DEV.co
Open-Source LLM · DavidAU

GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF

GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF is a 30B mixture-of-experts model with ~2B active parameters, quantized for CPU/GPU inference. It is a community-modified version of the original GLM-4.7-Flash with content-filter modifications ('heretic'ed') and custom quantization tuning. The model is optimized for creative writing, storytelling, and reasoning tasks. It requires careful prompt engineering to achieve desired output quality and runs best with specific sampler settings (smoothing_factor, temperature, top-p). Gated: false; MIT license.

Source: HuggingFace — huggingface.co/DavidAU/GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF
Unknown
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
41.7k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperDavidAU
ParametersUnknown
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads41.7k
Likes393
Last updated2026-06-12
SourceDavidAU/GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF

What GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF is

30B MOE architecture with 4 active experts (~2B parameters active); GGUF quantization variants (IQ4_NL, Q5_1, Q4_1, Q8_0 specialized; standard quants also available). Dual Imatrix optimization (NEO + Code datasets) and 16-bit precision output tensors. Context length: up to 200k (non-roped). Generated with LLAMACPP commit 7789 (Jan 21 2026). Known issues: Flash Attention offloaded to CPU causing low token generation speed; requires llama_HF config for text-generation-webui. Model is content-filter modified; baseline model: https://huggingface.co/zai-org/GLM-4.7-Flash.

Quickstart

Run GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="DavidAU/GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Creative Fiction and Storytelling

Designed for narrative generation (plot, subplot, scene continuation). Model supports vivid writing and multi-genre fiction. Excels with context of 8k–16k minimum; may 'polish' output mid-generation. Requires explicit prompt direction for desired content tone/style.

Deep Reasoning and Extended Analysis

Supports 'thinking' blocks with up to 131k max output tokens. Suitable for complex problem-solving, brainstorming, and iterative refinement. May require 16k+ context to capture full reasoning chain and polished final output.

Self-Hosted Private AI Applications

Runs on GPU/CPU due to low active parameter count (~2B). Ideal for organizations deploying private LLM infrastructure without cloud dependency. No gating; MIT license permits self-hosting and custom modification.

Running & fine-tuning it

Estimated VRAM: 8–16 GB for full-precision on GPU; 4–8 GB for quantized GGUF variants (Q4_1, Q5_1, IQ4_NL) depending on batch size and context length. CPU-only inference feasible (~2B active parameters) with slower token generation. Context handling up to 200k requires sufficient RAM/VRAM and batch-size constraints. Verify with your target quantization variant and inference engine (llama.cpp, KoboldCpp).

Not stated in card. Model is described as 'fine tune' capable (listed in tags), but no LoRA/QLoRA compatibility, training methodology, or fine-tuning guidance provided. Fine-tuning on custom GGUF quantized variants is not standard practice; would require conversion to full precision or source-format weights (referenced at https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be). Requires review for your use case.

When to avoid it — and what to weigh

  • High-Speed Production Chat/Inference — Flash Attention disabled due to CPU offload issues and slow token generation. Not suitable for latency-sensitive applications or high-throughput serving without inference optimization.
  • Factual Accuracy or Guardrailed Content Required — Model is content-filter modified ('uncensored'). Not designed for production systems requiring strict refusal guardrails, guardrailed safety boundaries, or factual consistency guarantees. Requires additional validation layers.
  • Out-of-Box Safe Commercial Deployment — Specialized 'uncensored' variant; will generate explicit, profane, or unfiltered content with minimal prompt direction. Organizations with strict content policies should use the non-uncensored GLM-4.7-Flash variant instead.
  • Immediate Production Use Without Testing — Card notes quant quality depends on LLAMACPP commit 7789+; older commits perform poorly. Requires validation on your target inference framework (llama.cpp, text-generation-webui, KoboldCpp, Silly Tavern) and careful sampler tuning before deployment.

License & commercial use

MIT license. Permissive OSI-approved license permits commercial use, modification, and distribution. No gating or access restrictions. Modifications (content-filter removal) are permitted under MIT.

MIT is a permissive OSI license and clearly allows commercial use. However, this model is a community modification ('uncensored/heretic'ed') of the base GLM-4.7-Flash (developed by zai-org). Commercial deployments should: (1) verify zai-org's original GLM-4.7-Flash license and commercial terms; (2) accept that content-filter removal may violate organizational policy or customer expectations in regulated industries; (3) ensure content governance and liability frameworks are in place for unfiltered model outputs. Recommend legal/compliance review for regulated industries.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceMedium
Security considerations

Content-filter removal increases risk of generating harmful, explicit, or unfiltered content in uncontrolled deployments. Model will generate profane, x-rated, or violent content with minimal prompt direction. No stated security audits, adversarial robustness testing, or content safety evaluation. Self-hosted deployment removes cloud-vendor security features (logging, DDoS protection, etc.). Inference via llama.cpp or KoboldCpp on exposed networks may leak model outputs or weights. Recommend: air-gapped deployment, input sanitization, output filtering for regulated use cases, and audit logging.

Alternatives to consider

GLM-4.7-Flash (zai-org, official)

Original, non-'uncensored' version with refusal guardrails intact. Use if you need safer content filtering, vendor-backed stability, or production guardrails. Same MOE architecture; no community modification risks.

Qwen2.5-32B or Llama-3.1-70B (open quantized variants)

Larger instruction-tuned alternatives with broader reasoning and code-gen capabilities. Standard quantization support, mature inference stacks, and wider adoption. Better for production systems requiring stable, vetted pipelines.

Mistral-7B or Mixtral-8x7B

Smaller, faster alternatives for creative/reasoning tasks with lower hardware overhead. Strong community support, comprehensive documentation, and production-grade serving infrastructure (vLLM, TGI). Trade-off: fewer 'thinking' tokens and slightly lower quality on complex reasoning.

Software development agency

Ship GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF with senior software developers

This model excels at creative writing and reasoning but requires careful sampler tuning and governance frameworks for unfiltered content. Assess hardware, inference framework compatibility (llama.cpp, KoboldCpp, Silly Tavern), and content policy fit before production. Review our advanced deployment guide or request a custom LLM evaluation.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF FAQ

Can I use this model commercially?
MIT license permits commercial use. However, this is a community-modified 'uncensored' variant of GLM-4.7-Flash (zai-org). Verify the original model's commercial terms. Content-filter removal may create liability or policy conflicts in regulated industries (healthcare, finance, legal). Legal review recommended for production deployments.
What GPU/VRAM do I need?
Estimated 8–16 GB VRAM for quantized GGUF variants (Q4_1, Q5_1, IQ4_NL) at 200k context. With smaller context (8k–16k) and 4-bit quants, 6–8 GB may suffice. CPU-only inference is feasible (~2B active params) but slow. Verify exact VRAM needs by testing your target quantization and context length with llama.cpp or KoboldCpp. Batch size and sequence length directly impact memory.
Why is token generation slow?
Flash Attention is offloaded to CPU in some inference scenarios, causing bottlenecks. Card recommends disabling Flash Attention until upstream llama.cpp/AI pipeline resolves the issue. Use KoboldCpp or text-generation-webui with sampler tuning (smoothing_factor 1.5, rep_pen 1.1) to improve output quality, not speed. Production inference may require alternative quantization or larger hardware.
How do I get the best output quality?
Model requires specific tuning: use smoothing_factor 1.5 (KoboldCpp, text-generation-webui, Silly Tavern), temperature 0.8–1.0, top_p 0.6–0.95, rep_pen 1.0–1.15, and min context 8k–16k. For content quality, provide explicit prompt direction (e.g., specify style, tone, language, content level). See DavidAU's 'Maximizing Model Performance' guide (linked in card) for advanced sampler settings and examples. Expect iterative refinement; model may 'polish' output mid-generation.

Custom software development services

From first prototype to production, DEV.co delivers software development services around tools like GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX-GGUF. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Deploy GLM-4.7-Flash-Uncensored Safely

This model excels at creative writing and reasoning but requires careful sampler tuning and governance frameworks for unfiltered content. Assess hardware, inference framework compatibility (llama.cpp, KoboldCpp, Silly Tavern), and content policy fit before production. Review our advanced deployment guide or request a custom LLM evaluation.