 is a state-of-the-art agentic software engineering model fine-tuned on top of Kwaipilot/KAT-Coder-V2.5-Dev (Sparse MoE 35B total / 3B active parameters). Quantized using LocalAI APEX (Adaptive Precision for EXpert Models)

Read Regis Tomkiel's full model card

umba-5.6.600 (KAT-Coder-V2.5-Dev 35B MoE)

umba-5.6.600 is a state-of-the-art agentic software engineering model fine-tuned on top of Kwaipilot/KAT-Coder-V2.5-Dev (Sparse MoE 35B total / 3B active parameters).

Training Architecture

  • Base Model: Kwaipilot/KAT-Coder-V2.5-Dev (Top Open-Source model on SWE-bench Verified: 69.4%)
  • Phase 1 (SFT): 600 steps trained on ~49,500 curated trajectories (~186k canonical tool calls, strict prompt assumptions, Fable Method, <think> deep CoT).
  • Phase 2 (DPO): 100 steps of contrastive direct preference optimization to eradicate hallucinated tool calls and language attractors.

APEX LocalAI Quantizations Included

Quantized using LocalAI APEX (Adaptive Precision for EXpert Models):

Filename APEX Flavor Target Profile Best Suited For
umba-5.6.600-tripod.gguf Quality Q6_K / Q5_K adaptive Maximum accuracy and coding precision (Production / GPUs with 24GB+ VRAM)
umba-5.6.600-dirt.gguf Compact-IQ4 Q4_K edge / IQ4_XS core Superior balance between memory usage and code syntax retention
umba-5.6.600-facelift.gguf Apex Compact Q4_K edge / Q3_K core High throughput, minimal memory footprint (16GB GPUs / Apple Silicon)

Usage with llama-server

llama-server -m umba-5.6.600-tripod.gguf -c 32768 -ngl 99 --port 8080

Identity and Version

Repository
Tomkiel/umba-5.6.600
Publisher
Regis Tomkiel
Task
Not stated by the source
Modality
Other
Library
Not stated by the source
Parameters
Not stated by the source
Languages
en, pt
Revision
e96d1ba97c32768985d8a42ea5d787f0fefd7eff
First published
2026-10-08
Last updated
2026-10-09

Files and Weights

2 files, 3.2 KB in total.

Documentation1 file · 1.6 KB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
README.mdDocumentation1.6 KB —
.gitattributesRepository1.6 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download from Regis Tomkiel

Released by Regis Tomkiel through its official repository on Hugging Face. Read the license.

Built From

  • Derived from Kwaipilot/KAT-Coder-V2.5-Dev

Questions About umba-5.6.600

Can I use umba-5.6.600 commercially?

Yes. umba-5.6.600 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.