The best local LLMs for the RX 9070 XT

Qwen3.8 27B (27B) fits, with little room to spare in the RX 9070 XT's memory at Q3_K_M, using about 15GB with 8K context and generating ~30 tokens/sec. It is the most capable model in our catalog that runs well on this card. It is also our top pick for coding, reasoning and working with images. It is also the largest model that runs fully on this GPU.

Qwen3.8 27B, released in August 2026, is one of the strongest open models that fits a single 24GB GPU, with image input, a 262K native context, and switchable thinking. It uses the same hybrid layout as the Qwen3.5 and 3.6 27B models, where three of every four layers use linear attention, so long contexts cost far less memory than on older dense models. At Q4_K_M it needs about 18GB with 8K context, leaving room for 64K or more of context on a 24GB card, and 32GB cards can run Q6_K. It supersedes Qwen3.6 27B and Qwen3 32B, but its long reasoning traces make speed on slower hardware a real consideration.

Qwen3.5 9B (9B) fits comfortably in the RX 9070 XT's memory at Q8_0, using about 12GB with 8K context and generating ~36 tokens/sec. It is the next-best model that runs well here, and the one to try if the top pick doesn't suit you.

Qwen3.5 9B is the largest of the small Qwen3.5 models released in March 2026, with built-in image understanding, a 262K context, and switchable thinking. It uses Qwen's hybrid attention, where most layers are linear-attention layers with a fixed-size state, so long contexts cost far less memory than on older 8B models. It is one of the strongest models that fits an 8GB GPU: at Q4_K_M it needs about 7GB with 8K context, and 12GB cards can run Q8_0. It replaces Qwen3 8B and, for most uses, Qwen3 14B.

Gemma 4 26B A4B (26B (3.8B active)) fits, with little room to spare in the RX 9070 XT's memory at Q3_K_M, using about 14GB with 8K context and generating ~90 tokens/sec. Pick it when responsiveness matters more than squeezing out maximum quality.

Gemma 4 26B A4B is Google's mixture-of-experts Gemma, with 26B total parameters and 3.8B active (128 experts, 8 used per token), image input, a 256K context, and a thinking mode. Because so few parameters are active, it generates text at small-model speed while matching much larger dense models on reasoning benchmarks. At Q4_K_M it needs about 17GB with 8K context, so it fits a 24GB GPU with plenty of room for context; on 16GB cards it runs at Q3_K_M or with a few experts in system RAM. It replaces Gemma 3 27B for most local uses.