---
pipeline_tag: image-text-to-text
base_model:
- thinkingmachines/Inkling-Small
base_model_relation: quantized
license: apache-2.0
library_name: transformers
tags:
- AxionML
- ModelOpt
- Inkling
- quantized
- NVFP4
- nvfp4
- sglang
- mirror
---

# AxionML Inkling-Small-NVFP4

> Mirrored by [AxionML](https://huggingface.co/AxionML) for open-source serving and deployment use cases. Part of AxionML's effort to provide ready-to-serve quantized models for the community.
>
> **Quantized by Thinking Machines.** The weights in this repository are an unmodified copy of [thinkingmachines/Inkling-Small-NVFP4](https://huggingface.co/thinkingmachines/Inkling-Small-NVFP4) (revision `b6a99534467840620d411e4cd4ad5819b2610d9c`). All credit for the quantization belongs to Thinking Machines.

This is an NVFP4-quantized version of [thinkingmachines/Inkling-Small](https://huggingface.co/thinkingmachines/Inkling-Small) (276B total parameters, 12B activated; text, image and audio input), quantized with [NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer).

**About NVFP4 quantization:** NVFP4 on Blackwell couples a compact E2M1 FP4 codebook with blockwise FP8 (E4M3) scaling over 16-element micro-blocks, so that 4-bit stored values remain numerically useful for neural-network computation. The E2M1 codebook provides a small, nonuniform set of representable magnitudes up to ±6 and relies on saturating behavior rather than IEEE NaN/Inf encodings to maximize usable range per bit. Using an FP8 block scale (rather than power-of-two-only E8M0) enables fractional scales and error-minimizing scale selection. On Blackwell Tensor Cores, native FP4 multipliers exploit E2M1 simplicity while higher-precision FP32 accumulation protects dot-product accuracy.

> **Ready for commercial and non-commercial use under [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)**, subject to Thinking Machines' [Acceptable Use Policy](https://thinkingmachines.ai/model-acceptable-use-policy).

## Model Summary

| | |
|:---|:---|
| **Architecture** | 42-layer decoder-only multimodal MoE; hybrid local/global attention |
| **Total Parameters** | 276B |
| **Activated Parameters** | 12B |
| **Experts** | 256 routed (6 active) + 2 shared |
| **Input** | Text, image, audio (16 kHz WAV) |
| **Checkpoint Size** | ~171 GB |

## Evaluation Results

| Benchmark | Inkling-Small |
|:---|:---:|
| SWE-bench Verified | 80.2 |
| SWE-bench Pro (Public) | 55.9 |
| Terminal-Bench 2.1 (best harness) | 64.7 |
| SciCode | 48.7 |
| GDPval-AA v2 | 1269 |
| MCP Atlas (public) | 79.6 |
| BrowseComp (w/ context) | 77.4 |
| Toolathlon Verified | 54.4 |
| GPQA Diamond | 89.5 |
| HLE (text only) | 31.6 |
| HLE (with tools) | 47.8 |

> Scores are from the [Inkling-Small model card](https://huggingface.co/thinkingmachines/Inkling-Small) (full-precision baseline).

## Quantization Details

- **Quantization format:** NVFP4 (group size 16), Thinking Machines' official NVFP4 release; attention, routers, vision/audio encoders and other excluded modules stay in BF16
- **KV cache:** not quantized

## Usage

### Deploy with SGLang

```bash
python3 -m sglang.launch_server \
    --model-path AxionML/Inkling-Small-NVFP4 \
    --tp 4 \
    --trust-remote-code
```

### Deploy with vLLM

```bash
vllm serve AxionML/Inkling-Small-NVFP4 \
    --tensor-parallel-size 4 \
    --trust-remote-code
```

Minimal commands; for tuned serving flags follow the official recipes: [SGLang](https://docs.sglang.io/cookbook/autoregressive/ThinkingMachines/Inkling-Small), [vLLM](https://recipes.vllm.ai/thinkingmachines/Inkling-Small), [TokenSpeed](https://lightseek.org/tokenspeed/recipes/models#Inkling).

## Limitations

The base model was trained on data that may contain toxic language and societal biases. The quantized model inherits these limitations. It may generate inaccurate, biased, or offensive content. Please refer to the [original model card](https://huggingface.co/thinkingmachines/Inkling-Small) and the [upstream quantized model card](https://huggingface.co/thinkingmachines/Inkling-Small-NVFP4) for full details.

## Credits

- **Base model:** [thinkingmachines/Inkling-Small](https://huggingface.co/thinkingmachines/Inkling-Small)
- **Quantization:** [thinkingmachines/Inkling-Small-NVFP4](https://huggingface.co/thinkingmachines/Inkling-Small-NVFP4) by Thinking Machines
- **Mirror:** [AxionML](https://huggingface.co/AxionML)