Developed by AxionML for open-source serving and deployment use cases. Part of AxionML's effort to provide ready-to-serve quantized models for the community.
This is an NVFP4-quantized version of Qwen/Qwen3.5-9B (9B parameters), quantized using NVIDIA TensorRT Model Optimizer. Weights and activations of linear layers are quantized to FP4, reducing disk size and GPU memory by ~4x compared to BF16.
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 strategies such as dual-pass evaluation comparing "map max to 6" versus "map max to 4 with clipping." On Blackwell Tensor Cores, native FP4 multipliers exploit E2M1 simplicity to reduce multiplier area while higher-precision FP32 accumulation protects dot-product accuracy.
Ready for commercial and non-commercial use under Apache 2.0.
Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.
Qwen3.5 Highlights
Qwen3.5 features the following enhancement:
Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.
Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.
Scalable RL Generalization: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.
Global Linguistic Coverage: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.
Next-Generation Training Infrastructure: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.
Benchmark Results
For more details, please refer to our blog post Qwen3.5.
Number of Linear Attention Heads: 32 for V and 16 for QK
Head Dimension: 128
Gated Attention:
Number of Attention Heads: 16 for Q and 4 for KV
Head Dimension: 256
Rotary Position Embedding Dimension: 64
Feed Forward Network:
Intermediate Dimension: 12288
LM Output: 248320 (Padded)
MTP: trained with multi-steps
Context Length: 262,144 natively and extensible up to 1,010,000 tokens.
Benchmark Results
Language
Qwen3.5-9B
Qwen3.5-9B-NVFP4
Knowledge & STEM
MMLU-Pro
82.5
81.2
MMLU-Redux
91.1
89.3
C-Eval
88.2
86.0
SuperGPQA
58.2
57.6
GPQA Diamond
81.7
79.4
Instruction Following
IFEval
91.5
89.2
IFBench
64.5
63.4
MultiChallenge
54.5
53.1
Long Context
AA-LCR
63.0
62.2
LongBench v2
55.2
54.1
Reasoning & Coding
HMMT Feb 25
83.2
81.8
HMMT Nov 25
82.9
81.4
LiveCodeBench v6
65.6
64.7
OJBench
29.2
28.8
General Agent
BFCL-V4
66.1
65.0
TAU2-Bench
79.1
77.9
VITA-Bench
29.8
29.4
DeepPlanning
18.0
17.8
Multilingualism
MMMLU
81.2
79.8
MMLU-ProX
76.3
75.2
NOVA-63
55.9
54.9
INCLUDE
75.6
74.1
Global PIQA
83.2
81.7
PolyMATH
57.3
55.9
WMT24++
72.6
69.9
MAXIFE
83.4
80.4
* TAU2-Bench: we follow the official setup except for the airline domain, where all models are evaluated by applying the fixes proposed in the Claude Opus 4.5 system card.
* MMLU-ProX: we report the averaged accuracy on 29 languages.
* WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL.
* MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings).
* Empty cells (--) indicate scores not yet available or not applicable.
Vision Language
Qwen3.5-9B
Qwen3.5-9B-NVFP4
STEM and Puzzle
MMMU
78.4
76.9
MMMU-Pro
70.1
68.8
MathVision
78.9
77.2
Mathvista(mini)
85.7
83.4
We-Math
75.2
72.4
DynaMath
83.6
80.6
ZEROBench
3.0
2.9
ZEROBench_sub
31.1
30.7
VlmsAreBlind
93.7
91.7
BabyVision
28.6/25.8
28.6/25.8
General VQA
RealWorldQA
80.3
77.9
MMStar
79.7
78.8
MMBenchEN-DEV-v1.1
90.1
87.7
SimpleVQA
51.2
49.8
HallusionBench
69.3
67.7
Text Recognition and Document Understanding
OmniDocBench1.5
87.7
86.6
CharXiv(RQ)
73.0
71.6
MMLongBench-Doc
57.7
56.3
CC-OCR
79.3
77.1
AI2D_TEST
90.2
88.7
OCRBench
89.2
86.1
Spatial Intelligence
ERQA
55.5
53.8
CountBench
97.2
95.8
RefCOCO(avg)
89.7
87.5
EmbSpatialBench
83.0
80.5
RefSpatialBench
58.5
56.9
LingoQA
80.4
78.0
Hypersim
13.5
13.2
Nuscene
11.8
11.4
Video Understanding
VideoMME(w sub.)
84.5
82.1
VideoMME(w/o sub.)
78.4
77.2
VideoMMMU
78.9
77.7
MLVU
84.4
83.3
MVBench
74.4
72.7
LVBench
70.0
68.1
MMVU
67.8
66.6
Visual Agent
ScreenSpot Pro
65.2
64.2
OSWorld-Verified
41.8
40.9
AndroidWorld
57.8
55.7
Tool Calling
TIR-Bench
45.6/31.9
45.6/31.9
V*
90.1/88.5
90.1/88.5
Medical VQA
SLAKE
79.0
78.0
PMC-VQA
57.9
56.7
MedXpertQA-MM
49.9
48.7
* MathVision: our model’s score is evaluated using a fixed prompt, e.g., “Please reason step by step, and put your final answer within \boxed{}.” For other models, we report the higher score between runs with and without the \boxed{} formatting.
* BabyVision: scores reported as "with CI / without CI".
* TIR-Bench and V*: scores reported as "with CI / without CI".
* Empty cells (--) indicate scores not yet available or not applicable.
Quantization Details
This model was quantized by applying NVFP4 to the weights and activations of linear operators within transformer blocks. The KV-cache is not quantized. Vision encoder weights are kept in their original precision.
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 for full details.
