blink-mimo-9b

blink-mimo-9b

Pick the next click from a screenshot. blink-mimo-9b reads it with its own vision tower and scores the offered targets as a typed choice, returning probabilities instead of generated text. It also takes text or JSON state with choice, noul (yes/no), or score questions. Each batch is one forward pass; large requests may need several batches.

Try it: Screen click · Space · Computer use · API · Docs · GitHub · On a Mac (MLX) · Customize · blink-4b · blink-27b

At a glance

Computer use

Screen click uses this model to score numbered targets in a screenshot as a typed choice.

How it reads a screenshot

MiMo's own vision encoder supplies image tokens; blink scores the offered clicks in a forward pass without generating text.

<details><summary>From pixels to probabilities</summary>

• A screenshot with numbered boxes goes through blink-mimo-9b's own vision encoder from XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. It has 27 encoder blocks, hidden 1152, and 333 vision tensors unchanged by training.
• The encoder turns the screenshot into image tokens.
• The language model reads the image tokens plus the typed choice question and offered targets.
• blink reads next-token logits for the offered options and applies an FP32 softmax. No text is generated.
• Enable self-hosted screenshot input with --vision.

</details>

Computer use in ten apps

blink-mimo-9b finished 133/230 tasks (58%) zero-shot. It's the model behind Screen click, which runs live in the Space. It uses its own vision encoder with --vision, not a borrowed tower. In the "Teach it your app" v1 experiment, 150 labelled screens lifted success on practised apps from 96/160 to 132/160 (60% to 83%). But it paused on only 35/76 truly risky clicks there, down from 62/62 before training. Apps it never saw went from 35/60 to 49/60. Screenshot training is still an experiment, not in the public trainer. These are original apps built for this showcase, not a public benchmark. Watch the runs · Videos and traces · Code.

Screenshots (opt-in, self-hosted)

Self-hosted image input is off by default: start serve.py with --vision to use MiMo's unchanged tower (image mode needs torchvision==0.28.0). TypeSafe's hosted Jev is text-only; screenshots are a blink self-hosted extension, not a hosted Jev feature.

Send a data:image/png;base64,... URI inside a string in state, or data URIs in a top-level images list, with a choice question whose criteria name the clickable targets. JPEG and WebP data URIs also work. Image URLs are never fetched.

These are development readouts, not a benchmark. With the default first image layout and five offered targets, target accuracy was:

Target accuracy is not completed-task success. The default layout was picked on previously seen development data. Try catalog, directory, lookup, or done; Next click (text) shows the page-elements-as-text path.

For text-only browser agents using page elements as text, blink-27b did best in local runs; see browser-agent setup. These local runs are not a benchmark.

Quickstart

This example uses text. For screenshots, use the opt-in server path above.

# pip install "torch==2.13.0" "transformers==5.17.0" "flash-linear-attention==0.5.2" "accelerate>=1.1.0" safetensors huggingface_hub
import os, sys
from huggingface_hub import hf_hub_download

os.environ["BLINK_MODEL"] = "thegovind/blink-mimo-9b"
os.environ["BLINK_REVISION"] = "v1.4"
sys.path.insert(0, os.path.dirname(hf_hub_download("thegovind/blink-mimo-9b", "blink.py", revision="v1.4")))
import blink

out = blink.decide(
    "Order #4411 arrived with a cracked screen. I want my money back, not another one.",
    {
        "intent": {
            "type": "choice",
            "instructions": "What does the customer want?",
            "criteria": {"refund": "Money back", "replacement": "A new unit", "info": "Information only"},
        },
        "urgent": {"type": "noul", "instructions": "Does this need a reply today?"},
        "anger": {"type": "score", "instructions": "How upset is the customer?", "criteria": ["calm", "annoyed", "angry"]},
    },
)
print(out["answers"]["intent"]["probabilities"])

Run it as a server

serve.py implements TypeSafe's text decision API at POST /v1/systemone and GET /v1/models. Point server-side Python or JavaScript SDKs at it with TYPESAFE_BASE_URL. It processes requests one at a time by default; use --batch-window-ms 5 for cross-request batching. GET /healthz reports readiness. The command below starts text-only; pass --vision for screenshots.

pip install "torch==2.13.0" "transformers==5.17.0" "flash-linear-attention==0.5.2" "accelerate>=1.1.0" safetensors huggingface_hub
hf download thegovind/blink-mimo-9b --revision v1.4 --local-dir blink-mimo-9b
python blink-mimo-9b/serve.py --model ./blink-mimo-9b --port 8000
# TypeSafe SDKs: export TYPESAFE_BASE_URL=http://127.0.0.1:8000 TYPESAFE_API_KEY=any

Or use Docker from the downloaded folder:

cd blink-mimo-9b
docker build -t blink-mimo-9b . && docker run --rm --gpus all -p 127.0.0.1:8000:8000 blink-mimo-9b

Earlier vLLM checks passed, but the final read matched earlier answers and still failed hard ECE (0.122 vs 0.121). No MiMo opt-in server ships; use serve.py.

Run it on a Mac (MLX)

blink_mlx.py keeps this repo's blink.py for prompts, option labels and answers, and runs the forward pass with MLX on Apple silicon. Text and JSON state only, bf16 weights only; it is not a server. It loads only the text side of this checkpoint.

pip install "mlx-lm>=0.31.3"
curl -O https://raw.githubusercontent.com/thegovind/blink/main/examples/mlx/blink_mlx.py
python blink_mlx.py --model thegovind/blink-mimo-9b

--check compares your Mac's answers with the Space's saved runs. The MLX guide covers memory and limits.

Customize it

Fine-tune blink-mimo-9b on your own decisions with a LoRA adapter. Check and split your rows with examples/train/prepare.py, mix in anchor rows so it keeps its general skills, train with the PyTorch trainer on a GPU, locally or in the cloud, or with examples/mlx/train_mlx.py on a Mac (checked against the PyTorch trainer on a small model, not yet timed on a Mac), and compare accuracy and calibration before and after. The trainers train its text side. jevlab.merge writes that side alone; jevlab.graft_vlm puts the trained weights back into the full checkpoint with the vision tower unchanged. Screenshot training is not published yet. The customize guide has the steps, and examples/rl covers learning from outcomes and RL.

Results

No official JevBench score for blink has been published. The JevBench numbers are public-item development proxies, not official scores or claims of rank or parity. Decision Index is a descriptive local run of the official kit, not a leaderboard submission; public training exposure affects its scores.

<details><summary>Model details: architecture, training, data</summary>

Architecture and readout

MiMo's text side has 32 decoder layers (24 Gated DeltaNet, 8 full-attention), hidden 4096, untied embeddings, and 8.95B text parameters. The tower has 27 encoder blocks, hidden 1152; all 333 vision tensors unchanged. Fine-tuning merged 43.3M LoRA parameters at rank 16, alpha 32; 179 other language-model tensors stayed unchanged. Targets: q_proj, k_proj, v_proj, o_proj; in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj; gate_proj, up_proj, down_proj.

Readout uses FP32 softmax on next-token logits for offered labels only. Probabilities are conditional on the options, not certified chances of success.

Training

One supervised epoch: 123,195 question rows, lr 5e-5, 615 steps. Cross-entropy trained on target distributions; no RL or preference optimization.

Data sources and licences

Source-repository licences do not settle rights in underlying texts.

Evaluation and limits

The DI-S selection sample scored 55.3; the archived Decision Index 0.1 local run scored 56.53. No MMLU-Pro or GPQA items were direct training sources, though SuperGPQA rows matched some added-request text. Public train splits overlap Decision Index, and semantic or pretraining overlap cannot be ruled out. The v1.2 blink.py and serve.py used its text side only. The screenshot readouts above are development measurements, not final held-out image evaluations.

Use --image-layout first|inline to place image tokens (first is the default), or --model-name blink-mimo-9b for a renamed folder. Neither flag switches on images. Text limits are 255 options per choice and 2–10 score levels; invalid images or over-limit requests return 422. English-centric; does not chat or explain answers, and long policies and date/number reasoning remain weak spots.

</details>

License

Code: Apache-2.0. Weights: non-commercial research and evaluation only; see each model card's license.

See LICENSE.md for the weight terms. The MiMo model card declares MIT without a separate upstream licence file (LICENSE-MiMo.md); its Qwen base is Apache-2.0 (LICENSE-Qwen).