DEV.co
Open-Source LLM · Qwen

Qwen3-235B-A22B-Instruct-2507

Qwen3-235B-A22B-Instruct-2507 is a 235-billion parameter mixture-of-experts (MoE) language model from Alibaba's Qwen team. It activates 22B parameters per inference, achieving competitive performance on reasoning, knowledge, coding, and multilingual tasks via Apache 2.0 open-source licensing. Native context window is 262K tokens, extendable to 1M with specialized techniques. Model is gated-free and available for local deployment or managed endpoints.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507
235.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
86.1k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperQwen
Parameters235.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads86.1k
Likes787
Last updated2025-09-17
SourceQwen/Qwen3-235B-A22B-Instruct-2507

What Qwen3-235B-A22B-Instruct-2507 is

Causal language model with MoE architecture: 235B total parameters, 22B activated, 94 layers, 64 query/4 KV attention heads (GQA), 128 experts (8 activated per token). Supports 262,144 native context length, extendable to 1,010,000 tokens via Dual Chunk Attention (DCA) and MInference sparse attention. Trained with pretraining and post-training stages. Requires transformers≥4.51.0 for architecture support. Deployable via vLLM (≥0.8.5), SGLang (≥0.4.6.post1), Ollama, LMStudio, llama.cpp, MLX-LM, and KTransformers.

Quickstart

Run Qwen3-235B-A22B-Instruct-2507 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Qwen/Qwen3-235B-A22B-Instruct-2507")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Complex Reasoning & Mathematical Problem Solving

Achieves 70.3% on AIME25 and 55.4% on HMMT25, significantly outperforming comparable models. Suitable for competitive math, scientific problem-solving, and logic-heavy tasks.

Long-Document Processing & Retrieval

Native 262K context (extendable to 1M) with DCA/MInference optimization. Ideal for document summarization, cross-document reasoning, and legal/financial document analysis.

Multilingual Agentic Systems

Scores 77.5% on MultiIF, 79.4% on MMLU-ProX, and supports function calling via Qwen-Agent framework. Suitable for deployed chatbots, customer service agents, and MCP-integrated tools across languages.

Running & fine-tuning it

ESTIMATE (verify with your exact deployment): Full precision (bfloat16): ~470 GB VRAM for inference. With quantization (int8/int4): 235–120 GB VRAM. For 262K context: reserve ≥100 GB additional for KV-cache. For 1M-token context: ~1000 GB total GPU memory (weights + KV-cache + activations). Tensor-parallelism across 8+ GPUs typical for production. Deployment examples use `--tp 8` and `--context-length 262144`.

Not explicitly documented. MoE models are compatible with LoRA/QLoRA in principle; Qwen team provides Qwen-Agent framework for tool-calling integration but does not explicitly detail fine-tuning pipelines. Recommend verifying LoRA compatibility with transformers library and testing on small batch before production fine-tuning. Context extension (1M tokens) requires config replacement and recompilation; full fine-tuning is not standard documented path.

When to avoid it — and what to weigh

  • Real-time, Ultra-low-latency Production Inference — Model size (235B) and MoE routing overhead require significant GPU capacity. Latency depends on hardware; verify deployment targets meet SLA requirements before committing.
  • Strict Closed-source or Proprietary Deployment Constraints — Licensed under Apache 2.0 (permissive open-source). If your deployment or use case requires non-open-source guarantees or vendor lock-in, consider proprietary alternatives (GPT-4, Claude).
  • Resource-Constrained Environments — Full model requires ~470 GB VRAM in bfloat16, or ~1000 GB for 1M-token context. Not suitable for edge devices, mobile, or single-GPU setups without quantization.
  • Zero-shot Performance Requirements Without Fine-tuning — Model is instruction-tuned but performance may vary on highly specialized domains. If your use case requires domain-specific accuracy out-of-the-box, benchmark against task-specific fine-tuned alternatives.

License & commercial use

Apache License 2.0 (OSI-compliant, permissive open-source). Permits commercial use, modification, and redistribution with attribution. No viral copyleft. Model weights and code are freely available; no gating or registration required.

Apache 2.0 is a permissive, OSI-approved license that explicitly permits commercial use, including proprietary deployment and SaaS integration. No restrictions on monetization, resale, or proprietary modifications. However: (1) verify your deployment infrastructure and any third-party dependencies comply with their own licenses; (2) Qwen team disclaims warranties; (3) ensure your use case complies with local regulations (e.g., export controls, content policy). No commercial support, SLA, or indemnification from Qwen unless separately contracted.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model inherits typical LLM safety considerations: (1) No explicit mention of red-teaming, jailbreak testing, or adversarial robustness in card. (2) No formal security audit or vulnerability disclosure program described. (3) Model is open-source; inspect code/weights for integrity if deploying in high-assurance environments. (4) Inference endpoints (vLLM/SGLang) introduce network attack surface; apply standard API security (rate-limiting, authentication, input validation). (5) MoE routing and sparse attention add complexity; test for information leakage or model poisoning risks in federated/multi-tenant setups. (6) Document any data or prompts fed to model if compliance/privacy-sensitive; Qwen does not disclose model telemetry policies.

Alternatives to consider

Deepseek-V3-0324

Similar scale (671B total, 37B activated), competitive reasoning benchmarks. Closed-source; proprietary licensing. Consider if you require vendor support or SLA guarantees.

Claude 3.5 Opus (non-thinking)

Proprietary, strong on long-context (200K) and alignment tasks. Hosted API with SLA/support. Trade-off: no local deployment, higher per-token cost, reduced transparency.

Llama 3.1-405B

Larger open-source alternative (405B parameters), strong coding/reasoning. Meta's Llama license (permissive but with acceptable-use clause). Better community ecosystem for quantization/serving; higher VRAM overhead.

Software development agency

Ship Qwen3-235B-A22B-Instruct-2507 with senior software developers

Compare hardware costs, fine-tuning feasibility, and alternative models using our evaluation framework. Contact us to architect a secure, scalable LLM deployment strategy tailored to your use case.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Qwen3-235B-A22B-Instruct-2507 FAQ

Can I use this model commercially in a SaaS product?
Yes. Apache 2.0 explicitly permits commercial use, including SaaS deployment and monetization. Ensure you attribute Qwen and comply with your jurisdiction's laws. No commercial support or SLA is provided; you assume all liability.
What GPU/hardware setup do I need to run this locally?
Minimum: 2x H100 80GB (with quantization) or 8x H100 for bfloat16 full precision. vLLM/SGLang examples use `--tp 8` (tensor-parallelism). For 1M-token context, expect ~1000 GB total VRAM. Quantization (4-bit) can reduce footprint to ~120 GB but may impact quality.
Does this model support fine-tuning?
Not explicitly documented. The model card does not provide fine-tuning pipelines or LoRA compatibility details. LoRA is theoretically feasible with latest transformers; test on a small dataset first. Full fine-tuning is resource-intensive; contact Qwen team or GitHub for guidance.
How does context extension to 1M tokens work, and what's the performance impact?
Uses Dual Chunk Attention (DCA) and MInference sparse attention. Requires replacing config.json with config_1m.json. Card reports up to 3× speedup vs. standard attention on long sequences. Trade-off: requires ~1000 GB GPU memory; verify quality on your domain-specific tasks (no 1M-token benchmarks provided in card).

Custom software development services

DEV.co helps companies turn open-source tools like Qwen3-235B-A22B-Instruct-2507 into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy Qwen3-235B for Your Enterprise?

Compare hardware costs, fine-tuning feasibility, and alternative models using our evaluation framework. Contact us to architect a secure, scalable LLM deployment strategy tailored to your use case.