Qwen3.8 27B
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 37 tok/s, typical error 50% · first token ~1.0s
Fits in 20 GB VRAM with room to spare. Best for coding, agent, vision, long context on RX 7900 XT.
The RX 7900 XT sits in a capacity gap NVIDIA does not serve: 20GB of VRAM, more than any 16GB card and cheaper than any 24GB one. That extra 4GB over a 5070 Ti lets 27B-class models load fully in VRAM at Q4, and 800 GB/s of bandwidth keeps generation quick. Like the XTX, it runs Ollama and llama.cpp through ROCm with a Vulkan fallback.
The best local LLM for the RX 7900 XT is Qwen3.8 27B (est. 37 tok/s, typical error 50%) on its 20GB VRAM. It uses ~16.5GB of VRAM; the RX 7900 XT handles dense models up to ~30B at Q4. A 14B model at Q4 runs at est. 60 tok/s, typical error 10%. (ModelFit, 2026)
Sizing rule: a Q4 model needs about 0.6 GB of VRAM per billion parameters for its weights, and ModelFit budgets 90% of this card's 20GB (18GB) for weights, context, and KV-cache. Real files often load more than that rule, so the per-size table below checks actual builds: a size fits only when the lightest Q4 build of that size in our catalog loads within 18GB. MoE rows, and sizes with no such build, use a typical Q4 file size. Other strong fits: Qwen3.5 9B Instruct (Q8) (9B, ~10.7GB) and Gemma 4 12B (Q8) (12B, ~12.8GB). On this card Qwen3.8 27B runs at est. 37 tok/s, typical error 50%. What it will not run: a 32B model does not fit in VRAM alone, so it spills to system RAM over PCIe and slows sharply.
Speeds are ModelFit estimates from memory bandwidth and model size, not measured benchmarks. Typical error: the median gap between the ModelFit speed model and measured runs on hardware the model was not fitted on, for this backend and model size.
Cite this page: ModelFit, RX 7900 XT 20GB Local LLM: Runs 27B at est. 37 tok/s, https://modelfit.io/gpu/rx-7900-xt/, updated November 2026, CC BY 4.0.
| Model size | Est. speed | Fit on 20GB |
|---|---|---|
| 7B | est. 102 tok/stypical error 10% | Fits in VRAM |
| 14B | est. 60 tok/stypical error 10% | Fits in VRAM |
| 20B MoE (3.6B active) | est. 114 tok/stypical error 25% | Fits in VRAM |
| 32B | est. 6 tok/srough estimate, offload not measured | CPU offload (slow) |
| 35B MoE (3B active) | est. 96 tok/srough estimate, offload not measured | CPU offload (slow) |
| 70B | est. 1 tok/srough estimate, offload not measured | CPU offload (slow) |
| 120B MoE (5.1B active) | est. 31 tok/srough estimate, offload not measured | CPU offload (slow) |
Decode speed, generic Q4_K_M model of each size at 4K context, 800 GB/s, typical error 10-25%
ModelFit estimates, not measured benchmarks: decode speed of a generic Q4_K_M model of each size at 4K context on the RX 7900 XT's 800 GB/s bandwidth. MoE rows scale by active parameters (decode reads only the active experts), so a 35B MoE runs far faster than a dense 32B. "CPU offload" sizes exceed the 20GB VRAM; dense models slow to a crawl there, MoE models degrade less because hot experts stay GPU-resident. Offload rows are rough estimates: no measured run covers offload, so they carry no typical error. What typical error means.
Qwen3.8 27B loads ~16.5 GB of weights; at 16k context the KV cache adds ~1.0 GB (still fits the ~18 GB usable VRAM), and at 64k it adds ~4.0 GB (exceeds the budget, use a smaller quant or a q8_0 KV cache).
The model leaves about 0.5 GB of the usable VRAM budget free after weights and 16k context.
KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.
One email when a new open-weight model is a better fit for 20 GB of VRAM, plus the Thursday weekly.
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
ModelFit may earn a commission on purchases through these links, at no extra cost to you.
ModelFit may earn a commission on purchases through these links, at no extra cost to you.
20GB GDDR6 at 800 GB/s is a unique spot in the lineup: 16GB cards top out at 14B-class dense models, while the 7900 XT fits a 27B at Q4 (~18GB) right at its usable budget, with the KV cache limiting context length. It cannot hold 32B Q4 (~20GB), so treat 27B as the practical ceiling. RDNA 3 is on the official ROCm support list, so Ollama, llama.cpp, and LM Studio all run it without CUDA. On the used market it usually undercuts every 24GB card, making it a strong value pick for the 14B-27B class.
| Hardware | Memory | Speed | Bandwidth |
|---|---|---|---|
| RTX 5070 Ti | 16 GB | est. 125 tok/stypical error 5% | 896 GB/s |
| RX 7900 XT | 20 GB | est. 91 tok/stypical error 10% | 800 GB/s |
| RX 7900 XTX | 24 GB | est. 104 tok/stypical error 10% | 960 GB/s |
| RTX 3090 | 24 GB | est. 117 tok/stypical error 5% | 936 GB/s |
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 37 tok/s, typical error 50% · first token ~1.0s
Fits in 20 GB VRAM with room to spare. Best for coding, agent, vision, long context on RX 7900 XT.
Best for: Quality, Coding, Reasoning·Pop: 86/100
Perf: est. 65 tok/s, typical error 10% · first token ~1.7s
Fits in 20 GB VRAM with room to spare. Best for quality, coding, reasoning on RX 7900 XT.
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 44 tok/s, typical error 10% · first token ~2.4s
Fits in 20 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 7900 XT.
Best for: Chat, Coding, Multimodal·Pop: 86/100
Perf: est. 111 tok/s, typical error 25% · first token ~0.3s
Fits in 20 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 7900 XT.
Best for: Local AI agents, privacy-first tool calling, MCP workflows·Pop: 80/100
Perf: est. 160 tok/s, typical error 25% · first token ~0.2s
Fits in 20 GB VRAM with room to spare. Best for local ai agents, privacy-first tool calling, mcp workflows on RX 7900 XT.
Best for: Chat, Reasoning, Vision·Pop: 74/100
Perf: est. 60 tok/s, typical error 10% · first token ~0.6s
Fits in 20 GB VRAM with room to spare. Best for chat, reasoning, vision on RX 7900 XT.
Best for: Agentic coding, Software engineering·Pop: 74/100
Perf: est. 39 tok/s, typical error 50% · first token ~1.0s
Fits in 20 GB VRAM with room to spare. Best for agentic coding, software engineering on RX 7900 XT.
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 68 tok/s, typical error 10% · first token ~0.6s
Fits in 20 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 7900 XT.
Best for: Chat, Coding, Reasoning·Pop: 85/100
Perf: est. 112 tok/s, typical error 25% · first token ~1.2s
Fits in 20 GB VRAM with room to spare. Best for chat, coding, reasoning on RX 7900 XT.
Best for: Quality, Coding, Reasoning·Pop: 86/100
Perf: est. 96 tok/s, typical error 10% · first token ~0.4s
Fits in 20 GB VRAM with room to spare. Best for quality, coding, reasoning on RX 7900 XT.
No stated VRAM figure or published run covers AMD for these models yet. By memory size: video models · image models.
No speed test run for this GPU yet.
Up to 27B parameter models at Q4 quantization. Its 20GB VRAM gives about 18GB usable, which holds a 27B Q4 (~18GB) fully in VRAM with context kept modest. 32B models need CPU offloading, which cuts speed sharply for dense models.
Yes. RDNA 3 (gfx1100) is officially supported by ROCm, which Ollama uses on Linux and Windows. llama.cpp and LM Studio also work via ROCm/HIP or Vulkan.
The 7900 XT has 4GB more VRAM (20 vs 16GB), so it runs 27B-class models the 5070 Ti cannot fit. The 5070 Ti has higher bandwidth (896 vs 800 GB/s) and the CUDA ecosystem, so it is faster on the 14B class both can hold. Pick by target model size.
The XTX if the budget allows: 24GB vs 20GB moves the ceiling from 27B to 32B-class models, and its 960 GB/s is 20% faster. The XT wins on value when the models you want stay at or under 27B.
The RX 7900 XT's 20GB of VRAM cannot fit a 32B model comfortably. The largest size class it fits is 20B MoE (3.6B active), at est. 114 tok/s, typical error 25%. The largest individual model in the catalog that fits is Muse Glimmer 30B (30B).
Time a model that fits your machine with Ollama, using the same prompts as every other run. The CLI shows your result and asks before it sends anything; your machine then joins the speed test leaderboard. It does not change the quality benchmark.
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.