This gemma4 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Open-weight model · Image and text to text
vllm-translategemma-4b-it
by Infomaniak Network SA Infomaniak-AI/vllm-translategemma-4b-it
vllm-translategemma-4b-it is an open-weight model for image and text to text from Infomaniak Network SA, released under Gemma Terms of Use. It has 5B parameters and a 131,072-token context. At 16-bit it needs about 11.9 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 743.2k downloads a month.
This is a modified version of google/translategemma-4b-it optimized for deployment with vLLM.
Parameters5B
Context131,072
Weights8.6 GB
Licensegemma
AccessOpen weights
Monthly Downloads743.2k
Runs On
What it takes to serve vllm-translategemma-4b-it (5B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 9.9 GB | 11.9 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 5.0 GB | 6.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.5 GB | 3.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 9, 2026.
vllm-translategemma-4b-it on every accelerator the SAVRN Index prices, at every precision
SAVRN's Notes on vllm-translategemma-4b-it
Infomaniak Network SA published this, not Google. It is google/translategemma-4b-it with the chat template rewritten so vLLM can take the source and target language codes inline in the message, marked by double arrows, rather than in separate fields. Image and text in, text out, 5 billion parameters, a 131,072-token window with a 1,024-token sliding window. The 16-bit weights are 9.9 GB and need 11.9 GB; 8-bit needs 6.0 GB and 4-bit needs 3.0 GB. The cheapest fit we list is one MI300X with 192 GB at $1.85 an hour.
The Gemma Terms of Use govern it whoever re-hosts the weights: commercial use is allowed provided the Prohibited Use Policy travels with every copy you pass on, so redistribution carries paperwork that an internal deployment does not. Confirm your serving stack honors the modified template, and know that our file holds no evaluations for it and no host prices.
Model Card
This is a modified version of google/translategemma-4b-it optimized for deployment with vLLM. The original TranslateGemma model requires a structured payload with dedicated sourcelangcode and targetlangcode fields: However, vLLM does not support these custom content parameters. To maintain compatibility, the chat template has been modified to encode language codes directly in the message content using a delimiter-based format: Format: >>{sourcelang} >>{targetlang} >>{texttotranslate} If you need to provide a custom prompt input The original model uses the new Transformers RoPE configuration format with separate attention type settings: This has been simplified for vLLM compatibility: The…
Excerpt from the card by Infomaniak Network SA, licensed gemma.
Configuration
- Architecture
- Gemma3ForConditionalGeneration
- Context length (tokens)
- 131,072
- Layers
- 34
- Hidden size
- 2,560
- Feed-forward size
- 10,240
- Attention heads
- 8
- Key/value heads
- 4
- Head dimension
- 256
- Vocabulary size
- 262,208
- Sliding window (tokens)
- 1,024
- RoPE base
- 1,000,000
- Model type
- gemma3
Identity and Version
- Repository
- Infomaniak-AI/vllm-translategemma-4b-it
- Publisher
- Infomaniak Network SA
- Task
- Image and text to text
- Modality
- Image and text
- Library
- transformers
- Parameters
- 5B parameters
- Languages
- Not stated by the source
- Revision
- cb3e0b2504f0db832982cca7fafd3ea4aaf777a0
- First published
- 2026-01-26
- Last updated
- 2026-01-27
Files and Weights
16 files, 8.6 GB in total. The weights are 2 files totalling 8.6 GB in safetensors.
Weights2 files · 8.6 GB
Configuration8 files · 104.8 KB
Tokenizer3 files · 39.2 MB
Documentation1 file · 19.3 KB
Other1 file · 17.3 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00002.safetensors | Weights | 5.0 GB | 6187566ec8d0 |
| model-00002-of-00002.safetensors | Weights | 3.6 GB | 40c6a896b227 |
| added_tokens.json | Configuration | 35 B | — |
| config.json | Configuration | 2.5 KB | — |
| generation_config.json | Configuration | 229 B | — |
| model.safetensors.index.json | Configuration | 90.6 KB | — |
| preprocessor_config.json | Configuration | 570 B | — |
| processor_config.json | Configuration | 70 B | — |
| special_tokens_map.json | Configuration | 662 B | — |
| test_chat_template.py | Configuration | 10.2 KB | — |
| README.md | Documentation | 19.3 KB | — |
| chat_template.jinja | Other | 17.3 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 33.4 MB | 7d4046bf0505 |
| tokenizer.model | Tokenizer | 4.7 MB | 1299c11d7cf6 |
| tokenizer_config.json | Tokenizer | 1.2 MB | — |
License and Download
- License
- gemma
- Access
- Open weights, no gate
- Download size
- 8.6 GB
Download from Infomaniak Network SA
Released by Infomaniak Network SA through its official repository on Hugging Face.
Built From
- Described by arXiv:2503.19786
- Described by arXiv:2601.09012
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 8.6 GB |
| 16-bit | 9.9 GB |
| 8-bit | 5.0 GB |
| 4-bit | 2.5 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Compare vllm-translategemma-4b-it
Questions About vllm-translategemma-4b-it
How much GPU memory does vllm-translategemma-4b-it need?
About 11.9 GB at 16-bit and 3 GB at 4-bit: the weights (5B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run vllm-translategemma-4b-it on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use vllm-translategemma-4b-it commercially?
Yes, with conditions. vllm-translategemma-4b-it is released under Gemma Terms of Use. Gemma models are released under Google's Gemma Terms of Use, which permit commercial use and redistribution subject to the Gemma Prohibited Use Policy, whose restrictions must be passed on to anyone the model is distributed to.
What is vllm-translategemma-4b-it's context length?
131,072 tokens, from the maximum position embeddings in its published configuration.
Similar Models
Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency. For more details, please refer to our blog post Qwen3.5. WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL. Empty cells (--) indicate scores not yet available or not applicable. Empty cells (--) indicate scores not…
Lodestar-4B is a 4-billion-parameter decision model. You give it a state (plain text, JSON or a long document) and one or more typed questions; it returns a probability for every listed option. Each answer is read from a single forward pass. The model never generates free text, so there is nothing to parse and no output length to budget for. (Probabilities rounded to three decimals.) The same call from Python, run inside the downloaded folder: A score question takes its levels as a list, lowest first, and returns {"type": "score", "score":, "probabilities": {"0": p0, "1": p1,...}}. Every question is rendered into one chat prompt (thinking disabled): The probability of each option is the…
Chandra 2 is a state of the art OCR model from Datalab that outputs markdown, HTML, and JSON. It is highly accurate at extracting text from images and PDFs, while preserving layout information. Try Chandra in the free playground, or use the hosted API for higher accuracy and speed. - 85.8% olmocr bench score (sota), 77.8% multilingual bench score (12% improvement over Chandra 1) - Significant improvements to math, tables, complex layouts - 90+ language support with major accuracy gains - Convert documents to markdown, HTML, or JSON with detailed layout information - Reconstructs forms accurately, including checkboxes - Strong performance with tables, math, and complex layouts - Extracts…
Expert-paged build of Vontra/Qwen3.8-Flash-Next-MLX-4bit. The weights that are read a fraction at a time live in their own containers, so a machine loads what it needs rather than all Total 105.46 GiB. Of that, 103.94 GiB is the source build, whose bytes moved into containers rather than being copied, and 1.52 GiB is the draft head, which no published build of this model carries. Where the weights fit they are filled from experts.bin and the model runs the stock path at stock speed; where they do not, they stream from disk. Reading the machine decides that, not a flag. To override that: GBXPAGING=off holds the experts resident, GBXPLE=off holds the n-gram table resident. Checked at build…
LT-OPD trains vision-language models with fewer visual tokens through on-policy self-distillation. This Qwen3.5-4B model combines CDPruner with an MLP that aggregates discarded tokens into their nearest retained tokens, preserving their original positions without adding tokens. Use Python 3.12 and the CUDA environment in the installation guide. Load the model with LT-OPD's visual-token compression runtime