EVIE-4.5B is an open-weight model for visual document retrieval from Tencent, released under Apache License 2.0. It has 4.5B parameters and a 262,144-token context. At 16-bit it needs about 10.9 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 2k downloads a month.
138 Multilingual Tasks Evaluated: Thoroughly evaluated across ViDoRe V1, V2, V3, and JinaVDR across 4 metric families (nDCG, Recall, MAP, MRR @1/5/10).
Runs On
What it takes to serve EVIE-4.5B (4.5B 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 | 9.1 GB | 10.9 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 4.5 GB | 5.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.3 GB | 2.7 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 8, 2026.
EVIE-4.5B on every accelerator the SAVRN Index prices, at every precision
Model Card
By Tencent, published under apache-2.0, revision 879232e4463a.
High-Precision Late-Interaction Retrieval • Dynamic Prefix-MRL (64D–2048D) • Training-Free HAC Token Compression
EVIE-4.5B (Prefix-MRL & HAC) • EVIE-8B (Flagship Teacher) • GitHub: Tencent/EVIE
Release Announcement: All model weights, training pipelines, token compression algorithms (HAC), and evaluation suites have been fully open-sourced in this repository. Full technical details, architectural ablations, and the formal research paper will be updated in an upcoming release.
Highlights
Configuration
- Architecture
- ColQwen3_5
- Context length (tokens)
- 262,144
- Layers
- 32
- Hidden size
- 2,560
- Feed-forward size
- 9,216
- Attention heads
- 16
- Key/value heads
- 4
- Head dimension
- 256
- Vocabulary size
- 248,320
- Model type
- qwen3_5
Identity and Version
- Repository
- tencent/EVIE-4.5B
- Publisher
- Tencent
- Task
- Visual document retrieval
- Modality
- Other
- Library
- colpali-engine
- Parameters
- 4.5B parameters
- Languages
- Not stated by the source
- Revision
- 879232e4463a848a05c1aae50acda5fb30d7239a
- First published
- 2026-09-04
- Last updated
- 2026-10-08
Files and Weights
97 files, 9.1 GB in total. The weights are 2 files totalling 9.1 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| 1_Dense/model.safetensors | Weights | 21.0 MB | 0af96e01abcb |
| model.safetensors | Weights | 9.1 GB | ef4d2da2055c |
| 1_Dense/config.json | Configuration | 214 B | — |
| 2_Normalize/config.json | Configuration | 90 B | — |
| 3_MultiVectorMask/config.json | Configuration | 27 B | — |
| code/compress/cluster.py | Configuration | 7.4 KB | — |
| code/compress/common.py | Configuration | 4.2 KB | — |
| code/compress/dump.py | Configuration | 19.0 KB | — |
| code/compress/eval_index.py | Configuration | 14.6 KB | — |
| code/compress/hac/__init__.py | Configuration | 620 B | — |
| code/compress/hac/core.py | Configuration | 9.2 KB | — |
| code/compress/ladder.py | Configuration | 3.1 KB | — |
| code/shared/aggregate_heads.py | Configuration | 4.8 KB | — |
| code/shared/data_loader.py | Configuration | 9.4 KB | — |
| code/shared/download_jinavdr.py | Configuration | 1.4 KB | — |
| code/shared/eval.py | Configuration | 58.1 KB | — |
| code/shared/merge_seeds.py | Configuration | 6.5 KB | — |
| code/shared/paths.py | Configuration | 1.3 KB | — |
| code/student/scripts/train.py | Configuration | 26.0 KB | — |
| code/teacher/scripts/train.py | Configuration | 12.7 KB | — |
| colpali/colpali_engine/__init__.py | Configuration | 22 B | — |
| colpali/colpali_engine/collators/__init__.py | Configuration | 63 B | — |
| colpali/colpali_engine/collators/visual_retriever_collator.py | Configuration | 4.1 KB | — |
| colpali/colpali_engine/data/__init__.py | Configuration | 97 B | — |
| colpali/colpali_engine/data/dataset.py | Configuration | 5.7 KB | — |
| colpali/colpali_engine/data/sampler.py | Configuration | 4.3 KB | — |
| colpali/colpali_engine/data/task_consistent_sampler.py | Configuration | 4.1 KB | — |
| colpali/colpali_engine/loss/__init__.py | Configuration | 156 B | — |
| colpali/colpali_engine/loss/ard.py | Configuration | 17.4 KB | — |
| colpali/colpali_engine/loss/late_interaction_losses.py | Configuration | 14.8 KB | — |
| colpali/colpali_engine/models/__init__.py | Configuration | 102 B | — |
| colpali/colpali_engine/models/qwen3_5/__init__.py | Configuration | 105 B | — |
| colpali/colpali_engine/models/qwen3_5/colqwen3_5/__init__.py | Configuration | 213 B | — |
| colpali/colpali_engine/models/qwen3_5/colqwen3_5/modeling_colqwen3_5.py | Configuration | 12.4 KB | — |
| colpali/colpali_engine/models/qwen3_5/colqwen3_5/processing_colqwen3_5.py | Configuration | 5.9 KB | — |
| colpali/colpali_engine/trainer/__init__.py | Configuration | 160 B | — |
| colpali/colpali_engine/trainer/ard_trainer.py | Configuration | 33.0 KB | — |
| colpali/colpali_engine/trainer/colmodel_training.py | Configuration | 3.6 KB | — |
| colpali/colpali_engine/trainer/contrastive_trainer.py | Configuration | 12.3 KB | — |
| colpali/colpali_engine/utils/__init__.py | Configuration | — | |
| colpali/colpali_engine/utils/_lik_backend.py | Configuration | 3.3 KB | — |
| colpali/colpali_engine/utils/maxsim.py | Configuration | 2.7 KB | — |
| colpali/colpali_engine/utils/processing_utils.py | Configuration | 2.7 KB | — |
| colpali/colpali_engine/utils/torch_utils.py | Configuration | 3.0 KB | — |
| colpali/colpali_engine/utils/transformers_wrappers.py | Configuration | 742 B | — |
| config.json | Configuration | 3.0 KB | — |
| config_sentence_transformers.json | Configuration | 184 B | — |
| examples/demo/build.py | Configuration | 4.0 KB | — |
| examples/toy/build.py | Configuration | 3.5 KB | — |
| examples/toy/hardneg/allpos/dataset_info.json | Configuration | 667 B | — |
| examples/toy/hardneg/allpos/state.json | Configuration | 247 B | — |
| examples/toy/hardneg/judged/dataset_info.json | Configuration | 644 B | — |
| examples/toy/hardneg/judged/state.json | Configuration | 247 B | — |
| infer.py | Configuration | 3.6 KB | — |
| modules.json | Configuration | 580 B | — |
| processor_config.json | Configuration | 1.2 KB | — |
| sentence_bert_config.json | Configuration | 649 B | — |
| LICENSE | Documentation | 21.5 KB | — |
| NOTICE | Documentation | 528 B | — |
| README.md | Documentation | 18.0 KB | — |
| colpali/LICENSE | Documentation | 1.1 KB | — |
| CITATION.cff | Other | 1.0 KB | — |
| additional_chat_templates/sentence_transformers.jinja | Other | 656 B | — |
| chat_template.jinja | Other | 7.8 KB | — |
| code/compress/run.sh | Other | 7.1 KB | — |
| code/requirements.lock.txt | Other | 222 B | — |
| code/shared/eval_run.sh | Other | 6.6 KB | — |
| code/shared/lib.sh | Other | 5.0 KB | — |
| code/student/run.sh | Other | 4.9 KB | — |
| code/student/scripts/eval_run.sh | Other | 241 B | — |
| code/student/scripts/launch.sh | Other | 7.4 KB | — |
| code/teacher/merge_alpha.sh | Other | 4.7 KB | — |
| code/teacher/run.sh | Other | 3.6 KB | — |
| code/teacher/scripts/eval_run.sh | Other | 228 B | — |
| code/teacher/scripts/launch.sh | Other | 4.4 KB | — |
| colpali/CITATION.cff | Other | 1.1 KB | — |
| colpali/pyproject.toml | Other | 2.0 KB | — |
| env.sh.example | Other | 1.1 KB | — |
| examples/demo/demo/evie_pages/data/test-00000-of-00001.parquet | Other | 166.2 KB | 741b2fde5966 |
| examples/demo/pages/cash.png | Other | 22.6 KB | — |
| examples/demo/pages/headcount.png | Other | 21.0 KB | — |
| examples/demo/pages/invoice_4412.png | Other | 29.0 KB | — |
| examples/demo/pages/org_chart.png | Other | 24.6 KB | — |
| examples/demo/pages/patent.png | Other | 25.2 KB | — |
| examples/demo/pages/q3_revenue.png | Other | 29.2 KB | — |
| examples/demo/pages/traceback.png | Other | 21.3 KB | — |
| examples/demo/pages/warehouse_b.png | Other | 20.9 KB | — |
| examples/demo/run.sh | Other | 1.5 KB | — |
| examples/toy/hardneg/allpos/data-00000-of-00001.arrow | Other | 2.1 KB | 341672c53438 |
| examples/toy/hardneg/judged/data-00000-of-00001.arrow | Other | 1.9 KB | 2568c828632f |
| examples/toy/part_000.parquet | Other | 22.9 KB | 798aa778eb5f |
| requirements.txt | Other | 144 B | — |
| .gitattributes | Repository | 446 B | — |
| .gitignore | Repository | 138 B | — |
| colpali/.gitignore | Repository | 3.3 KB | — |
| tokenizer.json | Tokenizer | 20.0 MB | 06b9509352d2 |
| tokenizer_config.json | Tokenizer | 1.2 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 9.1 GB
Released by Tencent through its official repository on Hugging Face. Read the license.
Built From
- Derived from tencent/EVIE-Preview-4.5B
- Trained on (disclosed) jinaai/jina-vdr
- Trained on (disclosed) vidore/vidore_benchmark
- Trained on (disclosed) vidore/vidore_benchmark_v2
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 9.1 GB |
| 16-bit | 9.1 GB |
| 8-bit | 4.5 GB |
| 4-bit | 2.3 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About EVIE-4.5B
How much GPU memory does EVIE-4.5B need?
About 10.9 GB at 16-bit and 2.7 GB at 4-bit: the weights (4.5B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run EVIE-4.5B 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 EVIE-4.5B commercially?
Yes. EVIE-4.5B is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
What is EVIE-4.5B's context length?
262,144 tokens, from the maximum position embeddings in its published configuration.