GLM-4.7-Flash-30B-A3B GGUF Released with Turbo Brilliance

GLM-4.7-Flash-30B-A3B GGUF Released with Turbo Brilliance

At a Glance

Item Value
Repository DavidAU/GLM-4.7-Flash-30B-A3B-Hellfire-Turbo-Brilliance-50x-AGI-NEO-GGUF
Publisher guide Z.ai (Zhipu AI): models and licenses
Published 2026-11-17
License apache-2.0
Formats GGUF
Paper arXiv:2508.06471
Source type Unverified (not confirmed by a primary source)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Please note that information regarding this model has not been officially confirmed and is currently reported as a public release by an individual developer.

It is reported that DavidAU has released a quantized model, “GLM-4.7-Flash-30B-A3B-Hellfire-Turbo-Brilliance-50x-AGI-NEO-GGUF", on Hugging Face, based on zai-org’s open-weight model “GLM-4.7-Flash". This model is said to be a derivative model that has a 30B-scale Mixture of Experts (MoE) structure, balancing inference efficiency with high performance, while applying the “Turbo Brilliance" system which incorporates numerous prompt syntaxes and inference modes.

According to “GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models“, a related paper on the GLM series that served as the base for this model, it features a hybrid inference approach supporting both thinking mode and direct response mode, and is designed to achieve high performance in agentic, reasoning, and coding (ARC) tasks.

Specifications

  • Total parameters: 30B
  • Active parameters: 3B per token (4 out of 64 experts active)
  • Architecture: Mixture of Experts (MoE / Glm4MoeLiteForCausalLM)
  • Context length: Up to 200k (recommended settings for use are 32k to 64k)

Performance

Comparison tables of official benchmarks listed on the model card of this model and the card of the base model “zai-org/GLM-4.7-Flash" are as follows. Please note that these are figures measured and published by the publisher of the base model, and since a GGUF format quantized version is provided this time, actual measured performance may degrade slightly due to the quantization process.

Benchmark GLM-4.7-Flash Qwen3-30B-A3B-Thinking-2507 GPT-OSS-20B
AIME 25 91.6 85.0 91.7
GPQA 75.2 73.4 71.5
LCB v6 64.0 66.0 61.0
HLE 14.4 9.8 10.9
SWE-bench Verified 59.2 22.0 34.0
τ²-Bench 79.5 49.0 47.7
BrowseComp 42.8 2.29 28.3

From these measurement results, it can be read that the original model, GLM-4.7-Flash, has remarkable strengths in tasks close to real environments and agent performance. It recorded 59.2 on SWE-bench Verified, which measures problem-solving capabilities on actual GitHub repositories, pulling far ahead of the comparison targets Qwen3-30B-A3B-Thinking-2507 (22.0) and GPT-OSS-20B (34.0). In addition, it holds a significant advantage over many competing models with 79.5 on τ²-Bench and 42.8 on BrowseComp, which are agent-type tasks involving multi-turn dialogue and tool use.

Furthermore, it scored 75.2 on GPQA, which targets Ph.D.-level difficult questions in physics, chemistry, biology, and other fields, and 14.4 on HLE, a collection of extremely difficult questions created by experts in various fields, maintaining top-tier figures even in areas requiring advanced knowledge.

On the other hand, the accuracy measured in solving US Math Olympiad preliminary problems, AIME 25, is 91.6, which is a close margin not far behind GPT-OSS-20B’s 91.7, though falling slightly short. Also, the LCB v6 score is 64.0, which falls below Qwen3-30B-A3B-Thinking-2507’s 66.0, indicating that there are fields where competing models have the advantage depending on the nature of the task.

Strengths and Use Cases

The original model, GLM-4.7-Flash, is designed specifically for agentic, reasoning, and coding (ARC), excelling in practical software development, multi-turn dialogue processing, and complex problem solving. Having recorded high figures in SWE-bench Verified, which simulates bug fixing in real repositories, and τ²-Bench, which evaluates tool integration, it is suitable for use as an autonomous coding assistant or workflow automation agent. It also possesses deep reasoning capabilities capable of handling advanced academic and scientific tasks such as GPQA and HLE.

In the newly released derivative model “GLM-4.7-Flash-30B-A3B-Hellfire-Turbo-Brilliance-50x-AGI-NEO-GGUF", the “Turbo Brilliance" system advocated by developer DavidAU is incorporated on top of these fundamental performances. According to the model card, this model is constructed assuming a wide range of use cases and functions such as the following:

  • Advanced Creative Writing and Roleplay:
    As indicated by the tags, it supports fiction writing, story creation, and general roleplaying. It is claimed to suppress overly restrictive responses by standard safety filters, enabling dynamic and powerful textual expression and character creation.

  • Diverse Thinking Modes and Persona Constructs:
    Aside from 27 types of inference modes and 27 types of instruct modes, it internalizes numerous AI/AGI constructs (role definitions) such as “Brainstorm", “Universal Teacher", “Agent", and “Coder". It is stated that by using designators like {REASON:xbrainstorm} within prompts, users can instantly switch to thinking styles tailored to their purpose, such as multi-faceted brainstorming or rigorous logical verification.

  • Lightweight Operations Leveraging Resource Efficiency:
    Due to the MoE structure where active parameters per token are kept to 3B out of a total 30B parameters, computational load is kept relatively low. According to the publisher’s explanation, it is reported that good generation speeds are easily maintained not only in GPU-only configurations, but also in offloading environments to CPU and main memory, or even in CPU/RAM-only environments.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 31.2B parameters (taken from the base model zai-org/GLM-4.7-Flash)

Your VRAM Quantization File size Est. memory needed
12GB (RTX 4070 / 3060 12GB, etc.) IQ2_M 9.9GB 11.8GB
16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) IQ3_M 13.0GB 15.6GB
24GB (RTX 4090 / 3090, etc.) Q4_K_M 17.6GB 21.1GB
32GB (RTX 5090, etc.) Q6_K 23.9GB 28.7GB
48GB (RTX 6000 Ada / A6000, etc.) Q8_0 30.9GB 37.1GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp as-is.

It is distributed in GGUF, so no conversion is needed.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compression: the IQ2_M build measures 2.72 bits per weight — about 17% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
AtomicChat/Maxwell-V1-Flash-GGUF 36.8B 24GB apache-2.0 Maxwell-V1-Flash-GGUF 35B MoE Model: 24GB+ VRAM (2026-11-18)
LiquidAI/LFM2.5-24B-A2B-Base 24.0B 80GB lfm1.0 LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-FP8 24.0B 32GB lfm1.0 LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-MLX-bf16 24.0B 80GB lfm1.0 LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-MLX-8bit 24.0B 32GB lfm1.0 LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17)

How to Get It

This model is distributed in GGUF format on the Hugging Face repository “DavidAU/GLM-4.7-Flash-30B-A3B-Hellfire-Turbo-Brilliance-50x-AGI-NEO-GGUF". It is not set as a gated model requiring prior agreement to terms of use, so anyone can directly download and use it. Distributed quantization variations include MTP quantized versions using NEO imatrix, and “MAX" versions where output tensors are set to full 16-bit precision.

Recommended generation parameters and quick-start settings guided in the model card are as follows:

  • Temperature: 1.0
  • Repetition Penalty: 1.0 (disabled)
  • Top-K: 64
  • Top-P: 0.95
  • Min-P: 0.05
  • Recommended context length: 32k to 64k (maximum context length 200k)

It is specified that by typing help alone and sending it at the start of a dialogue, you can interactively check the list of internal Turbo Brilliance systems and inference modes. Additionally, if the inference application being used allows enabling “MTP" settings, it is guided that generation speed improves by about 10% to 30%.

Note that as official inference and serving environments for the base model GLM-4.7-Flash, the latest branches of vLLM and SGLang are supported, while for GGUF formats like this model, loading from llama.cpp and various compatible frontends is anticipated.

Other Models for the Same Task

Recent text generation models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.

See all text generation models →

What to Read Next

Sources

This article contains unverified information. We will append an update note once it is confirmed by a primary source.