Gemma 4 strains single RTX 3090

The 3090 looks viable, but only if you accept that “fits” and “feels good” are different thresholds. Gemma 4’s 26B MoE is the safer local bet; the 31B dense model is where 24GB starts turning into a hard ceiling once KV cache and desktop overhead enter the picture.

• –Google says the quantized 26B/31B models are meant to run on consumer GPUs, with the 26B MoE activating only 3.8B parameters per token, which explains why it should feel much lighter than the dense model.
• –Community reports in the thread are consistent: the 26B MoE can hit roughly 80K to 128K context on a 3090 in careful setups, while the 31B dense model tends to land around low double-digit thousands of context before offload or slowdown becomes painful.
• –Speed is the tradeoff that matters most. One user reports about 98 tok/s for the 26B MoE and about 26 tok/s for the 31B dense at more modest contexts, while CPU offload pushes dense-model throughput down into unusable territory for interactive work.
• –For coding assistants, short-doc chat, and local-first agent workflows, a single 3090 is good enough. For “I want to actually use most of 256K” workloads, you’re really in 48GB-plus or multi-GPU territory.
• –If you only buy one card, buy it for the 26B MoE use case first; treating the 31B dense as the default local model on 24GB is optimistic.