Qwen3.6-35B-A3B Genesis Hermes V11: Complete GGUF Guide

Overview

Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V11-GGUF is a text-to-text GGUF model with native text, image, and video capability. LuffyTheFox produced it from the uncensored HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive base and transferred data from the Hermes finetune DJLougen/hermes-qwen3.5-35b-a3b-GGUF. The result combines an uncensored behavior profile with Hermes-oriented agent and function-calling behavior, then applies the maintainer’s Genesis tensor-repair process. Its architecture contains 35 billion total parameters but activates about 3 billion per forward pass through a mixture-of-experts design with 256 experts, 8 routed experts, and 1 shared expert per token. It uses Gated DeltaNet linear attention and full softmax attention in a 3:1 ratio across 40 layers, supports a native 262K-token context window, and can extend to 1M tokens with YaRN. The model uses the Apache-2.0 license, the GGUF file format, and the hermes library ecosystem; compatible runtimes include llama.cpp, LM Studio, koboldcpp, and other GGUF runtimes. The most important operational detail is that the model card recommends at least 128K context to preserve thinking capability, --jinja in llama.cpp for correct chat-template handling, F16 key/value cache quantization, maximum GPU offload, 8 active experts, and the V8 APEX quantization.

Best use cases

Local agent and tool-calling workflows. The Hermes transfer targets agent behavior, and the model card provides a JSON system prompt, a discussion-based formatting fix, and commands from the hermes-function-calling-v1 dataset. This makes the model a candidate for local assistants that emit structured tool calls, execute commands through an orchestration layer, and return schema-constrained JSON. You must validate JSON before execution and treat generated commands as untrusted because the uncensored behavior profile does not provide a safety boundary.

Coding and technical reasoning. The recommended thinking-mode settings include a dedicated coding profile: temperature 0.6, top-p 1.0, top-k 20, min-p 0.0, seed 42, disabled presence penalty, and disabled repeat penalty. The hybrid attention design and large total parameter count give it a broad reasoning and code-generation target, while the MoE routing keeps each forward pass near the compute profile of a much smaller active model. The card does not provide independent coding scores, so evaluate it against your repository, language mix, and test suite before deployment.

Long-context document analysis. The model supports 262K native context and a stated 1M-token extension through YaRN. Its Gated DeltaNet component is intended to handle long-context memory, and the Genesis process specifically scans ssm_conv1d tensors and repairs head balance. Use it for large codebases, long technical documents, and multi-document synthesis, but preserve at least 128K context when you want its thinking behavior according to the model card.

Unrestricted creative writing and roleplay. The base model is presented as uncensored, with the maintainer reporting 0 refusals across 465 tests. The non-thinking creative profile uses temperature 0.7, top-p 0.8, top-k 20, min-p 0.0, seed 42, and disabled penalties. This suits users who need fewer refusal interruptions and more permissive fictional content, but it also removes a useful safety filter and can increase operational risk.

Multimodal assistant experiments. The underlying architecture is described as natively multimodal for text, images, and video, and the model card states that vision support requires an mmproj file beside the main GGUF. The static example asks the model to generate an SVG of a pelican riding a bicycle, demonstrating text output in a structured vector format. The card does not provide image or video benchmark results, so treat multimodal support as an integration capability rather than proof of production-level visual quality.

Limitations

The model card makes strong claims about Genesis tensor repair, but it does not provide reproducible benchmark tables, ablation results, perplexity measurements, or independent evaluations that isolate V11 from the base and earlier Hermes versions. The reported “0/465 refusals” describes refusal behavior, not factual accuracy, coding quality, harmlessness, or resistance to prompt injection.

The model has 35B total parameters. Although only about 3B activate per token, memory use still depends on the complete quantized model, runtime buffers, context length, KV cache, and expert placement. The README gives no VRAM requirement, RAM requirement, tokens-per-second result, latency measurement, or practical batch-size guidance. Do not infer hardware suitability from the 3B active-parameter figure. The recommendation to force MoE weights from 40 layers onto the CPU indicates that mixed CPU/GPU placement may be part of the intended setup.

Long context carries a substantial memory cost. The card recommends at least 128K context for thinking, but it does not state a tested quality threshold at 262K or 1M tokens. YaRN extension can change quality and runtime behavior, and the model card supplies no validation results for the 1M-token setting.

The model is uncensored. It may produce unsafe, illegal, abusive, or otherwise inappropriate content, and it may follow dangerous instructions. It is a poor choice for unsupervised public-facing systems, autonomous command execution, or applications that require refusal consistency. The model card reports no bias audit, toxicity evaluation, privacy evaluation, or prompt-injection assessment.

Apache-2.0 permits commercial use, modification, and redistribution subject to the license terms, including preservation of notices and the license text. The license does not make generated outputs accurate, safe, or free from third-party rights concerns. You must also review the provenance and terms of the base and transferred finetune components before distributing a product.

The repository does not provide a Python inference example, a stated training dataset size, training-step count, or inference benchmark. It also does not establish that V11 supersedes every earlier Genesis release for every task. Compare versions on your workload rather than assuming the version number guarantees higher quality.

How it compares

Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V7-GGUF is the better comparison when you want an earlier Genesis Hermes release with the same broad base lineage and need to test regression risk. Pick V11 for the current release and its stated V8 APEX recommendation; pick V7 when an existing prompt, quant, or application was validated against that version. The provided material does not include quality, speed, cost, or memory measurements that prove V11’s advantage, so the tradeoff is version currency versus compatibility with prior testing.

Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V10-GGUF is a close predecessor for users comparing successive Genesis tensor-repair iterations. Choose V11 when you want the latest listed release; choose V10 when its supplied artifacts or observed outputs fit your workload better. No controlled benchmark or file-size comparison is provided, so claims about speed, cost, or quality superiority remain unverified.

Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V9-GGUF may suit reproducibility work that already uses V9 or researchers studying the effect of Genesis processing across releases. V11 is the stronger starting point for a new evaluation because it is the named V11 release and includes current setup guidance. The available information does not report differences in parameterization, quantization size, inference speed, or benchmark quality.

Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V6-GGUF is useful as an older baseline for longitudinal testing. Choose V6 when you need to measure whether later tensor repair or Hermes integration changes behavior; choose V11 for current agent, coding, and long-context experiments. The model card does not provide enough evidence to assign a concrete quality or cost advantage to either version.

Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V5-GGUF is the earliest listed comparison and includes referenced creative and LM Studio discussions that may help users with older integrations. V11 is preferable for a fresh deployment, while V5 can serve as a compatibility or regression baseline. Neither the current card nor the comparison data supplies controlled speed, VRAM, cost, or task-quality measurements.

Technical specifications

The confirmed technical details are:

• Task metadata: Text-to-Text tag; image-text-to-text pipeline tag.

• Library metadata: hermes.

• Parameters: 35B total; about 3B active per forward pass.

• MoE routing: 256 experts; 8 routed experts plus 1 shared expert per token.

• Attention: Gated DeltaNet linear attention plus full softmax attention, in a 3:1 ratio.

• Layer pattern: 10 repetitions of 3 DeltaNet-MoE layers plus 1 Attention-MoE layer.

• Context: 262K native tokens; extendable to 1M with YaRN; at least 128K recommended for thinking capability.

• Modalities: Text, image, and video are described as native capabilities.

• Quantization: V8 APEX is the recommended quant; the maintainer provides a quantization script with Unsloth profile support.

• Cache: Set K-cache and V-cache quantization to F16 for APEX.

• Expert placement: Force MoE weights from 40 layers onto the CPU; use maximum GPU offload; set active experts to 8.

• Runtime compatibility: llama.cpp, LM Studio, koboldcpp, and other GGUF-compatible runtimes.

• llama.cpp requirement: Use --jinja for correct chat-template handling.

• Vision requirement: Place the mmproj file beside the main GGUF.

• Genesis processing: Stage one scans ssm_conv1d tensors and repairs head balance. Stage two detects tensor noise with custom SVD while excluding token_embd.weight, output.weight, ffn_gate_inp_shexp.weight, 1D tensors, biases, and norms. The stated target preserves 99% of signal and learned gradient. Stage three scans blocks in chunks using three parameters and replaces zero chunks with a selected chunk that fits the tensor’s weight distribution.

• Source composition: HauhauCS uncensored base plus data transferred from the DJLougen Hermes finetune; the transfer covers about 2,000 blocks from two FFN expert tensors.

• Evaluation references: HermesBench is linked in the model card, but no numeric score appears in the supplied material.

• Compute used for Genesis development: Google Colab Free on a Tesla T4 GPU.

• Downloads: 1,061,057 at the time of the supplied metadata.

The Genesis method is numerical post-processing of GGUF files, not retraining or finetuning. The maintainer describes it as architecture-independent and applicable to GGUF models, but the supplied material does not independently verify that claim.

Model inputs and outputs

Inputs

• Text prompts using the supplied Qwen/Hermes chat template.

• System prompts, including the recommended Qwen identity prompt.

• Agent prompts that request JSON matching a supplied schema.

• Image and video inputs through the model’s multimodal path, with an mmproj file required for vision support.

• Long contexts up to 262K native tokens; YaRN is stated to support extension to 1M tokens.

• llama.cpp users should pass --jinja so the runtime applies the chat template.

Outputs

• Code and technical explanations.

• JSON for agentic workflows when the system prompt and schema are applied.

• SVG text, such as the pelican-on-a-bicycle example.

• Multimodal responses are claimed by the architecture description, but the supplied card does not define a separate output format or visual benchmark.

• Applications should parse and validate JSON or SVG before use; the model does not guarantee syntactic validity.

Getting started

The supplied material does not include a Python loading or inference example, and it identifies GGUF runtimes rather than a Python model class. Use a GGUF-compatible runtime such as llama.cpp or LM Studio, load the V8 APEX quant, apply the supplied chat template, enable thinking, and start with:

You are Qwen (Tongyi Qianwen), a large language model developed by Alibaba Group's Tongyi Lab.

For llama.cpp, use --jinja, keep at least 128K context for thinking, set K and V cache quantization to F16, offload as much as possible to the GPU, set 8 active experts, and configure 40 layers for CPU MoE placement as recommended by the model card. For coding in thinking mode, use temperature 0.6, top-p 1.0, top-k 20, min-p 0.0, seed 42, disabled presence penalty, and disabled repeat penalty.

Frequently asked questions

Q: Can I use Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V11-GGUF commercially?

A: The metadata lists Apache-2.0. Commercial use is permitted under that license, subject to its notice and redistribution requirements. Review the provenance and terms of the base and transferred finetune components before shipping a product.

Q: What hardware or VRAM do I need?

A: The model card gives no VRAM or RAM number. The 35B total-parameter size means you must size hardware for the complete quantized model, runtime buffers, and context, despite about 3B active parameters per forward pass.

Q: How fast is inference?

A: No tokens-per-second, latency, or throughput result is provided. Speed depends on quantization, CPU/GPU placement, context length, runtime, and hardware.

Q: Which quantization should I use?

A: The model card recommends V8 APEX. It also provides a quantization script with Unsloth profile support for users who want to create their own quants.

Q: Does it support images and video?

A: The specifications describe native text, image, and video support. Vision support requires an mmproj file beside the main GGUF, and the card does not provide multimodal quality scores.

Q: How should I run it with llama.cpp?

A: Use --jinja for chat-template handling. The card recommends F16 K/V cache quantization, maximum GPU offload, 8 active experts, and 40 CPU-forced MoE layers for the APEX setup.

Q: Can I fine-tune this model?

A: The supplied information does not document a finetuning recipe or framework. The Genesis process itself is not finetuning; it is numerical post-processing of GGUF files. The card does provide a quantization script with Unsloth profile support, which is a quantization tool rather than a finetuning method.

Q: Is the model safe for autonomous agents?

A: Treat it as unsafe without an external policy layer. The model is advertised as uncensored and reports 0 refusals in a 465-prompt test, so validate outputs, constrain tools, sandbox commands, and require approval for consequential actions.

This is a simplified guide to an AI model called Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V11-GGUF maintained by LuffyTheFox. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.