Dolphin 2.9.1 Yi 1.5 34B · 10 GPUs measured first-party · llama.cpp Q4_K_M · Updated July 2026

What GPU Do You Need for Dolphin 2.9.1 Yi 1.5 34B?

Dolphin 2.9.1 Yi 1.5 34B is the heavyweight of the Dolphin line, an uncensored tune of 01.AI's Yi 34B, the biggest judge in our database. Measured on 10 GPUs (llama.cpp, Q4_K_M): 85 tok/s on the B300, ~21GB peak VRAM, same hardware class as the dense 32Bs.

Benchmarked weights: bartowski/dolphin-2.9.1-yi-1.5-34b-GGUF

Maximum brain, maximum bluntness, maximum diversity. Three things justify the biggest Dolphin. It's the smartest uncensored model we benchmark, so its critiques carry 32B-class depth. It's built on Yi, a lineage almost nobody else in your stack will share, so in a consensus panel it disagrees in genuinely novel ways where Qwen-family models fail together. And it slots into the same 24GB budget as the other big dense models: ~21GB floor, one weight class, interchangeable rotation. Its age is also its character: Dolphin 2.9.1 comes from the earlier, blunter era of the line: in our experience the older Dolphins on older bases were the most genuinely uncensored, and the newer generations have drifted more neutral. If you want the classic no-filter Dolphin temperament at maximum brain size, this vintage is it.

Chart notes. 85 tok/s peak actually edges out both dense 32Bs we run (83 apiece), Yi's architecture generates a touch quicker per parameter. The H100's 77 tok/s at 199W (0.39 tok/W) leads efficiency, and the familiar dense-model advice applies: this class is bandwidth-bound and single-stream, so consumer flagship cards deliver most of the datacenter experience, spend on VRAM and bandwidth, not on rented scale.

Our verdict

Dolphin 2.9.1 Yi 34B: 85 tok/s peak, ~21GB floor, the biggest, bluntest reviewer in our lineup, on a base model nothing else in your stack shares. If your consensus panel needs one voice guaranteed to think differently, this is it, and any 24GB card seats it.

FAQ

Why the Yi base?

Diversity. Yi 34B shares training lineage with neither Qwen nor Llama nor Mistral, so in a multi-model panel its errors don't correlate with the others'. For consensus setups, an uncorrelated strong opinion is the most valuable seat.

What GPU does it need?

24GB. Measured peak was ~21GB at Q4_K_M. Same class as DeepSeek 32B and Qwen 32B, so one RX 7900 XTX / RTX 3090 / 4090 hosts your entire big-model rotation sequentially.

How does it compare to the 32B reasoning models on speed?

Slightly faster: 85 tok/s peak vs 83 for both dense 32Bs, and it doesn't burn reasoning tokens, so effective throughput is meaningfully higher per answer.

Is it better than the 24B Dolphins as a judge?

Deeper, yes; cheaper, no. The Mistral 24Bs fit ~15GB and run 40% faster. Our split: 24B Dolphin as the everyday referee, this 34B when the verdict itself is high-stakes.

Does uncensored matter more at this size?

Arguably yes, a 34B's critiques are sharp enough to be worth hearing unfiltered. Alignment hedging costs more when the underlying judgment is this good.

Why run a 34B judge instead of just a bigger answer model?

Because verification and generation are different jobs. A second big generator gives you a second opinion with correlated blind spots; a big uncensored judge from a different lineage (Yi) audits both answers with uncorrelated instincts and no diplomatic softening. In our experience the panel-plus-judge structure beats one-more-generator every time.