Qwen3-Swallow-32B-RL-v0.2-AWQ-INT4

Qwen3-Swallow-32B-RL-v0.2-AWQ-INT4 is the 4-bit AWQ-quantized release of Qwen3-Swallow-32B-RL-v0.2, developed by Tokyo Institute of Science (tokyotech-llm) in collaboration with AIST under the Swallow LLM project. Built on Alibaba's Qwen3-32B, it is a dense bilingual Japanese–English reasoning model trained through continual pre-training (209.7B tokens at 32K context), supervised fine-tuning, and reinforcement learning with verifiable rewards (RLVR). The RL variant is the recommended production checkpoint and achieves state-of-the-art results among open 32B-class models on the Swallow LLM Leaderboard v2 (February 2026), with particularly strong Japanese knowledge, translation, and reasoning.

The AWQ-INT4 variant uses activation-aware quantization calibrated on validated RL dataset outputs to reduce VRAM from ~65GB (FP16) to roughly 18–24GB, making 32B reasoning inference feasible on a single consumer GPU. Official HuggingFace guidance recommends vLLM or SGLang serving with the qwen3 reasoning parser, generation parameters of temperature 0.6, top_p 0.95, top_k 20, min_p 0, and a maximum context length of 32,768 tokens. Quantized benchmark scores are listed as coming soon; performance is expected to closely track the full-precision RL model.

On the Swallow LLM Leaderboard (zero-shot reasoning evaluation), the base Qwen3-Swallow-32B-RL-v0.2 scores 77.8% on MMLU-Pro (English), 63.8% on LiveCodeBench (English coding), 98.2% on MATH-500, 82.5% on AIME 2024–25, and 92.9% on JHumanEval (Japanese HumanEval). The model defaults to chain-of-thought reasoning and does not support reasoning on/off toggling or native tool use. Released under Apache 2.0 on 2026-02-20, with GPTQ variants withdrawn on 2026-02-23 due to quality degradation.