gemma-4-12B-it-qat-q4_0-unquantized-assistant
Gemma 4 is a family of open models built by Google DeepMind, designed for text and image input (with audio supported on some variants) and text output. It features a context window of up to 256K tokens and is available in both pre-trained and instruction-tuned variants.
Fact sheet
Parameters
Context length
Architecture
Shape
Weights stored as
Task
Languages
Built on
Released
Last updated
Downloads (last 30 days)
Licence and access
Licence
Access
Security
Hugging Face security scan
Weight files by format
Facthory’s checks
Every weight file is safetensors or GGUF.
No file flagged.
The card declares its licence.
The card reports safety or harm evaluations with their results.
GPU calculator
How much memory this model needs to serve at a quantization, context and number of concurrent sequences, which GPUs hold it, and how fast it can generate. Estimates from the model’s published size and shape, with the assumptions below.
Weights
Context per sequence: 8K tokens
Concurrent sequences: 1
KV cache
Memory
1 GB
Weights: 0.8 GB
KV cache: 0.1 GB
Runtime: 0.1 GB
| GPU | Memory | GPUs needed | Per sequence | All sequences |
|---|---|---|---|---|
| NVIDIA L4 | 24 GB | 1 × | 197 tokens/s | 197 tokens/s |
| NVIDIA L40S | 48 GB | 1 × | 568 tokens/s | 568 tokens/s |
| NVIDIA RTX 6000 Ada | 48 GB | 1 × | 631 tokens/s | 631 tokens/s |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB | 1 × | 1,178 tokens/s | 1,178 tokens/s |
| NVIDIA A100 80GB SXM | 80 GB | 1 × | 1,340 tokens/s | 1,340 tokens/s |
| NVIDIA H100 SXM | 80 GB | 1 × | 2,202 tokens/s | 2,202 tokens/s |
| NVIDIA H200 SXM | 141 GB | 1 × | 3,155 tokens/s | 3,155 tokens/s |
| AMD Instinct MI300X | 192 GB | 1 × | 3,484 tokens/s | 3,484 tokens/s |
| AMD Instinct MI325X | 256 GB | 1 × | 3,944 tokens/s | 3,944 tokens/s |
Assumptions
- Runtime memory (activations, buffers) is 10% of weights and KV cache, plus 1 GB per GPU for the framework.
- A GPU is filled to at most 90% of its memory; tensor parallelism uses 1, 2, 4 or 8 GPUs in one server.
- Generation is limited by memory bandwidth at 60% of the vendor’s peak: each step reads the weights once and every sequence’s KV cache. A mixture-of-experts model reads only its active experts, so it runs faster than shown.
- GPU memory and bandwidth are the vendors’ published figures.
Profile
Kind
Lineage
Family
Quantization
Uses
In and out
Written on 10 Nov 2026 by Workers AI (@cf/zai-org/glm-4.7-flash) from the model card and the hub’s metadata: a reading of the publisher’s text, not checked by Facthory.
Published benchmarks
The model card publishes no benchmark results.
EU AI Act
What the EU AI Act asks of a general-purpose AI model depends on who places it on the EU market and how. These facts bear on it:
- Licence: apache-2.0
- Access: Weights download without approval
- Parameters: 422.9M
General-purpose AI model (Article 3(63))
Not determined. No estimate of its training compute is published, and the Commission’s indicative criterion (10²³ FLOP) needs one. If it is a general-purpose AI model, the obligations below bind its provider.
Systemic risk (Article 51)
Not presumed. No estimate of its training compute is published; without one, or a designation by the Commission under Article 52, Article 55 does not apply.
Open-source exemption (Article 53(2))
Applies. A free and open-source licence, weights anyone can download and no systemic risk: points (a) and (b) of Article 53(1) and Article 54 do not bind its provider. The copyright policy and the training-content summary of points (c) and (d) still do.
General-Purpose AI Code of Practice
Google signed it, by the European Commission’s list of signatories. The Code is voluntary; signing it shows how a provider means to meet Articles 53 and 55.
What the card and the provider document
Lifted by the open-source exemption (Article 53(2)).
