Best AMD GPU for local LLMs

The short answer

We judge every card at 256K context, because a local model with a short memory is not worth running. If a 32 GB workstation card is too much, a 16 GB RX 9060 XT runs the same 35B with part of it in system RAM. It is slower, but it works. 96 GB DDR5 is the house RAM size for most builds.

The picks

Street prices change. We omit today’s asking prices unless we read the retailer on 2026-11-02. AMD’s R9700 launch MSRP is $1,299. Bandwidth figures are AMD specs.

Why the RX 9070 XT is not the middle pick for most people

It shares the R9700’s chip family and 640 GB/s bandwidth but has half the VRAM. The 35B does not fit, so the card spends time waiting on system RAM. Two 9070 XTs give you 32 GB split across two cards, need an x8/x8 board, and splitting a model across cards was slower than one card in every row we measured on two R9700s on the local desktop. Buy the 9070 XT if you also game. For LLMs only, the 9060 XT gets you the same offload setup for less.

• RX 9070 XT
• RX 9060 XT
• One vs two GPUs

The W7800 question

The Radeon PRO W7800 is another 32 GB Radeon, on older RDNA 3, with 576 GB/s (AMD spec) versus the R9700’s 640 GB/s. Get it only if you need a particular seller’s own-stock return policy or cannot find an R9700. Otherwise the R9700 is the better card for this job. We do not print a street comparison here.

• PRO W7800

Cards to skip new

The RX 7900 XTX, 7900 XT, 7900 GRE and 7600 XT are last-generation for this job. Prefer used or skip them. See Buying used Radeon.

• Buying used Radeon

Buy nothing if…

You already have a 16 GB or larger Radeon and at least 64 GB of RAM — try expert offload first. You want the bigger Qwen 3.8 Flash — that is mostly a RAM decision (two R9700s plus 96 GB runs Q4 at 256K).

• Can my Radeon run it?
• Flash requirements

Our pick: Radeon AI PRO R9700 (32 GB)

AMD launch MSRP is $1,299. We omit today’s street ask. Newegg listings are often marketplace sellers — that seller’s warranty and return policy apply. Check the listing before you buy.

Check R9700 listings

FAQ

Is AMD good for local LLMs in 2026?

Yes for inference on Linux with llama.cpp (ROCm or Vulkan). Some NVIDIA-only kernels are still missing on RDNA 4. See the lab log. Measured 35B numbers on this site come from the local desktop R9700.

Lab log · Radeon vs NVIDIA

Is 16 GB enough?

For MoE models like Qwen3.6-35B, yes with expert offload and 64 GB or more of RAM, at lower speed. Not for fully on-GPU 256K.

35B requirements

R9700 or 9070 XT?

Same chip family and bandwidth; the R9700 has twice the VRAM, and that is what decides LLM fit.

R9700 · 9070 XT

Sources

• S-LAB — Local desktop 35B llama.cpp cards — our lab · checked 2026-11-02 · Measured on the local desktop
• S-AMD — AMD Radeon AI PRO R9700 product page — 32 GB, 640 GB/s, launch MSRP — third-party · checked 2026-11-02 · AMD launch MSRP $1,299
• S-35B — 35B requirements at 256K — our lab · checked 2026-11-02 · site hub