Mapika/Decider Vision 2B

• License: Apache-2.0
• Release date: Not documented for the current weights; 2026-09-16 is the first HF publication (repository creation), not a verified date for this revision.
• Modalities: Input: text, image; Output: typed decisions (choice, score, noul)
• Original model: Mapika/decider-2b-vision

Description

Decider Vision combines Decider 2B v5 language weights with the Qwen3.5-2B vision-language model. It scores supplied options for questions about photos, diagrams or game frames in one forward pass, without generating captions or chat responses. Training combines visual multiple-choice tasks, game frames and text replay, followed by PPO on Pong and Breakout.

Run

self serve decider-vision:2b

Here is a link to the API docs: Decision API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/systemone \
	-H 'Content-Type: application/json' \
	-d '{
		"state": "Inspect this photograph.",
		"images": ["https://upload.wikimedia.org/wikipedia/commons/3/3a/Cat03.jpg"],
		"questions": {
			"animal": {
				"type": "choice",
				"instructions": "Which animal is shown?",
				"criteria": {"cat": "A cat", "dog": "A dog"}
			}
		}
	}'

The server returns a JSON decision for the animal question.

Engine Parameters

Served by the llama.cpp-based ggmlc-custom-decider engine and adapter. These are startup settings, not per-request API fields.

Parameter	Default	What it controls
context_size	8192	Context tokens including prompt and image tokens; accepts integers from 512 to 262144.
threads	llama.cpp default	CPU workers; accepts integers from 1 to 1024.
temperature	1.0	Softmax temperature on option-letter logits, not text sampling; accepts 0.01 to 100.
gpu_layers	-1 (all); 0 on CPU devices	Layers offloaded to the GPU; accepts integers from -1 to 10000.
flash_attn	auto	Flash attention mode: auto, on, or off.
image_min_tokens	Unset (projector metadata)	Minimum tokens per image; accepts integers from 1 to 65536.
image_max_tokens	Unset (projector metadata)	Maximum tokens per image; accepts integers from 1 to 65536.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry.

q4 (default)

File	Role	Size	SHA-256	Origin
decider-2b-vision.Q4_K_M.gguf	model	1.19 GiB (1274397056 bytes)	b3e38aa90856d6b93df0d2b186e941e57fa6fd4b022a3a89fe538ff01ebd9985	mradermacher/decider-2b-vision-GGUF
decider-2b-vision.mmproj-Q8_0.gguf	mmproj	347.8 MiB (364664448 bytes)	79e636085e90f1b4bdc5df22003cde0dd0d88086a04b7c101e6bd7d1cc9f2c91	mradermacher/decider-2b-vision-GGUF

Total download: 1.53 GiB (1639061504 bytes).

q8

File	Role	Size	SHA-256	Origin
decider-2b-vision.Q8_0.gguf	model	1.87 GiB (2012012928 bytes)	11e07e4c6a56856e05ec5b4c98308a8e869607bfaad9e70b94c9a6439c595b9c	mradermacher/decider-2b-vision-GGUF
decider-2b-vision.mmproj-Q8_0.gguf	mmproj	347.8 MiB (364664448 bytes)	79e636085e90f1b4bdc5df22003cde0dd0d88086a04b7c101e6bd7d1cc9f2c91	mradermacher/decider-2b-vision-GGUF

Total download: 2.21 GiB (2376677376 bytes).

Notes

• Up to 10 options per question; one image per request through images. Both weights and the Qwen-VL projector are required.
• Images must be public HTTPS URLs or JPEG, PNG, or WebP data URIs by default; HTTP and private addresses require explicit server opt-ins.
• English only; not a chat model or captioner. Text-only behavior comes from Decider 2B v5, not newer text releases.
• Upstream visual calibration and game results cover specific datasets, not your images; validate confidence on your own inputs.
• Download size is not the total runtime memory requirement.
• Origin links use main because these registry entries pin hashes, not repository revisions.

Benchmarks

No submitted benchmark reports yet.