Qwen3-235B-A22B

Qwen3-235B-A22B is a sparse MoE decoder. Total parameters are 235B; 22B activate per forward pass (8 of 128 experts routed per token). It uses 94 transformer layers, Grouped Query Attention, SwiGLU activations, Rotary Positional Embeddings with frequency scaling, and RMSNorm pre-normalization — the same modern stack as the Qwen3 dense models, scaled up and made sparse. Pre-training ran on roughly 36 trillion tokens spanning 119 languages and dialects, followed by a multi-stage post-training pipeline that fuses reasoning (long-CoT) and non-reasoning behavior into one model with a runtime switch. Native context is 131K. Alibaba disclosed these figures in the Qwen3 launch blog and Technical Report (arXiv 2505.09388), so unlike most frontier models the architecture is fully public.

The 22B active footprint keeps inference cost and latency closer to a 30B dense model than to a true 235B, while the sparse capacity delivers frontier-adjacent quality. Reasoning and math are the headline strengths (cap_reasoning 8.5, cap_math 8.7): MMLU-Pro 82.8, GPQA Diamond 70.0, AIME 2025 81.5 with thinking mode on. Coding is strong (cap_coding 8.0) — LiveCodeBench v5 70.7, best among open-weight at release. Tool-use, parallel function calls, and structured JSON output are natively trained (cap_function_calling 8.0, cap_agentic 7.5). The defining edge is multilingual coverage (cap_multilingual 9.0): Chinese, Japanese, Korean, Arabic, and major Indic-language performance materially exceeds DeepSeek and Llama peers — this is the open-weight model teams reach for when the workload mixes English and Asian languages. Long-context is honest to roughly 64-96K before degradation (cap_long_context 7.0). It has no vision or live-data access (cap_vision 0, cap_realtime_data 0). Safety calibration is mid (cap_safety_calibration 6.0): refusal behavior is Western-comparable except on PRC-sensitive political topics.

AIME 2024 was reported at 85.7 and Arena Hard at 95.6 in the Technical Report; CodeForces rating 2056, BFCL v3 70.8. The LMArena Elo of 1431 is for the refreshed 2507 checkpoint (Aug 2025), not the original April release. Scores reflect thinking mode on; non-thinking scores are lower. Never invented — keys without a verifiable score are null in the data layer.

Median throughput around 68 tokens/sec with time-to-first-token near 0.78s on tuned inference stacks (llm-stats aggregate). In practice, latency varies sharply with the thinking toggle: non-thinking responses feel like a normal 30B-class model; thinking mode can run 5-30x longer because reasoning chains are long. For interactive product surfaces, gate thinking mode behind an explicit "reason" path. Batch throughput is strong on vLLM/SGLang with FP8.

Open weights on Hugging Face and ModelScope under Apache 2.0 — no MAU threshold, full commercial use, redistribution, and fine-tuning rights. FP16 serving needs roughly 470GB VRAM (8x H100 / H200 class); FP8 and AWQ quantizations bring the practical floor to roughly 240GB. GGUF exists for llama.cpp but the MoE is impractical on consumer hardware. Hosted by Together, Fireworks, DeepInfra, Hyperbolic, Novita, and OpenRouter; first-party via Alibaba Cloud Model Studio (DashScope), which offers an international endpoint outside mainland China. Data-residency note: the official DashScope mainland endpoint routes through Alibaba Cloud in China; the international endpoint and third-party providers (US/EU-hosted) avoid that, and self-hosting eliminates egress entirely.

No formal published safety framework or tier label. Alibaba does not train on third-party inference inputs when self-hosted; first-party API data handling follows Alibaba Cloud terms with opt-out. No SOC2/HIPAA certifications attach to the open weights themselves (a hosting provider may carry its own). No built-in content moderation layer. Refusal calibration is Western-comparable on general topics, with notably stricter refusal/deflection on PRC-sensitive political content — an honest enterprise consideration for US public-sector or politically sensitive consumer surfaces.

SDKs via OpenAI-compatible clients (Python, TypeScript). Deep open-source support: vLLM, SGLang, Ollama, llama.cpp, MLX, Transformers, plus LangChain, LlamaIndex, Axolotl, and LLaMA-Factory for fine-tuning. Hosted by Together, Fireworks, DeepInfra, Hyperbolic, Novita, and OpenRouter; first-party via Alibaba Cloud Model Studio. Popularity is mainstream — Qwen3 is one of the most-downloaded open-weight families on Hugging Face and a frequent top-of-leaderboard open model.