Qwen3.8 27B
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 44 tok/s, typical error 50% · first token ~0.9s
Fits in 24 GB VRAM with room to spare. Best for coding, agent, vision, long context on RX 7900 XTX.
The RX 7900 XTX is the AMD counterpart to the RTX 3090: 24GB of VRAM at 960 GB/s, often bought used. That capacity runs 32B parameter models entirely in VRAM, and since token generation is bandwidth-bound, its 960 GB/s matches the 3090's 936 GB/s on paper, though ROCm decode lands a little behind CUDA in ModelFit's estimates (est. 104 tok/s, typical error 10% vs est. 117 tok/s, typical error 5% on 8B models). Ollama and llama.cpp both support it via ROCm, with Vulkan as a fallback, so the CUDA-free stack is no longer the obstacle it once was.
The best local LLM for the RX 7900 XTX is Qwen3.8 27B (est. 44 tok/s, typical error 50%) on its 24GB VRAM. It uses ~16.5GB of VRAM; the RX 7900 XTX handles dense models up to ~33B at Q4. A 32B model at Q4 runs at est. 34 tok/s, typical error 50%. (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 24GB (21.6GB) 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 21.6GB. MoE rows, and sizes with no such build, use a typical Q4 file size. Other strong fits: Gemma 4 26B-A4B (26B, ~16GB) and Gemma 4 12B (Q8) (12B, ~12.8GB). On this card Qwen3.8 27B runs at est. 44 tok/s, typical error 50%. What it will not run: a 35B MoE (3B active) 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 XTX 24GB Local LLM: Runs 32B at est. 34 tok/s, https://modelfit.io/gpu/rx-7900-xtx/, updated November 2026, CC BY 4.0.
| Model size | Est. speed | Fit on 24GB |
|---|---|---|
| 7B | est. 115 tok/stypical error 10% | Fits in VRAM |
| 14B | est. 69 tok/stypical error 10% | Fits in VRAM |
| 20B MoE (3.6B active) | est. 129 tok/stypical error 25% | Fits in VRAM |
| 32B | est. 34 tok/stypical error 50% | Fits in VRAM |
| 35B MoE (3B active) | est. 108 tok/srough estimate, offload not measured | CPU offload (slow) |
| 70B | est. 2 tok/srough estimate, offload not measured | CPU offload (slow) |
| 120B MoE (5.1B active) | est. 36 tok/srough estimate, offload not measured | CPU offload (slow) |
Decode speed, generic Q4_K_M model of each size at 4K context, 960 GB/s, typical error 10-50%
ModelFit estimates, not measured benchmarks: decode speed of a generic Q4_K_M model of each size at 4K context on the RX 7900 XTX's 960 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 24GB 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 ~22 GB usable VRAM), and at 64k it adds ~4.0 GB (still fits).
The model leaves about 4.1 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 24 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.
24GB GDDR6 at 960 GB/s puts the RX 7900 XTX in the same capacity class as the RTX 3090 and RTX 4090: 32B models at Q4 (~20GB) load fully in VRAM with room for context. RDNA 3 (gfx1100) is officially supported by ROCm, which Ollama uses on Linux and Windows; llama.cpp also runs it via ROCm/HIP or Vulkan. Where the card really stands out is dual-card rigs: two used 7900 XTX give 48GB of pooled VRAM, enough to hold a 35B-class MoE at Q8 with its full context, usually for less than an NVIDIA pair of the same capacity. The trade-off versus NVIDIA is software maturity outside llama.cpp: some backends (vLLM, ExLlama) still favor CUDA.
| Hardware | Memory | Speed | Bandwidth |
|---|---|---|---|
| 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 |
| RTX 4090 | 24 GB | est. 133 tok/stypical error 5% | 1008 GB/s |
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 44 tok/s, typical error 50% · first token ~0.9s
Fits in 24 GB VRAM with room to spare. Best for coding, agent, vision, long context on RX 7900 XTX.
Best for: Chat, Coding, Multimodal·Pop: 86/100
Perf: est. 126 tok/s, typical error 25% · first token ~0.2s
Fits in 24 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 7900 XTX.
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 52 tok/s, typical error 10% · first token ~2.0s
Fits in 24 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 7900 XTX.
Best for: Agentic, Coding, Long context·Pop: 78/100
Perf: est. 145 tok/s, typical error 25% · first token ~0.3s
Fits in 24 GB VRAM with room to spare. Best for agentic, coding, long context on RX 7900 XTX.
Best for: Reasoning, Coding, Agent scenarios·Pop: 90/100
Perf: est. 168 tok/s, typical error 25% · first token ~0.2s
Fits in 24 GB VRAM with room to spare. Best for reasoning, coding, agent scenarios on RX 7900 XTX.
Best for: Local AI agents, privacy-first tool calling, MCP workflows·Pop: 80/100
Perf: est. 176 tok/s, typical error 25% · first token ~0.2s
Fits in 24 GB VRAM with room to spare. Best for local ai agents, privacy-first tool calling, mcp workflows on RX 7900 XTX.
Best for: Chat, Agents, Vision·Pop: 60/100
Perf: est. 43 tok/s, typical error 50% · first token ~0.9s
Fits in 24 GB VRAM with room to spare. Best for chat, agents, vision on RX 7900 XTX.
Best for: Agents, Deep research, Coding·Pop: 60/100
Perf: est. 43 tok/s, typical error 50% · first token ~0.9s
Fits in 24 GB VRAM with room to spare. Best for agents, deep research, coding on RX 7900 XTX.
Best for: Coding, Agentic tasks, Reasoning·Pop: 78/100
Perf: est. 132 tok/s, typical error 25% · first token ~0.3s
Fits in 24 GB VRAM with room to spare. Best for coding, agentic tasks, reasoning on RX 7900 XTX.
Best for: Agentic coding, Long-horizon tasks·Pop: 72/100
Perf: est. 160 tok/s, typical error 25% · first token ~0.2s
Fits in 24 GB VRAM with room to spare. Best for agentic coding, long-horizon tasks on RX 7900 XTX.
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 32B parameter models at Q4 quantization. Its 24GB VRAM gives about 21.6GB usable for weights, the same capacity class as the RTX 3090 and RTX 4090. Good picks are Qwen3.6 27B and DeepSeek-R1 32B. 35B-class MoE models at Q4 (~20-22GB) sit right at the limit: Qwen3.5 35B-A3B fits, Qwen3.6 35B-A3B may spill slightly to system RAM.
Yes. Ollama supports the RX 7900 XTX through ROCm on Linux and Windows, since RDNA 3 (gfx1100) is on the official ROCm support list. llama.cpp and LM Studio also run it via ROCm/HIP or the Vulkan backend. No CUDA is needed.
Yes. Two cards pool 48GB of VRAM, and a 35B-class MoE at Q8 weighs about 38.7GB. Hybrid linear-attention models like Qwen3.6 35B-A3B cache KV only on their sparse full-attention layers, roughly 20 KB per token, so even the full 262k context adds only ~5GB. Weights plus full context fit inside the two cards with no offloading.
They are close. Both have 24GB and run the same 32B-class models; the 7900 XTX has slightly higher bandwidth (960 vs 936 GB/s) and tends to cost less used, while the 3090 is estimated 13% faster on 8B models (est. 117 tok/s, typical error 5% vs est. 104 tok/s, typical error 10%). The 3090 also keeps the edge in software ecosystem: CUDA-only backends and broader community tooling. For an Ollama/llama.cpp workflow, either works well.
Not for the main runtimes. Ollama, llama.cpp, and LM Studio all support RDNA 3 through ROCm or Vulkan, and the 7900 XTX is on the official ROCm support list. Friction remains in CUDA-first projects like some vLLM and training workflows, so check your specific stack before buying.
ModelFit estimates a 32B model on the RX 7900 XTX at Q4_K_M runs at est. 34 tok/s, typical error 50%. The current 27B-class pick in the catalog is Qwen3.8 27B (ollama run qwen3.8:27b).
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.