Qwen3.6-35B-A3B Genesis Hermes V7 — Ollama

An Ollama-ready repack of LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-GGUF, available in both APEX and APEX-Compact variants and configured for long-context coding/agent workflows and multimodal use.

Attribution notice

This Ollama package is a repack/distribution configuration. It does not claim authorship of the underlying model, training, uncensoring work, Hermes transfer, Genesis tensor repair, GGUF conversion, or upstream chat template.

Please credit the upstream authors and projects listed below.

Available Variants

This Ollama package is intended to provide both major V7 GGUF variants:

Overview

• Architecture: Qwen3.6 MoE / qwen35moe
• Total parameters: ~35B
• Active parameters: ~3B per forward pass
• Experts: 256
• Active experts: 8 routed + 1 shared
• Native context: 262,144 tokens
• Ollama profile context: 224,000 tokens
• Multimodal: Text + image/video architecture upstream
• Vision: Included through the corresponding mmproj/projector
• Primary use: Coding agents, tool/function calling, long-context development work
• Available GGUF variants: APEX and APEX-Compact
• License: Apache-2.0 upstream

Model Lineage / Credits

This model has multiple upstream contributors. They should not be collapsed into a single “author”.

1. Original foundation model — Qwen Team / Alibaba

Qwen/Qwen3.6-35B-A3B

Original Qwen3.6-35B-A3B foundation/post-trained model by the Qwen (Tongyi Lab / Alibaba Group) team.

Source:

https://huggingface.co/Qwen/Qwen3.6-35B-A3B

Qwen describes the model as a 35B-total / ~3B-active Mixture-of-Experts model with a vision encoder and a native 262,144-token context window.

2. Uncensored base — HauhauCS

HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

HauhauCS produced the uncensored/aggressive derivative used as the base for the Genesis Hermes release.

Source:

https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

3. Hermes agent finetune source — DJLougen

DJLougen/hermes-qwen3.5-35b-a3b-GGUF

LuffyTheFox states that the Genesis Hermes model uses this Hermes-oriented finetune as a source for agent/function-calling behavior, with Hermes-related data transferred into the HauhauCS uncensored base.

Source:

https://huggingface.co/DJLougen/hermes-qwen3.5-35b-a3b-GGUF

4. Genesis Hermes V7 — LuffyTheFox

LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-GGUF

LuffyTheFox created the Genesis Hermes V7 derivative and GGUF release.

According to the upstream model card, the release:

• is based on the HauhauCS uncensored Qwen3.6 model;
• incorporates Hermes-agent-related data from the DJLougen model;
• uses LuffyTheFox’s Genesis tensor repair / calibration process;
• provides GGUF variants including APEX and APEX-Compact;
• provides a corresponding vision mmproj;
• provides/references a Hermes-compatible chat template.

Primary source:

https://huggingface.co/LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-GGUF

Primary upstream author for this V7 derivative / GGUF release: LuffyTheFox.

5. Hermes function-calling ecosystem — NousResearch

The upstream V7 model card references:

NousResearch/hermes-function-calling-v1

Source:

https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1

This dataset contains structured-output, function-calling and agentic examples used by the Hermes ecosystem.

What This Ollama Package Changes

This repository/package only adapts the upstream release for convenient Ollama use.

Depending on the selected tag, it packages:

• the selected APEX or APEX-Compact GGUF;
• its corresponding Vision mmproj/projector;
• Ollama model metadata/manifest;
• long-context configuration;
• coding/Hermes-oriented sampling defaults.

It does not claim to retrain or independently create the underlying neural-network weights.

Recommended Ollama Profile

The coding-agent variants are configured with:

num_ctx             224000
temperature         0.6
top_p               0.95
top_k               20
min_p               0.05
seed                42
presence_penalty    0
repeat_penalty      1.08

These sampling values follow the Hermes agent recommendation published on the upstream V7 model card, with Ollama’s disabled presence penalty represented as 0.

The upstream author recommends keeping at least 128K context for thinking behavior. This Ollama package uses 224K, below Qwen3.6’s native 262,144-token context ceiling while retaining additional runtime headroom.

Which Variant Should I Use?

APEX-Compact

Use APEX-Compact when:

• VRAM is limited;
• you want to keep a large context window such as 224K;
• you want lower memory pressure for coding-agent workloads;
• you want more headroom for KV cache and Vision projector memory.

Select the APEX-Compact tag on this model page when memory efficiency and long-context headroom are the priority.

APEX

Use APEX when:

• model quality takes priority over memory usage;
• your system has sufficient GPU VRAM and/or system RAM;
• you are willing to trade additional memory usage for the larger APEX representation.

Select the APEX tag on this model page when maximum model quality is the priority.

Usage

Choose the APEX or APEX-Compact tag directly from this Ollama model page, depending on your memory/quality requirements.

The same model can be used through the Ollama CLI, native Ollama API, or OpenAI-compatible API after pulling the desired tag.

Vision

The upstream Qwen3.6 architecture is multimodal, and the Genesis Hermes V7 release explicitly requires the corresponding mmproj file for Vision when using split-GGUF runtimes.

Both the APEX and APEX-Compact Ollama variants are intended to include the matching projector so the published Ollama model can preserve Vision capability.

After pulling a tag, use ollama show locally and confirm that the model reports Vision capability before relying on image input.

Chat Template / Tool Calling

The upstream V7 model card provides a chat_template.jinja and credits froggeric and qweefchief for the template.

For agentic/coding use, correct template handling matters because role formatting, tool definitions, tool calls and tool results must remain consistent.

When repackaging for Ollama, verify each published variant with ollama show --modelfile and confirm that the TEMPLATE is not reduced to a trivial {{ .Prompt }} template.

Important Upstream Notes

The upstream V7 author recommends:

• APEX as the preferred quality-oriented quantization family;
• APEX-Compact when lower memory usage is desired;
• at least 128K context for thinking;
• correct Jinja/template handling in llama.cpp;
• loading the matching mmproj for Vision.

This Ollama package provides both APEX and APEX-Compact, allowing users to choose between memory efficiency and quality.

Attribution Summary

Upstream Links

• Qwen3.6-35B-A3B
  https://huggingface.co/Qwen/Qwen3.6-35B-A3B

• HauhauCS uncensored base
  https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

• DJLougen Hermes source
  https://huggingface.co/DJLougen/hermes-qwen3.5-35b-a3b-GGUF

• LuffyTheFox Genesis Hermes V7
  https://huggingface.co/LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-GGUF

• NousResearch Hermes Function Calling V1
  https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1

License

The upstream repositories identify the model lineage/release under the Apache License 2.0.

When redistributing this Ollama package, retain upstream attribution and review the license and usage terms of each source repository.

Disclaimer

This is a community Ollama repack of an upstream derivative model.

The Ollama package maintainer is not affiliated with or endorsed by Qwen/Alibaba, HauhauCS, DJLougen, LuffyTheFox, or NousResearch, unless explicitly stated otherwise by those parties.

All model names and project names belong to their respective authors/maintainers.