available
Weights
Pre-trained and instruction-tuned weights are released.
Inspect this claim’s source ↗Google DeepMind / FAMILY RECORD
Gemma 4: license, size and what is open.
Google’s Gemma 4 generation spans edge and workstation sizes; this atlas record uses the 26B mixture-of-experts checkpoint as its reference point.
Sizes and download
26B total / 4B active (reference checkpoint)
- Gemma 4 26B-A4B
- Total: 26B. Active: 4B. Source ↗
About 15.63 GB of selected documented weights · File listing ↗
Selected artifact: google/gemma-4-26B-A4B-it-qat-q4_0-gguf. Format: GGUF Q4_0 QAT plus multimodal projector. File evidence checked 2026-10-11. Selected artifact license: apache-2.0 ↗
Google-published QAT variant, not original BF16 checkpoint. 14,439,363,584-byte language GGUF plus 1,194,828,160-byte projector. Distinct artifact; does not recheck original evaluation or nine-door findings. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.
Can my computer run it?
Compatibility, runtime memory and speed are untested by Bright.
Choose your computer and inspect documented candidates →Nine things to check
Source check: . Downloads, when known, carry their own check date.
Ask about an image → · Turn speech into text → · Search my own documents →
Checked release: Gemma 4 26B-A4B
Sources checked 2026-09-19
- Initial Gemma 4 model release
- · Primary source
- Google Gemma 4 announcement
- · Primary source
The 26B-A4B reference checkpoint was in the initial March 31 release; later Gemma 4 updates are separate events.
The profile describes this checkpoint, not every release carrying the family name. Developer documentation is not independent reproduction.
THE MODEL / THE TERMS
Know what you are picking up.
- Parameters
- 26B total / 4B active (reference checkpoint)
- Architecture
- Hybrid local/global-attention multimodal transformer; the 26B reference variant uses mixture-of-experts routing.
- Modalities
- text · image
- Context
- Up to 256K tokens on the 26B reference variant
- Weights license
- Apache-2.0
- Commercial use
- Permitted by Apache-2.0.
- Memory
- Google documents a 12B variant running in 16 GB; requirements for the 26B-A4B variant depend on format and context.
- Known quantizations
- Official quantization-aware-trained checkpoints
Specifications are developer-reported. Primary release documentation ↗ · All sources and terms
NINE DIMENSIONS / NO OPENNESS SCORE
What is actually available?
Each component has its own evidence. “Not established” means this source review did not verify availability; it is not a claim that nothing exists.
available
Architecture
The official card describes the dense and MoE variants and attention design.
Inspect this claim’s source ↗available
Inference code
Google documents supported runtimes and deployment routes.
Inspect this claim’s source ↗not established
Training code
The complete code used for the original training runs was not confirmed.
Inspect this claim’s source ↗partial
Training recipe
The model card documents methods but does not provide a full reproduction recipe.
Inspect this claim’s source ↗partial
Data information
Data categories and filtering are described at a high level.
Inspect this claim’s source ↗not established
Training data
No sufficiently specific public artifact was confirmed in this review.
Inspect this claim’s source ↗partial
Evaluation
Official benchmark methodology and results are published; full independent reproduction is separate.
Inspect this claim’s source ↗available
Commercial use
Gemma 4 is released under Apache-2.0.
Inspect this claim’s source ↗WHY THIS RELEASE MATTERS
What to explore.
- Reasoning and tool use
- Vision input
- Wide size range
- Official QAT checkpoints
WHAT THE RECORD DOES NOT ESTABLISH
Keep the limits.
- Audio is limited to the E2B, E4B, and 12B variants
- Training data is described but not released
Arena placement cited by Google is a third-party leaderboard result, but this record does not independently reproduce it.
CONNECTED EVIDENCE
A release becomes a starting point.
- Combining image, audio, video, and text at the edge
Gemma 3n is a multimodal model designed to accept text, images, video, and audio while producing text on resource-constrained devices.
Weights are governed by Gemma terms, not an OSI-approved software license asserted here; training data is summarized rather than released. Device fit and quality vary by hardware, language, and task.
Thematic or family connections, not proof that an older application uses the selected checkpoint.
FOLLOW THE RECORD
Sources, not superlatives.
- Gemma release log ↗ (institution)
- Gemma 4: Byte for byte, the most capable open models ↗ (institution)
- Gemma 4 model card ↗ (report)
Source check: 2026-09-19. AI-assisted editorial review, with no named external expert or independent hardware validation claimed. Licenses and model cards can change. Review the linked terms for the specific artifact and intended use.
Understand open source vs open weights ↗