Core Announcement: Qwen3-2507 Series Launch

Tongyi Lab has released the Qwen3-2507 series of large language models, featuring two variants and three sizes:

• Two variants: Qwen3-Instruct-2507 (non-thinking mode for general chat) and Qwen3-Thinking-2507 (thinking mode for complex reasoning)
• Three sizes: Qwen3-235B-A22B (MoE), Qwen3-30B-A3B (MoE), and Qwen3-4B (dense)

All model weights are open-sourced and available via Hugging Face or ModelScope. The 235B and 30B versions rolled out from late July to early August 2025, with the 4B models released on August 6.

Models support 256K-token long context, extendable to 1 million tokens (available since August 8, 2025), enabling document summarization and long-text generation tasks.

Performance Enhancements

Qwen3-Instruct-2507, as the evolved non-thinking variant, demonstrates significant improvements in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage. Multi-language long-tail knowledge coverage has expanded, and alignment with user preferences in subjective tasks is markedly better. The 256K-token context capability extends to a maximum of 1 million tokens.

Qwen3-Thinking-2507 continues the thinking model lineage, achieving state-of-the-art results among open-weight thinking models on reasoning benchmarks including logical reasoning, mathematics, science, coding, and academic evaluations. Its general capabilities—instruction following, tool usage, text generation, and human preference alignment—have also improved alongside long-context handling.

A notable feature is the hybrid dense/MoE architecture: 235B-A22B and 30B-A3B use Mixture-of-Expert (MoE) design (activating only a subset of parameters for efficiency), while the 4B is a traditional dense model better suited for resource-constrained deployments.

Usage and Technical Details

Models load via Hugging Face Transformers with transformers≥4.51.0 required. Example code is publicly provided:

1
2
3
4
5
6
7
8

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen3-30B-A3B-Instruct-2507"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

Note: Instruct-2507 models do not generate blocks by default (no enable_thinking=False needed). Thinking-2507 outputs reasoning blocks explicitly.

MoE models work with local frameworks like llama.cpp, Ollama, LM Studio, and server deployment tools including SGLang, vLLM, and TGI.

Deployment Guidance

Choose immediately if you need:

• Strong reasoning capabilities: Select Thinking-2507 for math, coding, or academic tasks
• Long-context processing: 256K-token baseline with 1M-token上限 suitable for multi-document analysis
• Resource-constrained environments: The 4B dense model offers lightweight deployment

Consider waiting if:

• Your production use case demands extremely high reasoning reliability: Despite leading in open-weight categories,Thinking-2507 should still be benchmarked against commercial closed models
• You need multimodal capabilities: This release covers text-only models; multimodal versions are unmentioned

Final Thoughts

The Qwen3-2507 series has matured the deliberate separation of reasoning and general chat modes, with open models now matching commercial peers in thinking ability. The combination of MoE architecture and million-token context represents a shift from “parameter competition” to “practical efficiency” in large model development, setting new benchmarks for open-source practicality.