Instructions to use jankin123/gameplayer-ov2-8B-beta with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jankin123/gameplayer-ov2-8B-beta with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="jankin123/gameplayer-ov2-8B-beta")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("jankin123/gameplayer-ov2-8B-beta", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jankin123/gameplayer-ov2-8B-beta with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jankin123/gameplayer-ov2-8B-beta"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jankin123/gameplayer-ov2-8B-beta",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/jankin123/gameplayer-ov2-8B-beta

SGLang

How to use jankin123/gameplayer-ov2-8B-beta with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jankin123/gameplayer-ov2-8B-beta" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jankin123/gameplayer-ov2-8B-beta",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jankin123/gameplayer-ov2-8B-beta" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jankin123/gameplayer-ov2-8B-beta",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use jankin123/gameplayer-ov2-8B-beta with Docker Model Runner:
```
docker model run hf.co/jankin123/gameplayer-ov2-8B-beta
```

Gameplayer-OV2-8B

This package contains two standalone, complete BF16 LLaVA-OneVision-2 8B models. Each model loads directly from its own directory.

Background

`models/general`

general is the shared multi-game FPS model. It uses one set of weights across different FPS environments and receives each environment's available controls through the prompt capability manifest. It is the default model for new cross-game inference.

`models/specific`

specific is the ViZDoom-focused model. It is intended for the established ViZDoom scenarios and for reproducing the included qualitative examples.

Both models use the same compact action-chunk interface:

<p>m=<mode> t=<target> h=<8|12|24> i=<interrupt></p>
<a>k=<keys> x=<dx> y=<dy> l=<0|1> r=<0|1> m=<0|1></a>
...
<eoc>

Inference Scheme

Inference is synchronous and event-driven:

Pause the complete game simulation at a decision boundary.
Capture the latest causal RGB observation.
Build a prompt containing the task and environment capability manifest.
Generate one action chunk while the simulation remains paused.
Parse and validate the complete chunk.
Install the first validated action, acknowledge it, and resume simulation.
Execute until the chunk expires or a perception event interrupts it.
Pause again and request a new chunk.

The environment must not advance while the model is generating. This prevents the player from becoming unresponsive when model latency is high.

Inference Configuration

Setting	`general`	`specific`
Precision	BF16	BF16
Observation	Latest causal RGB frame	Latest causal RGB frame
Context	Latest round, one turn	Latest round, one turn
Plan contract	`target_v1`	`target_v1`
Horizons	`8,12,24`	`8,12,24`
Scheduler	Single pass, event or expiry	Single pass, event or expiry
Sampling	Temperature 0, top-p 1	Temperature 0, top-p 1
Max new tokens	640	640
Action guard	Strict capability validation; optional format retry	Canonicalize recoverable syntax
Simulation during generation	Paused	Paused

A compatible FPS runtime checkout must provide the codec/bitcost libraries, game environments, OpenCV, ffmpeg, Transformers, and FlashAttention 2. Set FORMAL to that runtime's formal experiment directory. The scripts use $FORMAL/.venv-v31/bin/python by default; override PYTHON_BIN when needed.

Quick Start

Clone the model package, enter its root, and select the compatible runtime:

git clone https://huggingface.co/jankin123/gameplayer-ov2-8B-beta
cd gameplayer-ov2-8B-beta
export FORMAL=/path/to/formal-runtime

General model in ViZDoom

bash scripts/infer_vizdoom.sh general \
  deadly_corridor.cfg 20273821 general_deadly

Disable the optional format retry when raw generation behavior is required:

FORMAT_RETRY=0 bash scripts/infer_vizdoom.sh general \
  deadly_corridor.cfg 20273821 general_deadly_raw

Specific model in ViZDoom

bash scripts/infer_vizdoom.sh specific \
  deadly_corridor.cfg 20271001 specific_deadly

The complete interface is:

infer_vizdoom.sh MODEL SCENARIOS SEEDS [TAG] [MAX_DECISIONS] [RECORD_DIR]

MODEL is general or specific. SCENARIOS and SEEDS accept comma-separated lists. Results are written under outputs/vizdoom/.

General model in Red Eclipse

bash scripts/infer_redeclipse.sh 20269100 general_redeclipse

The complete interface is:

infer_redeclipse.sh [SEED_BASE] [TAG] [RUNS] [DECISIONS] [BOTS]

The default configuration uses one run, 64 decisions, four bots, and the event_or_expiry scheduler. Results are written under outputs/redeclipse/.

Reproduce the included ViZDoom examples

bash scripts/reproduce_specific_vizdoom_examples.sh

This runs the following deterministic settings without overwriting the bundled videos:

Scenario	Seed
`deadly_corridor.cfg`	20271001
`rocket_basic.cfg`	20271001
`defend_the_line.cfg`	20271002

New videos are written to reproduced_specific_demos/<timestamp>/. Encoded MP4 bytes may differ across ffmpeg versions; compare the scenario, seed, action contract, game outcome, and synchronized clock configuration.

Included examples

demos/vizdoom/ours_v27_deadly_corridor_seed20271001_success.mp4
demos/vizdoom/ours_v27_rocket_basic_seed20271001_success.mp4
demos/vizdoom/ours_v27_defend_the_line_seed20271002_success.mp4

Verify the model shards and bundled demonstrations after download:

sha256sum -c SHA256SUMS

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Image-Text-to-Text

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support