Gemma 4 12B
Local OKExcellentA · 44
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 58 tok/s, typical error 10% · first token ~0.4s
Fits in 16 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 9070 XT.
Best Local AI Models for AMD RX 9070 XT (16GB)
The RX 9070 XT has 16GB of GDDR6 at 640 GB/s. That fits dense models up to 14B at Q4, with an 8B model at est. 79 tok/s, typical error 10% and a 14B model at est. 51 tok/s. Ollama runs it through ROCm on Linux, where it is on the official support list (gfx1201); its Windows ROCm list does not include the card, so on Windows Ollama uses its Vulkan backend.
Quick answer
The best local LLM for the RX 9070 XT is Gemma 4 12B (est. 58 tok/s, typical error 10%) on its 16GB VRAM. It uses ~8GB of VRAM; the RX 9070 XT handles dense models up to ~14B at Q4. A 14B model at Q4 runs at est. 51 tok/s, typical error 10%. (ModelFit, 2026)
Sizing rule: a Q4 model needs about 0.6 GB of VRAM per billion parameters, and ModelFit budgets 90% of this card's 16GB for weights, context, and KV-cache. The per-size table below uses that same budget. Other strong fits: Qwen3.5 9B Instruct (Q8) (9B, ~10.7GB) and Mellum2.1 12B-A2.5B Thinking (12.1B, ~8.1GB). On this card Gemma 4 12B runs at est. 58 tok/s, typical error 10%. 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.
$ollama run gemma4:12b
- Top pick
- Gemma 4 12B
- Decode speed
- est. 58 tok/s, typical error 10%
- VRAM needed
- ~8 GB
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 9070 XT 16GB Local LLM: ROCm on Linux, 14B at est. 51 tok/s, https://modelfit.io/gpu/rx-9070-xt/, updated October 2026, CC BY 4.0.
Specs
VRAM
16 GB GDDR6
Speed (8B Q4)
est. 79 tok/stypical error 10%
Bandwidth
640 GB/s
Architecture
RDNA 4
Max model size
Up to ~14B dense
Compatibility
10 excellent, 0 workable
Table 1
RX 9070 XT estimated tokens/sec by model size
Q4_K_M · ModelFit estimate| Model size | Est. speed | Fit on 16GB |
|---|---|---|
| 7B | est. 88 tok/stypical error 10% | Fits in VRAM |
| 14B | est. 51 tok/stypical error 10% | Fits in VRAM |
| 20B MoE (3.6B active) | est. 99 tok/stypical error 25% | Fits in VRAM |
| 32B | est. 5 tok/srough estimate, offload not measured | CPU offload (slow) |
| 35B MoE (3B active) | est. 39 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. 27 tok/srough estimate, offload not measured | CPU offload (slow) |
Fig. 1RX 9070 XT: estimated speed by model size
Decode speed, generic Q4_K_M model of each size at 4K context, 640 GB/s, typical error 10-25%
Fits in 16 GBCPU offload (slow)
7B88
14B51
20B MoE (3.6B active)99
32B5
35B MoE (3B active)39
70B1
120B MoE (5.1B active)27
tok/s est.ModelFit estimates, not measured benchmarks: decode speed of a generic Q4_K_M model of each size at 4K context on the RX 9070 XT's 640 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 16GB 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.
Fig. 2
Context costs VRAM too.
Gemma 4 12B loads ~8 GB of weights; at 16k context the KV cache adds ~3.0 GB (still fits the ~14 GB usable VRAM), and at 64k it adds ~12.0 GB (exceeds the budget, use a smaller quant or a q8_0 KV cache).
Weights 8.0 GBKV cache · 16k context 3.0 GBHeadroom 3.4 GBRuntime reserve 1.6 GB
The model leaves about 3.4 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.
Get told when a model beats Gemma 4 12B on your RX 9070 XT
One email when a new open-weight model is a better fit for 16 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.
Where to buy the RX 9070 XT
Storage & accessories for your model library
- Internal NVMe SSD · 2TBA Gen4 M.2 drive keeps your whole GGUF and quant collection on fast local storage, loading models straight off NVMe.Check price on Amazon
- USB4 NVMe Enclosure40Gbps external storage fast enough to run models from. Pair it with an M.2 drive for a portable model vault.Check price on Amazon
ModelFit may earn a commission on purchases through these links, at no extra cost to you.
RX 9070 XT VRAM for AI: what actually fits?
ModelFit budgets about 14.4GB of the 16GB for weights and context: 14B models at Q4 (~11GB) fit with room for context, and 20B-class mixture-of-experts models load too. Dense 27B and 32B models spill into system RAM. Against the RTX 5070 Ti (16GB at 896 GB/s) it has less bandwidth and runs est. 79 tok/s, typical error 10% against est. 125 tok/s, typical error 5% on 8B. For 27B models fully in VRAM, the 20GB RX 7900 XT is the AMD step up.
Table 2
RX 9070 XT vs top GPUs
| Hardware | Memory | Speed | Bandwidth |
|---|---|---|---|
| RTX 5060 Ti | 16 GB | est. 68 tok/stypical error 5% | 448 GB/s |
| RX 9070 XT | 16 GB | est. 79 tok/stypical error 10% | 640 GB/s |
| 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 |
Recommended models
10 modelsQwen3.5 9B Instruct (Q8)
Local OKOKB · 26
Best for: Quality, Coding, Reasoning·Pop: 86/100
Perf: est. 56 tok/s, typical error 10% · first token ~1.0s
Fits in 16 GB VRAM with room to spare. Best for quality, coding, reasoning on RX 9070 XT.
Mellum2.1 12B-A2.5B Thinking
Local OKExcellentA · 44
Best for: Agentic coding, Small machines·Pop: 76/100
Perf: est. 131 tok/s, typical error 25% · first token ~0.2s
Fits in 16 GB VRAM with room to spare. Best for agentic coding, small machines on RX 9070 XT.
Gemma 4 E4B (Q8)
Local OKExcellentA · 48
Best for: On-device, Mobile, Chat·Pop: 82/100
Perf: est. 84 tok/s, typical error 10% · first token ~0.9s
Fits in 16 GB VRAM with room to spare. Best for on-device, mobile, chat on RX 9070 XT.
NVIDIA Nemotron 3.5 Bolt E4B
Local OKExcellentA · 45
Best for: Voice assistants, Tool calling, Chat·Pop: 75/100
Perf: est. 100 tok/s, typical error 10% · first token ~0.3s
Fits in 16 GB VRAM with room to spare. Best for voice assistants, tool calling, chat on RX 9070 XT.
Gemma 4 12B (Q8)
Local OKOKB · 11
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 38 tok/s, typical error 10% · first token ~1.3s
Fits in 16 GB VRAM with room to spare. Best for chat, coding, multimodal on RX 9070 XT.
Ministral 3 14B
Local OKExcellentA · 37
Best for: Chat, Reasoning, Vision·Pop: 74/100
Perf: est. 52 tok/s, typical error 10% · first token ~0.4s
Fits in 16 GB VRAM with room to spare. Best for chat, reasoning, vision on RX 9070 XT.
Qwen3.5 9B Instruct
Local OKExcellentA · 51
Best for: Quality, Coding, Reasoning·Pop: 86/100
Perf: est. 83 tok/s, typical error 10% · first token ~0.3s
Fits in 16 GB VRAM with room to spare. Best for quality, coding, reasoning on RX 9070 XT.
LFM2 24B-A2B Instruct
Local OKOKB · 3
Best for: Local AI agents, privacy-first tool calling, MCP workflows·Pop: 80/100
Perf: est. 141 tok/s, typical error 25% · first token ~0.2s
Fits in 16 GB VRAM with room to spare. Best for local ai agents, privacy-first tool calling, mcp workflows on RX 9070 XT.
Ornith 1.0 9B
Local OKExcellentA · 61
Best for: Agentic coding on small machines·Pop: 76/100
Perf: est. 83 tok/s, typical error 10% · first token ~0.3s
Fits in 16 GB VRAM with room to spare. Best for agentic coding on small machines on RX 9070 XT.
Video and image models on the RX 9070 XT
No stated VRAM figure or published run covers AMD for these models yet. By memory size: video models · image models.
Compare with top GPUs
Compatible model families
- QwenAlibaba Cloud: Qwen3.8 27B local flagship (Apache 2.0, vision)
- LlamaMeta: Most popular open-weight model family
- DeepSeekDeepSeek AI: V4.1 Flash: multimodal 552B MoE, API only
- MistralMistral AI: Excellent performance-per-parameter ratio
- GemmaGoogle DeepMind: Gemma 4 E4B: 4.5B effective, runs on 8GB Macs
- PhiMicrosoft: Best quality-per-gigabyte at small sizes
RX 9070 XT FAQ: common questions
Does Ollama support the RX 9070 XT?
Yes. On Linux, Ollama lists the RX 9070 XT (gfx1201) among its ROCm cards and needs the AMD ROCm v7 driver. Its Windows ROCm list does not include the card, so on Windows Ollama runs it through its Vulkan backend, which needs no extra setup there.
How much VRAM does the RX 9070 XT have for LLMs?
The RX 9070 XT has 16GB of GDDR6 at up to 640 GB/s (AMD specifications). ModelFit budgets about 14.4GB of it for weights and context, enough for dense models up to 14B at Q4 quantization.
What size LLM can I run on an RX 9070 XT?
Dense models up to 14B at Q4, such as Qwen3 14B (~11GB) at est. 51 tok/s, typical error 10%, plus 20B-class mixture-of-experts models. 8B models run at est. 79 tok/s. Dense 27B and 32B models need a 20GB or 24GB card to stay in VRAM.
RX 9070 XT vs RTX 5070 Ti for local AI?
Both have 16GB, so they hold the same models. The RTX 5070 Ti streams 896 GB/s against 640 GB/s and runs CUDA, so it is 58% faster on 8B models (est. 125 tok/s, typical error 5% vs est. 79 tok/s, typical error 10%) and works with CUDA-only tools. The 9070 XT covers the main runtimes, Ollama and llama.cpp, on ROCm or Vulkan.
How fast is a 27B-class model on the RX 9070 XT?
The RX 9070 XT's 16GB of VRAM cannot fit a 32B model comfortably. The largest size class it fits is 20B MoE (3.6B active), at est. 99 tok/s, typical error 25%. The largest individual model in the catalog that fits is LFM2 24B-A2B Instruct (24B).
Related guides & benchmarks
More local AI guides
- Best LLM for 16GB VRAMEvery 16 GB card side by side, each with its best pick and speed.
- How Much RAM (or VRAM) Do You Need for a Local LLM?The model-size-to-memory matrix: what each VRAM and RAM tier actually runs.
- Best LLM for MacBook (Apple Silicon)Unified-memory Macs run bigger models per dollar. Picks by RAM tier, M1 to M5.
- Local AI image generator on MacCompare image models by RAM and app.
- Local image generation on MacImage models and apps by Mac RAM.
Browse all NVIDIA GPUs for AI
Real-world calibration
Have this hardware? Share a measured run.
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.
$ npx @modelfit/cli@latest bench
Want personalized recommendations?
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.