autotrust/GEV-26B-Decide-NVFP4 — new model trending #30 on Hugging Face

quantization from `config.json` / `hf_quant_config.json` and logs `modelopt_fp4` and the selected NVFP4 MoE kernel.

| GPU | how NVFP4 runs |

|---|---|

| B200 / B300 / GB200 (SM100) | native W4A4: FlashInfer TRT-LLM NVFP4 MoE kernels (tested, B200) |

| RTX PRO 6000, RTX 5090, DGX Spark (SM12x) | native W4A4: CUTLASS / FlashInfer CUTLASS FP4 MoE kernels (not tested here) |

| H100 / H200 / A100 (SM80–90) | Marlin W4A16: weights in FP4, activations bf16 (not tested here) |

* **Memory.** The weights take 17.1 GiB, so the whole model fits on a 24 GB GPU with room left for a short context, and

on a 32 GB RTX 5090 or a 48 GB card with long contexts and many concurrent requests. Set `--gpu-memory-utilization`

and `MAX_MODEL_LEN` to fit the card, e.g. `MAX_MODEL_LEN=32768`.

* **Same patch as the bf16 model.** The one-engine setup needs `patches/vllm-gemma4-lm-head-lora.patch` (LoRA on

Gemma-4's tied `lm_head`); tested with a vLLM development build from September 2026.

* **System 1 prompt.** Gemma-4 reads System 1 prompts after `<bos>`; `serve_decide.py` adds it. If you call

`/v1/completions` yourself, start the prompt with `<bos>` and pass `top_k: 0` and `top_p: 1.0`.

* **transformers.** The `transformers` + `peft` path below loads bf16 weights; use the bf16 repository

[autotrust/GEV-26B-Decide](https://huggingface.co/autotrust/GEV-26B-Decide) for it.

## Decision Index

| | Decision Index 0.2.1 (balanced skill) | balanced raw | breadth skill |

|---|---:|---:|---:|

| **autotrust/GEV-26B-Decide, adaptive thinking** | **62.48** | 70.66 | 62.00 |

| TypeSafe Jev 1.13 (board) | 57.91 | — | — |

| area (skill) | Knowledge & Reasoning | Language | Retrieval & Classification | Tools & Automation | Arts & Taste |

|---|---:|---:|---:|---:|---:|

| GEV-26B-Decide, adaptive thinking | **0.602** | 0.636 | 0.679 | 0.697 | 0.415 |

How the score was computed (our scoring with the kit's `score --edition 0.2.1`, not a board entry):

* **Knowledge & Reasoning:** all ten benchmarks with adaptive thinking (per-benchmark table under [Adaptive thinking](#adaptive-thinking)).

* **Other four areas:** System 1 only (thinking off), from the complete System 1 run of these weights (all 150,759

requests, 0 errors), results in

[`autotrust/jev-decision-index-results`](https://huggingface.co/datasets/autotrust/jev-decision-index-results)

(`runs/jev-gemma4-26b-a4b`, the weights' previous name). Thinking was tried on five of their benchmarks and is not used

there: it adds little to classification, retrieval and tool selection (see

[Outside Knowledge & Reasoning](#outside-knowledge--reasoning)).

* The board requires a median latency of at most 1,000 ms per request, measured on an RTX PRO 6000. System 1 answers in

about 45 ms on a B200; adaptive thinking is far slower on the Knowledge & Reasoning benchmarks (see

[Latency](#latency)).

Details: `reports/decision_index_adaptive.json`, `reports/adaptive_latency_summary.json`.

## New (3 October 2026): computer use and robot arm, 60–85 ms per step

The fastest vision model of the family. Every step below is **one System 1 decision** (thinking off): a screenshot or a

camera image in, a probability for every action out, in a single forward pass on one B200.

**Computer use: screenshot → which element to click.** A real browser (headless Chromium). Every clickable element gets a

numbered box; System 1 picks the next click (or "the task is complete"), the browser clicks it, and the loop repeats.

<video src="https://huggingface.co/autotrust/GEV-26B-Decide/resolve/main/videos/computer_use_shop.mp4" controls autoplay loop muted playsinline width="100%"></video>

**95% of 60 random multi-step tasks completed** (shop, settings, mail; 3–7 clicks each) in **about 85 ms per click**:

the same success rate as JEV-27B-VL, 3× faster. The colour swatches carry no text, so that click is decided from the

screenshot alone.

**Robot arm: pick and place from a camera image.** At every step System 1 looks at the top camera image and answers two

questions: is the target left or right of the gripper, and above or below it? The arm moves accordingly and halves its

step whenever an answer flips. It grasps the cube, carries it and drops it in the tray (MuJoCo simulation).

<video src="https://huggingface.co/autotrust/GEV-26B-Decide/resolve/main/videos/robot_arm_pick_place.mp4" controls autoplay loop muted playsinline width="100%"></video>

61 ms per decision, so a whole pick and place takes 4–8 seconds of model time. It completed 40% of 20 random scenes:

close to the target its left/right answers are less precise than JEV-27B-VL's, so more grasps miss. Once grasped, 8 of 9

cubes ended in the tray.

Same scenes and tasks for every model in the family:

| | **GEV-26B-Decide** | [JEV-27B-VL](https://huggingface.co/autotrust/JEV-27B-VL) | [JEV-9B](https://huggingface.co/autotrust/JEV-9B) |

|---|---:|---:|---:|

| computer use: numbered boxes + element text (60 tasks) | 95% | 95% | 95% |

| time per click | **≈ 85 ms** | ≈ 260 ms | ≈ 200 ms |

| robot arm: pick and place (20 scenes) | 40% | **75%** | 50% |

| time per robot-arm decision | **61 ms** | 239 ms | 163 ms |

* For computer use, give it the element text (as an accessibility tree would). With the numbered boxes alone it

completes 15% and often declares the task complete too early.

* Use System 1 for simple visual questions inside a control loop. Asked to pick one of 8 motor commands directly, it

completed 0 of 10 scenes.

Demo code: [JEV-9B `vl/demos/`](https://huggingface.co/autotrust/JEV-9B/tree/main/vl/demos) (set `JEV_URL` to this

server). Per-episode results: [`reports/demos/`](reports/demos).

## Overview

**GEV-26B-Decide answers typed questions with a calibrated probability for every option; with thinking switched on, it

thinks only when it needs to.** System 1 decides in one forward pass (about 45 ms). When its leading option is uncertain, System 2 (the same

backbone in Gemma-4 thinking mode) reasons over the question, and the reasoning is folded into the final probabilities.

One set of weights, one vLLM engine, for text and images.

| | what it does | output |

|---|---|---|

| **System 1** | typed decisions: yes/no · pick one of 2–256 options · rate 0–5, over text and images; prompts up to 256K tokens | a calibrated probability for every option, in one forward pass |

| **Adaptive thinking** (opt-in: `thinking: "auto"`) | System 1 first; below 0.8 confidence, System 2 thinks and its answer is folded in | calibrated probabilities |

| **System 2** | the unmodified `google/gemma-4-26B-A4B-it`, optionally thinking step by step, text and images | text / reasoning |

GEV-26B-Decide was previously published as `autotrust/JEV-Gemma4-26B-A4B`; the weights are the same.

> **Two models, two organisations.** **TypeSafe Jev 1.13** is the hosted, closed model made by TypeSafe AI.

> **autotrust/GEV-26B-Decide** is an independent open-weights model built by AutoTrust AI; it is not affiliated with,

> endorsed by, or a product of TypeSafe AI.

## Watch it think

Each puzzle has one correct answer. System 1 answers in one pass; with `thinking: "auto"`, System 2 thinks when System 1

is below 0.8 confidence, and its answer is folded into the probabilities. The videos show puzzles that System 1 got wrong;

the tables below count all puzzles.

**Minesweeper.** Which hidden cell is certainly safe? System 1 is at chance (21.0 % against 25 %); with thinking, 86.0 %.

These thoughts usually reach the 8,192-token budget, and the answer read at that point is still right most of the time.

<video src="https://huggingface.co/autotrust/GEV-26B-Decide/resolve/main/videos/think_minesweeper.mp4" controls autoplay loop muted playsinline width="100%"></video>

**Connect Four.** Which column wins now, or stops the opponent from winning next move? 54.0 % → 99.5 %.

<video src="https://huggingface.co/autotrust/GEV-26B-Decide/resolve/main/videos/think_connect4.mp4" controls autoplay loop muted playsinline width="100%"></video>

**Wordle.** Which word still fits all the colour feedback? 52.0 % → 100 %.

<video src="https://huggingface.co/autotrust/GEV-26B-Decide/resolve/main/videos/think_wordle.mp4" controls autoplay loop muted playsinline width="100%"></video>

**Sudoku.** Which digit belongs in the highlighted cell? 76.0 % → 99.5 %.

<video src="https://huggingface.co/autotrust/GEV-26B-Decide/resolve/main/videos/think_sudoku.mp4" controls autoplay loop muted playsinline width="100%"></video>

Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/autotrust/GEV-26B-Decide-NVFP4