Cialtion/SimpleTool-VLM-FP8

SimpleTool-VLM

The multimodal line is maintained separately from the ICML 2026 SimpleTool / RealtimeTool release. The original research implementation remains at HaxxorCialtion/SimpleTool.

Paper: RealtimeTool: Parallel Decoding for Real-Time LLM Function Calling · ICML 2026 paper page · ICML poster

SimpleTool-VLM is a research release of a Qwen3-VL-4B model adapted for structured tool calls from images and text. It emits a function name and only the argument heads required by the supplied tool schema. The public release contains inference code and reproducibility notes; model weights are hosted separately on Hugging Face.

Highlights

• Direct and Adaptive modes.
• Function plus dynamically sized arg1...argN heads (N is the maximum property count in the supplied tools).
• Image input is passed as pixels; invoice values are not copied into hidden prompt fields.
• Raw branch outputs and legality are preserved for auditing.

Release tracks

Track	Status	Scope
SimpleTool-VLM	Available	Qwen3-VL-4B image/text to structured tool calls; Direct is the reproducible baseline.
SimpleTool-VLM-Adaptive	Research track	Short adaptive content before the tool heads; experiments and limitations are documented separately.
SimpleTool-VLA	Coming soon	Vision-language-action interfaces and embodied control integrations.

The tracks share the positional SimpleTool protocol, but they are separate research claims. Results from the Adaptive track should not be read as a replacement for the validated Direct baseline.

Visual case studies

The invoice example image and the 4090 invoice recording show image pixels entering a real structured tool-call loop. The video is a demonstration of the protocol, not an OCR or accounting benchmark.

The Panda grasp preview is included as an embodied research preview. Its prompt explicitly supplied the table geometry, object size, grasp order, heights, gripper settings, and a calibrated RGB-D localization tool. It therefore demonstrates a prompt-assisted, tool-mediated grasp loop, not autonomous visual planning or learned trajectory discovery.

For interface context, see the SimpleTool-VLM vs Jev comparison.

Intended use

Research on real-time multimodal structured decisions, tool calling, and embodied-agent interfaces. This is not an OCR benchmark or a general invoice accounting system. Validate every action before execution.

Model variants

• BF16 merged checkpoint: Cialtion/SimpleTool-VLM.
• W4A16 checkpoint: Cialtion/SimpleTool-VLM-W4A16 (when the upload is complete).
• FP8 checkpoint: Cialtion/SimpleTool-VLM-FP8 (when the upload is complete).

The quantized variants use the same task adapter and tokenizer contract; backend support depends on the serving stack.

Quick start

The lightweight HTTP implementation is in inference/. It accepts one PNG, text messages, and OpenAI-style function tools. See examples/invoice_request.json. The service returns prediction.raw, prediction.legal, events, and server timings.

Adaptive track

The Adaptive track is described in adaptive_cot/README.md. It is intentionally kept separate because short generated content is not a guaranteed reasoning trace and current distribution-shifted evaluations do not show a stable advantage over Direct.

VLA roadmap

vla/README.md records the planned SimpleTool-VLA interface. The prompt-assisted Panda grasp case is a research preview with an explicit claim boundary; the autonomous VLA release is coming soon.

Limitations

This release does not claim strict adapter-to-merged numerical parity, arbitrary-image robustness, OCR superiority, or native Qwen chain-of-thought. Adaptive content is a short task-controlled field. The included invoice demo is an end-to-end protocol demonstration, not a financial approval system.

License and attribution

Code and release documentation are Apache-2.0. The checkpoint remains subject to the Qwen3-VL base-model license and any upstream terms. Users must review upstream terms before redistribution or commercial deployment.