Qwen3-235B-A22B-Instruct-2507

Qwen3-235B-A22B-Instruct-2507 is an updated version of the flagship MoE model in the Qwen 3 series. This 235-billion-parameter model activates only 22 billion parameters at each inference step. The architecture consists of 94 transformer layers with 128 experts, of which only 8 are activated per token.

Unlike previous versions of Qwen, the new 2507 model has completely abandoned the hybrid thinking mode in favor of a highly optimized non-thinking mode. This decision was made based on user feedback indicating a preference for faster responses without the generation of <think> blocks. As a result, response speed has dramatically increased, along with significant improvements in output quality. In mathematical benchmarks, the model achieved remarkable gains: AIME25 (70.3 vs. 24.7 in the previous version) and HMMT25 (55.4 vs. 10.0). Particularly impressive is its performance on ZebraLogic (95.0), demonstrating near-perfect accuracy in logical reasoning tasks. In programming, the model also significantly outperforms its predecessor, achieving state-of-the-art results on LiveCodeBench and MultiPL-E. Overall, across numerous benchmarks, the model surpasses leading competitors such as GPT-4o, DeepSeek-V3, and Kimi K2.

Additionally, developers have released Qwen3-235B-A22B-Instruct-2507-FP8—an FP8 quantized version of the model. This innovative technique reduces memory requirements by approximately 50% while preserving nearly all of the original model's quality. The FP8 format outperforms traditional INT8 approaches, especially for large-scale models, offering a superior balance between accuracy and efficiency.

Another key technological advancement of Qwen3-235B-A22B-Instruct-2507 is native support for a context length of 262,144 tokens. This capability enables entirely new use cases—from analyzing lengthy documents and codebases to conducting multi-hour conversations while maintaining contextual understanding and high response accuracy even with a fully filled context window. Therefore, there are strong grounds to believe that the new Qwen3-235B-A22B-Instruct-2507 model positions itself as the leading open-source solution for a broad range of enterprise applications.

Announce Date: 21.07.2025
Parameters: 235B
Experts: 128
Activated at inference: 22B
Context: 263K
Layers: 94
Attention Type: Full or Sliding Window Attention
VRAM requirements: 172.7 GB using 4 bits quantization
Developer: Qwen
Transformers Version: 4.51.0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Qwen3-235B-A22B-Instruct-2507 capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.