Nvidia’s New OpenClaw‑Optimized Model Cracks Top‑5 on PinchBench – Free to Use

PinchBench Ranking and OpenClaw Compatibility

Nemotron‑3‑Super (120B total, 12B active) achieved an 85.6% success rate on PinchBench, placing it in the top five and making it the only open‑source model among flagship closed‑source models such as GPT‑5.4 and Claude Opus 4.5.

Task‑Level Success Rates

Basic

, Calendar, Coding, File Ops: 100% Data Analysis: 98% Research: 90% Comprehension: 91% Organization: 89% Creativity: 18% Memory: 0% Context: 70%

The model excels at file‑read/write, script generation, and multi‑step workflow execution, but it lacks long‑context memory and creative generation.

PinchBench Evaluation Criteria

File read/write operations

Code modification and refactoring

Tool calling and API interaction

Multi‑step complex tasks

Self‑repair after errors

These capabilities align directly with the requirements of AI coding agents such as OpenClaw.

Hardware and Inference Parameters

Total parameters: 120B

Activated parameters: 12B

Architecture: LatentMoE (Mamba‑2 + MoE + Attention)

Context window: 1 M tokens

Minimum GPU: 1× B200‑80GB or 1× DGX Spark

Inference mode: supports enable_thinking=True/False Quantization: NVFP4 (training‑time FP4 precision)

Architecture Details

Mamba‑2 (state‑space model) : linear‑complexity handling of long sequences, enabling the 1 M token context.

LatentMoE : activates only 12 B of the 120 B parameters per token via a low‑dimensional latent routing space, improving precision while reducing compute.

Attention layers : retained at critical positions to preserve essential information.

Multi‑Token Prediction (MTP) : predicts multiple future tokens during training, allowing speculative decoding and faster inference.

NVFP4 Quantization Benchmarks

Benchmark          BF16   FP8   NVFP4
MMLU‑Pro           83.73  83.63 83.33
HMMT Feb25 (w/ tools) 94.73 94.38 95.36
GPQA (no tools)    79.23  79.36 79.42
LiveCodeBench v6   78.69  78.44 78.44
IFBench            72.58  72.32 73.30
Arena‑Hard‑V2      73.88  76.06 76.00
RULER‑500 @128k    96.79  96.85 95.99

On HMMT, GPQA, and IFBench the NVFP4 version matches or exceeds the BF16 baseline, demonstrating that training‑time low‑precision quantization retains accuracy while reducing memory usage.

Training Methodology (Fully Open‑Source)

Pre‑training data: >25 T tokens, fully public (Nemotron Pre‑Training Datasets).

Post‑training data: SFT + RL datasets, fully public (Nemotron Post‑Training v3).

Training scripts: available on GitHub.

Evaluation: NeMo Evaluator SDK reproduces all benchmark results.

RL environment: NeMo Gym with asynchronous GRPO multi‑environment reinforcement learning.

Training proceeds in three stages: pre‑training → SFT (synthetic code, tool use, instruction following) → RL (math, code, science, tool usage across multiple environments).

Local Deployment Example (vLLM)

# vLLM deployment
vllm serve $MODEL_CKPT \
  --async-scheduling \
  --served-model-name nvidia/nemotron-3-super \
  --dtype auto \
  --kv-cache-dtype fp8 \
  --tensor-parallel-size 1 \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin "./super_v3_reasoning_parser.py" \
  --reasoning-parser super_v3

Recommended inference parameters: temperature=1.0, top_p=0.95. The service exposes an OpenAI‑compatible endpoint for agents such as OpenCode.

Practical Limitations

The minimum hardware (B200‑80GB or DGX Spark) exceeds typical consumer GPUs; a 4090 cannot run the model. For most developers, API access is more realistic.

While the 85.6% PinchBench score is strong, real‑world projects may introduce additional complexities (specific language frameworks, long‑running multi‑turn dialogs, stability under diverse workloads) that require empirical verification.

Emerging Open‑Source Agent Models

Qwen‑3.5‑122B‑A10B adopts a similar MoE‑based hybrid architecture (122 B total, 10 B active), indicating a broader shift toward high‑capacity models with limited active parameters for efficient agent backbones.

HuggingFace model page (full deployment guide): https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4