Unsloth Releases Qwen3.6-27B-NVFP4: Enhanced Throughput and Agentic Coding for Developers

What Changed

Unsloth has introduced unsloth/Qwen3.6-27B-NVFP4, a new NVFP4 quantized checkpoint of the Qwen3.6-27B model. This release emphasizes performance enhancements, particularly a reported 2.5x faster throughput compared to other NVFP4 quantizations. The model is calibrated on a mixture of Unsloth's proprietary dataset and the UltraChat dataset. It is designed to operate on GPUs with 24GB of VRAM.

Beyond performance, Qwen3.6-27B-NVFP4 incorporates substantial upgrades in its core capabilities. Key among these are improved Agentic Coding, enabling the model to handle frontend workflows and repository-level reasoning with greater fluency. Additionally, a new feature called Thinking Preservation allows the model to retain reasoning context from historical messages, aiming to streamline iterative development and reduce overhead. This release follows the Qwen3.5 series and represents the first open-weight variant of Qwen3.6, built on community feedback to prioritize stability and practical utility.

Technical Details

The unsloth/Qwen3.6-27B-NVFP4 model is a Causal Language Model with a Vision Encoder. It has 27 billion parameters, a hidden dimension of 5120, and 64 layers. The architecture includes Gated DeltaNet and Gated Attention mechanisms. The Gated DeltaNet features 48 linear attention heads for V and 16 for QK, each with a head dimension of 128. The Gated Attention has 24 attention heads for Q and 4 for KV, with a head dimension of 256. Rotary Position Embedding is used with a dimension of 64. The Feed Forward Network has an intermediate dimension of 17408.

Crucially, this checkpoint includes a Multi-Token Prediction (MTP) module, allowing it to act as its own speculative draft for faster decoding. The model supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens. Unsloth provides specific instructions for running the model with vLLM, recommending vllm>=0.25.0, flashinfer-python>=0.6.13, and nvidia-cutlass-dsl>=4.5.2 for optimal performance. They explicitly advise against using the Marlin backend due to its slower inference speed, advocating for native vLLM, cute-DSL, CUTLASS, or flashinfer_trtllm backends. For DGX Spark environments, specific environment variables (export CUTE_DSL_ARCH=sm_121a) and flashinfer_b12x moe-backend are required to prevent 2x slower inference.

Benchmark Analysis

Unsloth conducted NVFP4 accuracy benchmarks across MMLU-Pro, AIME 2025, and GPQA, comparing their NVFP4 quantization against NVIDIA NVFP4, FP8, and BF16. The results indicate that Unsloth NVFP4 maintains comparable accuracy:

Throughput benchmarks demonstrate the performance advantage of Unsloth's approach. When serving the 27B model, Unsloth's cute-DSL (auto) backend achieved 6,863 throughput tokens/second, significantly outperforming NVIDIA's marlin (auto) backend at 2,403 throughput tokens/second. Similarly, for a 35B-A3B model, Unsloth's cute-DSL + trtllm (auto) reached 15,636 throughput tokens/second, compared to NVIDIA's marlin (auto) at 8,721 throughput tokens/second.

Original Qwen3.6 BF16 reference benchmarks highlight the model's capabilities across various domains. In Language tasks, Qwen3.6-27B achieved a SWE-bench Verified score of 77.2, SWE-bench Pro at 53.5, and Terminal-Bench 2.0 at 59.3. For Knowledge, it scored 86.2 on MMLU-Pro and 93.5 on MMLU-Redux. In STEM & Reasoning, GPQA Diamond reached 87.8 and AIME26 scored 94.1.

For Vision Language tasks, Qwen3.6-27B demonstrated strong performance: MMMU at 82.9, MathVista mini at 87.4, and RealWorldQA at 84.1. Document Understanding scores included CharXiv RQ at 78.4 and OCRBench at 89.4. Spatial Intelligence benchmarks showed ERQA at 62.5 and CountBench at 97.8. Video Understanding metrics included VideoMME(w sub.) at 87.7 and VideoMMMU at 84.4. The Visual Agent benchmark V* achieved 94.7, and AndroidWorld scored 70.3.

Developer Implications

Developers can leverage unsloth/Qwen3.6-27B-NVFP4 for applications requiring high-throughput text generation and advanced agentic capabilities. The model's compatibility with popular inference frameworks like vLLM, SGLang, KTransformers, and Hugging Face Transformers simplifies integration into existing workflows. The provided vLLM run instructions, including speculative decoding configurations, offer clear guidance for deployment.

The enhanced agentic coding features, particularly for frontend and repository-level tasks, suggest potential for automating complex development processes or building more sophisticated AI assistants for coding. The