Open-weight model
tars-paper-v4-20261003-13b-mask-s42-ep1
by Pham Ba Cong pbcong/tars-paper-v4-20261003-13b-mask-s42-ep1
tars-paper-v4-20261003-13b-mask-s42-ep1 is an open-weight model from Pham Ba Cong. It has 13.4B parameters and a 4,096-token context. At 16-bit it needs about 32 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
Epoch-1 full fine-tuning checkpoint in original LLaVA format. Load with the pinned TARS/LLaVA loader, not an HF LoRA loader. Paper profiles are reconstructions with documented assumptions, not author checkpoints.
Runs On
What it takes to serve tars-paper-v4-20261003-13b-mask-s42-ep1 (13.4B 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 | 26.7 GB | 32.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 13.4 GB | 16.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 6.7 GB | 8.0 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 18, 2026.
Model Card
Epoch-1 full fine-tuning checkpoint in original LLaVA format. Load with the pinned TARS/LLaVA loader, not an HF LoRA loader. Paper profiles are reconstructions with documented assumptions, not author checkpoints.
Excerpt from the card by Pham Ba Cong.
Configuration
- Architecture
- LlavaLlamaForCausalLM
- Context length (tokens)
- 4,096
- Layers
- 40
- Hidden size
- 5,120
- Feed-forward size
- 13,824
- Attention heads
- 40
- Key/value heads
- 40
- Vocabulary size
- 32,000
- RoPE base
- 10000
- Stored precision
- bfloat16
- Model type
- llava_llama
Identity and Version
- Repository
- pbcong/tars-paper-v4-20261003-13b-mask-s42-ep1
- Publisher
- Pham Ba Cong
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- 13.4B parameters
- Languages
- Not stated by the source
- Revision
- c5ee8a4f31d34e26797f3f8f3fd29288bfdd8ac4
- First published
- 2026-10-03
- Last updated
- 2026-10-03
Files and Weights
16 files, 26.7 GB in total. The weights are 6 files totalling 26.7 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00006.safetensors | Weights | 5.0 GB | 5cbf9e0c9e34 |
| model-00002-of-00006.safetensors | Weights | 5.0 GB | 8fd545a688a4 |
| model-00003-of-00006.safetensors | Weights | 5.0 GB | f82aa7ed709a |
| model-00004-of-00006.safetensors | Weights | 4.9 GB | 12922c681471 |
| model-00005-of-00006.safetensors | Weights | 4.9 GB | 405189f1a7e1 |
| model-00006-of-00006.safetensors | Weights | 1.9 GB | f6736dfd9def |
| checksums.json | Configuration | 1.4 KB | — |
| config.json | Configuration | 1.4 KB | — |
| generation_config.json | Configuration | 154 B | — |
| model.safetensors.index.json | Configuration | 79.1 KB | — |
| reproduction.json | Configuration | 50.0 KB | — |
| special_tokens_map.json | Configuration | 438 B | — |
| README.md | Documentation | 360 B | — |
| .gitattributes | Repository | 1.5 KB | — |
| tokenizer.model | Tokenizer | 499.7 KB | 9e556afd4421 |
| tokenizer_config.json | Tokenizer | 964 B | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 26.7 GB
Released by Pham Ba Cong through its official repository on Hugging Face.
Built From
- Derived from liuhaotian/llava-v1.5-13b
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 26.7 GB |
| 16-bit | 26.7 GB |
| 8-bit | 13.4 GB |
| 4-bit | 6.7 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About tars-paper-v4-20261003-13b-mask-s42-ep1
How much GPU memory does tars-paper-v4-20261003-13b-mask-s42-ep1 need?
About 32 GB at 16-bit and 8 GB at 4-bit: the weights (13.4B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run tars-paper-v4-20261003-13b-mask-s42-ep1 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.
What is tars-paper-v4-20261003-13b-mask-s42-ep1's context length?
4,096 tokens, from the maximum position embeddings in its published configuration.