---
base_model: browser-use/bu-30b-a3b-preview
base_model_relation: quantized
license: other
license_name: modified-mit-browser-use
license_link: https://huggingface.co/browser-use/bu-30b-a3b-preview/blob/main/LICENSE
tags:
  - nvfp4
  - awq
  - modelopt
  - browser-use
  - agent
  - vision-language
  - moe
  - quantized
pipeline_tag: image-text-to-text
library_name: transformers
---

# bu-30b-a3b-preview NVFP4-AWQ (LITE)

A 4-bit NVFP4 + AWQ-lite quantization of
[browser-use/bu-30b-a3b-preview](https://huggingface.co/browser-use/bu-30b-a3b-preview) — the 30B Qwen3-VL-MoE browser-agent model — produced with
[NVIDIA TensorRT-Model-Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer)
v0.43.

**What's notable about this quant**

This is (as of upload) the first **NVFP4_AWQ** quantization of any browser-agent VLM on the Hub, and the first NVFP4 quant of this model with documented calibration provenance. Existing NVFP4 / INT4-AWQ quants of `bu-30b-a3b-preview` either lack calibration data disclosure or calibrate against generic text corpora; this one was calibrated **on-distribution**, using 602 real multimodal browser-use trajectories generated by the full-precision model itself.

The calibration-data argument is the load-bearing claim of this quant — it's documented in detail below.

## Why NVFP4 for this model

- **Native acceleration on Blackwell.** RTX 5090, PRO 6000, B100/B200, GB10 all have native FP4 tensor cores (sm_100+). On Blackwell-class hardware NVFP4 weights execute at ~2× the throughput of FP8.
- **Memory.** ~17 GB vs ~58 GB at BF16. Fits comfortably on a single RTX 5090 (32 GB) with headroom for the 32K-token context window.
- **Accuracy-preserving 4-bit format.** NVFP4's two-level scales (FP8 E4M3 block scales at block size 16, plus FP32 per-tensor scale) substantially outperform naive INT4 in accuracy, and AWQ's activation-aware per-channel scaling protects the weight channels that matter most.

## Quantization Recipe

**Base config**: `NVFP4_AWQ_LITE_CFG` from `modelopt.torch.quantization.config`.

**Module-scoped exclusions (kept at BF16 precision)**:

| Module pattern | Reason |
|---|---|
| `*visual*` | Vision encoder (ViT tower) is small relative to MoE decoder; disproportionate accuracy loss for minimal memory savings. Standard practice. |
| `*mlp.gate.*` | MoE router — tiny logit perturbations cascade into expert misrouting. Already excluded in `NVFP4_AWQ_LITE_CFG`. |
| `*lm_head*` | Output projection. Already excluded. |
| `*router*`, `*block_sparse_moe.gate*` | Generic router patterns (covers Mixtral-style MoE architectures). Already excluded. |

All 128 MoE experts (`model.language_model.layers.*.mlp.experts.*`) and attention matrices are quantized to NVFP4 weights + NVFP4 activations (W4A4). The `model.visual.*` ViT tower (depth 27, hidden 1152) stays in BF16.

## Calibration Data

**602 samples** of real browser-use agent trajectories:

| Category (BU_Bench V1) | Tasks | Samples | Weight (rationale) |
|---|---|---|---|
| GAIA | 8 | ~200 | Research + reasoning — dominant agent workload |
| OM2W2 | 6 | ~150 | Open-ended info gathering |
| BrowseComp | 5 | ~130 | Cross-source comparison |
| WebBenchREAD | 5 | ~80 | Clean DOM activations |
| InteractionTests | 1 | ~15 | Signal floor for form/interaction regime |

**Collection process:**

1. Full-precision bu-30b-a3b-preview served via vLLM 0.17 at `--dtype bfloat16`.
2. 3 parallel `browser-use` v0.12.6 agents with `enable_planning=True` and `use_vision=True` ran 25 tasks sampled from the official [browser-use/benchmark](https://github.com/browser-use/benchmark) BU_Bench V1 set.
3. Per-category step caps: 40 for GAIA/OM2W2/BrowseComp, 25 for WebBenchREAD/InteractionTests.
4. A proxy between the agents and vLLM captured every `/v1/chat/completions` request payload (including image parts) to JSONL.
5. Samples with total tokens < 1000 (keepalive/error artifacts, 3) or blank screenshots (variance < 150, 16) were filtered out.

**Sample-level statistics** (staged calibration, 602 samples, Qwen3-VL tokenizer + true vision-token expansion):

| Metric | Value |
|---|---|
| Total tokens | min=3, p25=11.2K, median=13.4K, p75=15.8K, p90=18.1K, max=35.4K |
| 8-16K bucket | 439 samples (73%) |
| 16-32K bucket | 144 samples (24%) |
| 32K+ samples | 6 (long-context tail) |
| Samples with screenshot | 93.6% |
| Non-degenerate screenshots | 97.2% |
| DOM element count (median / max) | 136 / 941 |

The calibration distribution was committed to **before** running the analyzer on the exploratory data — weights reflect the target user population (researchers and educators running a local agent), not post-hoc curve-fitting to whatever tasks happened to look interesting.

## Serving

### ⚠ vLLM support

As of **vLLM 0.19.1 / main**, the `ModelOpt` quantization loader does **not** accept `quant_algo: NVFP4_AWQ` — the supported list is only `['FP8', 'FP8_PER_CHANNEL_PER_TOKEN', 'FP8_PB_WO', 'NVFP4', 'MXFP8', 'MIXED_PRECISION']`. Renaming the algo to plain `NVFP4` would load but produce mathematically wrong inference because the 18,480 `pre_quant_scale` tensors that carry AWQ's per-channel activation rescaling would not be applied.

If you want a vLLM-loadable variant, use the sibling repo **[`Code4me2/bu-30b-a3b-preview-NVFP4`](https://huggingface.co/Code4me2/bu-30b-a3b-preview-NVFP4)** (plain NVFP4, no AWQ, slightly lower accuracy but same memory footprint).

### TensorRT-LLM (recommended)

This format is produced by and natively supported by [NVIDIA TensorRT-Model-Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer) + TensorRT-LLM. Build an NVFP4 engine:

```bash
trtllm-build --checkpoint_dir Code4me2/bu-30b-a3b-preview-NVFP4-AWQ \
    --quant_format nvfp4 \
    --max_seq_len 32768
```

See the [TRT-LLM NVFP4 guide](https://nvidia.github.io/TensorRT-LLM/reference/precision.html) for more details.

### SGLang

SGLang's ModelOpt integration supports NVFP4_AWQ when built against the matching ModelOpt version — consult their docs for the current status.

## Intended Use

This model is a drop-in replacement for `bu-30b-a3b-preview` within the
[browser-use](https://github.com/browser-use/browser-use) library. It is
trained/tuned specifically for browser-use's indexed-DOM + structured-action
format. Using it outside that flow (or with a different harness / freeform
CDP scripting) will produce substantially worse results than the
quantization accuracy alone would suggest.

## Evaluation

_Evaluation numbers (MMLU, GSM8K, MM-Bench, BU_Bench V1 subset) will be
added after running against BF16 baseline. See methodology below._

Planned eval suite:
- MMLU (general knowledge, 5-shot)
- GSM8K (math reasoning, 0-shot chain-of-thought)
- MM-Bench (vision-language, 0-shot)
- BU_Bench V1 held-out tasks (agent-specific, using the same browser-use harness)

## Reproduction

- Base model: `browser-use/bu-30b-a3b-preview`
- Quantization tool: `nvidia-modelopt==0.43.0`
- Quantization config: `NVFP4_AWQ_LITE_CFG` with `*visual*` excluded (ViT stays BF16); router (`*mlp.gate.*`) already excluded by the config default
- Calibration samples: 512 / 602 (shuffled, seed=42). 6 samples above 32K tokens skipped (aligned with `--max-model-len`)
- Host: single RTX PRO 6000 Blackwell, 98GB
- Calibration wall time: ~14h (70 min cache activation stats + 12h AWQ scale search + 10 min export)

### ModelOpt patch for Qwen3-VL-MoE support

ModelOpt 0.43 does not natively know how to export quantized checkpoints for `Qwen3VLMoeForConditionalGeneration`. Three patches were required (included in the model repo as `modelopt_patch.py`):

1. `get_expert_linear_names()` in `layer_utils.py` — recognize `Qwen3VLMoe*` and return `[gate_proj, up_proj, down_proj]`
2. `get_experts_list()` in `layer_utils.py` — recognize `qwen3vlmoe*` model_type
3. `_export_transformers_checkpoint()` in `unified_export_hf.py` — wrap the `QuantQwen3VLMoeTextExperts` container with a transparent iterable proxy so the existing iterable dispatch walks the un-BMM'd per-expert `ModuleList`s, while `__call__` and attribute access still delegate to the real experts module for the internal dummy forward pass

Reference code + calibration harness: [GitHub link TBD]

## Attribution & License

Derived from [`browser-use/bu-30b-a3b-preview`](https://huggingface.co/browser-use/bu-30b-a3b-preview), which is distributed under a **Modified MIT License** by Browser Use Inc. with a commercial-use restriction: **use is not permitted for organizations whose annual consolidated revenue exceeds USD 1 million for the preceding month**. That restriction propagates to this derivative. Commercial users above the revenue threshold must obtain a license from Browser Use Inc. (`support@browser-use.com`) or use Browser Use's hosted services.

The original LICENSE file is included alongside the weights.

## Acknowledgements

- **Browser Use** for the base model and the open benchmark suite
- **NVIDIA Model Optimizer** for the NVFP4_AWQ calibration tooling
- **Qwen team** for the Qwen3-VL-MoE architecture