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/Qwen3.5-4B (4B 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: 9216
LM Output: 248320 (Tied to token embedding)
MTP: trained with multi-steps
Context Length: 262,144 natively and extensible up to 1,010,000 tokens.
Benchmark Results
Language
Qwen3.5-4B
Qwen3.5-4B-NVFP4
Knowledge & STEM
MMLU-Pro
79.1
77.9
MMLU-Redux
88.8
87.0
C-Eval
85.1
83.0
SuperGPQA
52.9
52.4
GPQA Diamond
76.2
74.1
Instruction Following
IFEval
89.8
87.5
IFBench
59.2
58.1
MultiChallenge
49.0
47.8
Long Context
AA-LCR
57.0
56.2
LongBench v2
50.0
49.0
Reasoning & Coding
HMMT Feb 25
74.0
72.7
HMMT Nov 25
76.8
75.4
LiveCodeBench v6
55.8
55.1
OJBench
24.1
23.7
General Agent
BFCL-V4
50.3
49.4
TAU2-Bench
79.9
78.7
VITA-Bench
22.0
21.7
DeepPlanning
17.6
17.4
Multilingualism
MMMLU
76.1
74.8
MMLU-ProX
71.5
70.4
NOVA-63
54.3
53.3
INCLUDE
71.0
69.6
Global PIQA
78.9
77.5
PolyMATH
51.1
49.8
WMT24++
66.6
64.1
MAXIFE
78.0
75.2
* 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
Qwen3.5-4B
Qwen3.5-4B-NVFP4
STEM and Puzzle
MMMU
77.6
76.1
MMMU-Pro
66.3
65.1
MathVision
74.6
73.0
Mathvista(mini)
85.1
82.8
We-Math
75.4
72.6
DynaMath
83.3
80.3
ZEROBench
3.0
2.9
ZEROBench_sub
26.3
26.0
VlmsAreBlind
92.6
90.6
BabyVision
16.0/19.1
16.0/19.1
General VQA
RealWorldQA
79.5
77.1
MMStar
78.3
77.4
MMBenchEN-DEV-v1.1
89.4
87.0
SimpleVQA
43.4
42.2
HallusionBench
65.0
63.5
Text Recognition and Document Understanding
OmniDocBench1.5
86.2
85.1
CharXiv(RQ)
70.8
69.5
MMLongBench-Doc
54.2
52.9
CC-OCR
76.7
74.6
AI2D_TEST
89.6
88.2
OCRBench
85.0
82.0
Spatial Intelligence
ERQA
54.0
52.4
CountBench
96.3
94.9
RefCOCO(avg)
88.1
85.9
EmbSpatialBench
81.3
78.9
RefSpatialBench
54.6
53.1
LingoQA
74.4
72.2
Hypersim
12.5
12.2
Nuscene
9.9
9.6
Video Understanding
VideoMME(w sub.)
83.5
81.1
VideoMME(w/o sub.)
76.9
75.7
VideoMMMU
74.1
73.0
MLVU
82.8
81.7
MVBench
71.2
69.6
LVBench
66.4
64.6
MMVU
64.9
63.7
Visual Agent
ScreenSpot Pro
60.3
59.3
OSWorld-Verified
35.6
34.9
AndroidWorld
58.6
56.5
Tool Calling
TIR-Bench
38.9/29.9
38.9/29.9
V*
84.3/86.4
84.3/86.4
Medical VQA
SLAKE
76.1
75.1
PMC-VQA
55.5
54.4
MedXpertQA-MM
42.9
41.9
* 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.
