Your rig

Radeon RX 6900 XT · 16 GB

256 GB system RAM · RDNA 2 · 512 GB/s · 14.4 GB usable VRAM
Edit hardware Export as JSON

Context length

KV cache at f16, 8K tokens

Quality floor
Q8_0 — near-lossless Q6_K Q5_K_M Q4_K_M — recommended Q3_K_M Q2_K — damaged
KV cache type
System RAM (GB)
Runtime
Ollamallama.cppLM StudiovLLM

4 of 5 runtimes drive this device.

Allow CPU offload Include partial-offload fits

Offloaded layers run at 60 GB/s host bandwidth — expect single-digit tokens per second.

Which one should I actually run?

Best models for your Radeon RX 6900 XT

Updated 8 Oct 2026 1 new this month

What do you want to do? Each row is the best-ranked model on that use case's shortlist that fits at Q8_0 and 8K. The ordering is editorial; the sizes and speeds are computed. How the shortlists work.

RecommendationModelWhyQuant · totalTok/s est.
Best overall Qwen3.5 9B
Alibaba · Apache 2.0
The 8–12 GB default of 2026 — beats every older 8B, with vision. Q8_0 · 10.4 GB 32 Runs great
Best for coding Qwen3.5 9B
Alibaba · Apache 2.0
The best coding model for 8–12 GB cards. Q8_0 · 10.4 GB 32 Runs great
Best reasoning gpt-oss 20B
OpenAI · Apache 2.0
Low/medium/high reasoning effort in 13 GB. MXFP4 · 11.6 GB 64 Runs great
Best for writing Ministral 3 8B
Mistral AI · Apache 2.0
Mistral's voice in 8B, Apache 2.0. Q8_0 · 10.5 GB 35 Runs great
Best vision Qwen3.5 9B
Alibaba · Apache 2.0
Vision on an 8–12 GB card. Q8_0 · 10.4 GB 32 Runs great
Best for agents gpt-oss 20B
OpenAI · Apache 2.0
Harmony-format tool calling, native 4-bit. MXFP4 · 11.6 GB 64 Runs great
Best translation Qwen3.5 9B
Alibaba · Apache 2.0
Strong CJK and European translation in 8–12 GB. Q8_0 · 10.4 GB 32 Runs great
Fastest good model gpt-oss 20B
OpenAI · Apache 2.0
3.6B active, native 4-bit. MXFP4 · 11.6 GB 64 Runs great
Best long context Qwen3.5 4B
Alibaba · Apache 2.0
262K on 8 GB. Q8_0 · 9.2 GB
at 128K context
67 Runs great

Everything that fits

Runs great

30 of 80 · fits with headroom for a longer prompt
ModelParamsQuantWeights +KV 8KTotal VRAM headroom Tok/sRuntime
gpt-oss 20B 20.9B MoE / 3.6B act MXFP4 10.8 GB 0.19 GB 11.6 GB
2.8 GB free
64 Ollama Run
Qwen3.5 9B 9.65B Q8_0 9.5 GB 0.25 GB 10.4 GB
4.0 GB free
32 Ollama Run
Ornith 1.5 9B 9.41B Q8_0 9.3 GB 0.25 GB 10.2 GB
4.2 GB free
33 Ollama Run
Ministral 3 8B 8.92B Q8_0 8.8 GB 1.06 GB 10.5 GB
3.9 GB free
35 Ollama Run
Granite 4.1 8B 8.79B Q8_0 8.7 GB 1.25 GB 10.5 GB
3.9 GB free
36 Ollama Run
LFM2.5 8B-A1B 8.47B MoE / 1.5B act Q8_0 8.4 GB 0.09 GB 9.1 GB
5.3 GB free
80 Ollama Run
Fara 7B 8.29B Q8_0 8.2 GB 0.44 GB 9.2 GB
5.2 GB free
38 llama.cpp Run
Qwen3 8B 8.19B Q8_0 8.1 GB 1.13 GB 9.8 GB
4.6 GB free
38 Ollama Run
Llama 3.1 8B Instruct 8.03B Q8_0 7.9 GB 1.00 GB 9.5 GB
4.9 GB free
39 Ollama Run
Gemma 4 E4B 8.0B Q8_0 7.9 GB 0.14 GB 8.7 GB
5.7 GB free
39 Ollama Run
Ling 3.0 Tiny 7.9B-A1.3B 7.9B MoE / 1.3B act Q8_0 7.8 GB 0.05 GB 8.5 GB
5.9 GB free
93 llama.cpp Run
DeepSeek-R1-Distill-Qwen 7B 7.62B Q8_0 7.5 GB 0.44 GB 8.6 GB
5.8 GB free
41 Ollama Run
Qwen2.5-Coder 7B 7.62B Q8_0 7.5 GB 0.44 GB 8.6 GB
5.8 GB free
41 Ollama Run
Olmo 3 7B Instruct 7.3B Q8_0 7.2 GB 2.50 GB 10.3 GB
4.1 GB free
43 Ollama Run
Mistral 7B Instruct v0.3 7.25B Q8_0 7.2 GB 1.00 GB 8.8 GB
5.6 GB free
43 Ollama Run
Gemma 4 E2B 5.1B Q8_0 5.0 GB 0.07 GB 5.7 GB
8.7 GB free
61 Ollama Run
Qwen3.5 4B 4.66B Q8_0 4.6 GB 0.25 GB 5.5 GB
8.9 GB free
67 Ollama Run
Gemma 3 4B 4.3B Q8_0 4.3 GB 0.30 GB 5.2 GB
9.2 GB free
73 Ollama Run
Qwen3 4B 4.02B Q8_0 4.0 GB 1.13 GB 5.7 GB
8.7 GB free
78 Ollama Run
Ministral 3 3B 3.85B Q8_0 3.8 GB 0.81 GB 5.2 GB
9.2 GB free
81 Ollama Run
Phi-4-mini 3.8B 3.84B Q8_0 3.8 GB 1.00 GB 5.4 GB
9.0 GB free
82 Ollama Run
Granite 4.1 3B 3.4B Q8_0 3.4 GB 0.63 GB 4.6 GB
9.8 GB free
92 Ollama Run
Llama 3.2 3B Instruct 3.21B Q8_0 3.2 GB 0.88 GB 4.7 GB
9.7 GB free
98 Ollama Run
LFM2.5 2.6B 2.7B Q8_0 2.7 GB 0.13 GB 3.4 GB
11.0 GB free
116 llama.cpp Run
Qwen3.5 2B 2.27B Q8_0 2.2 GB 0.09 GB 2.9 GB
11.5 GB free
138 Ollama Run
Qwen3 1.7B 1.72B Q8_0 1.7 GB 0.88 GB 3.2 GB
11.2 GB free
182 Ollama Run
Llama 3.2 1B Instruct 1.24B Q8_0 1.2 GB 0.25 GB 2.1 GB
12.3 GB free
253 Ollama Run
Gemma 3 1B 1.0B Q8_0 1.0 GB 0.04 GB 1.6 GB
12.8 GB free
313 Ollama Run
Qwen3.5 0.8B 0.87B Q8_0 0.9 GB 0.09 GB 1.6 GB
12.8 GB free
360 Ollama Run
Qwen3 0.6B 0.6B Q8_0 0.6 GB 0.88 GB 2.1 GB
12.3 GB free
522 Ollama Run

Runs — tight

3 models · fits at 8K, drop context or quant to go longer
ModelParamsQuantWeights +KV 8KTotal VRAM headroom Tok/sRuntime
Gemma 3 12B 12.2B Q8_0 12.1 GB 0.81 GB 13.5 GB
0.9 GB free
26 llama.cpp Run
Mistral NeMo 12B 12.2B Q8_0 12.1 GB 1.25 GB 13.9 GB
0.5 GB free
26 llama.cpp Run
Gemma 4 12B 12B Q8_0 11.9 GB 0.81 GB 13.3 GB
1.1 GB free
26 llama.cpp Run

Won't fit in VRAM

47 models · some are possible with CPU offload
ModelParamsQuantWeights +KV 8KTotal Over budget Tok/sRuntime
Qwen3.8-Flash-Next 180B-A6B 180B MoE / 6B act Q8_0 178.1 GB 0.19 GB 178.9 GB +164.5 GB ~2.5 llama.cpp Why
Qwen3.5 122B-A10B 125B MoE / 10B act Q8_0 123.7 GB 0.19 GB 124.5 GB +110.1 GB ~1.6 llama.cpp Why
Nemotron 3 Super 120B-A12B 124B MoE / 12B act Q8_0 122.7 GB 0.06 GB 123.4 GB +109.0 GB ~1.3 llama.cpp Why
Ling 3.0 Flash 124B-A5B 124B MoE / 5.1B act Q8_0 122.7 GB 0.06 GB 123.4 GB +109.0 GB ~3.1 llama.cpp Why
Mistral Small 4 119B-A6B 119B MoE / 6.5B act Q8_0 117.8 GB 0.18 GB 118.5 GB +104.1 GB ~2.4 llama.cpp Why
gpt-oss 120B 117B MoE / 5.1B act MXFP4 60.5 GB 0.29 GB 61.4 GB +47.0 GB ~6.6 llama.cpp Why
Llama 4 Scout 109B-A17B 109B MoE / 17B act Q8_0 107.9 GB 1.50 GB 110.0 GB +95.6 GB ~0.9 llama.cpp Why
Llama 3.3 70B Instruct 70.6B Q8_0 69.9 GB 2.50 GB 73.0 GB +58.6 GB ~0.6 llama.cpp Why
DeepSeek-R1-Distill-Llama 70B 70.6B Q8_0 69.9 GB 2.50 GB 73.0 GB +58.6 GB ~0.6 llama.cpp Why
Qwen3.6 35B-A3B 35.9B MoE / 3.3B act Q8_0 35.5 GB 0.16 GB 36.3 GB +21.9 GB ~6.5 llama.cpp Why
Ornith 1.5 35B-A3B 35.9B MoE / 3.3B act Q8_0 35.5 GB 0.16 GB 36.3 GB +21.9 GB ~6.5 llama.cpp Why
LLM-jp 4 33B Thinking 33.2B Q8_0 32.9 GB 2.00 GB 35.5 GB +21.1 GB ~1.6 llama.cpp Why
Qwen3 32B 32.8B Q8_0 32.5 GB 2.00 GB 35.1 GB +20.7 GB ~1.6 llama.cpp Why
DeepSeek-R1-Distill-Qwen 32B 32.8B Q8_0 32.5 GB 2.00 GB 35.1 GB +20.7 GB ~1.6 llama.cpp Why
Qwen2.5-Coder 32B 32.8B Q8_0 32.5 GB 2.00 GB 35.1 GB +20.7 GB ~1.6 llama.cpp Why
Olmo 3.1 32B Instruct 32.2B Q8_0 31.9 GB 1.25 GB 33.7 GB +19.3 GB ~1.7 llama.cpp Why
Nemotron 3.5 Lightning 30B-A3B 31.6B MoE / 3.2B act Q8_0 31.3 GB 0.05 GB 31.9 GB +17.5 GB ~7.2 llama.cpp Why
Gemma 4 31B 31.3B Q8_0 31.0 GB 2.03 GB 33.6 GB +19.2 GB ~1.8 llama.cpp Why
GLM-4.7-Flash 30B-A3B 31.2B MoE / 3B act Q8_0 30.9 GB 0.41 GB 31.9 GB +17.5 GB ~7.6 llama.cpp Why
Qwen3 30B-A3B 30.5B MoE / 3.3B act Q8_0 30.2 GB 0.75 GB 31.5 GB +17.1 GB ~6.9 llama.cpp Why
Qwen3 Coder 30B-A3B 30.5B MoE / 3.3B act Q8_0 30.2 GB 0.75 GB 31.5 GB +17.1 GB ~6.9 llama.cpp Why
Muse Glimmer 30B 29.8B Q8_0 29.5 GB 0.18 GB 30.3 GB +15.9 GB ~2.1 llama.cpp Why
Granite 4.1 30B 28.9B Q8_0 28.6 GB 2.00 GB 31.2 GB +16.8 GB ~2.0 llama.cpp Why
Qwen3.8 27B 27.8B Q8_0 27.5 GB 0.50 GB 28.6 GB +14.2 GB ~2.3 llama.cpp Why
Qwen3.6 27B 27.8B Q8_0 27.5 GB 0.50 GB 28.6 GB +14.2 GB ~2.3 llama.cpp Why
Gemma 3 27B 27.4B Q8_0 27.1 GB 1.03 GB 28.7 GB +14.3 GB ~2.3 llama.cpp Why
Gemma 4 26B-A4B 26.5B MoE / 3.8B act Q8_0 26.2 GB 0.51 GB 27.3 GB +12.9 GB ~6.7 llama.cpp Why
Devstral Small 2 24B 24B Q8_0 23.7 GB 1.25 GB 25.6 GB +11.2 GB ~2.9 llama.cpp Why
Mistral Small 3.2 24B 23.6B Q8_0 23.4 GB 1.25 GB 25.2 GB +10.8 GB ~3.0 llama.cpp Why
Qwen3 14B 14.8B Q8_0 14.6 GB 1.25 GB 16.5 GB +2.1 GB ~10 llama.cpp Why
DeepSeek-R1-Distill-Qwen 14B 14.8B Q8_0 14.6 GB 1.50 GB 16.7 GB +2.3 GB ~9.6 llama.cpp Why
Qwen2.5-Coder 14B 14.8B Q8_0 14.6 GB 1.50 GB 16.7 GB +2.3 GB ~9.6 llama.cpp Why
Phi-4 14B 14.7B Q8_0 14.5 GB 1.56 GB 16.7 GB +2.3 GB ~9.7 llama.cpp Why
Ministral 3 14B 13.9B Q8_0 13.8 GB 1.25 GB 15.6 GB +1.2 GB ~14 llama.cpp Why
Kimi K3 2.8T-A104B 2780B MoE / 104B act Q8_0 2750.9 GB 0.21 GB 2751.7 GB +2737.3 GB ~0.1 llama.cpp Why
Qwen3.8 2.4T-A95B 2446B MoE / 95B act Q8_0 2420.4 GB 0.72 GB 2421.7 GB +2407.3 GB ~0.1 llama.cpp Why
DeepSeek V4 Pro 1.6T-A49B 1650B MoE / 49B act Q8_0 1632.7 GB 0.54 GB 1633.9 GB +1619.5 GB ~0.3 llama.cpp Why
Kimi K2.6 1T-A32B 1027B MoE / 32B act Q8_0 1016.2 GB 0.54 GB 1017.4 GB +1003.0 GB ~0.4 llama.cpp Why
GLM-5.2 744B-A40B 753B MoE / 40B act Q8_0 745.1 GB 0.69 GB 746.4 GB +732.0 GB ~0.4 llama.cpp Why
GLM-5.3 744B-A40B 753B MoE / 40B act Q8_0 745.1 GB 0.69 GB 746.4 GB +732.0 GB ~0.4 llama.cpp Why
DeepSeek-R1 671B 671B MoE / 37B act Q8_0 664.0 GB 0.54 GB 665.1 GB +650.7 GB ~0.4 llama.cpp Why
DeepSeek V4.1 Flash 552B-A16B New 552B MoE / 16B act Q8_0 546.2 GB 0.04 GB 546.9 GB +532.5 GB ~0.9 — Why
Qwen3.5 397B-A17B 403B MoE / 17B act Q8_0 398.8 GB 0.23 GB 399.6 GB +385.2 GB ~0.9 llama.cpp Why
Ornith 1.5 397B-A17B 397B MoE / 17B act Q8_0 392.8 GB 0.23 GB 393.7 GB +379.3 GB ~0.9 llama.cpp Why
GLM-5.3-Flash 320B-A18B 321B MoE / 18B act Q8_0 317.6 GB 0.09 GB 318.3 GB +303.9 GB ~0.8 llama.cpp Why
DeepSeek V4 Flash 284B-A13B 284B MoE / 13B act Q8_0 281.0 GB 0.38 GB 282.0 GB +267.6 GB ~1.1 llama.cpp Why
Qwen3 235B-A22B 235B MoE / 22B act Q8_0 232.5 GB 1.47 GB 234.6 GB +220.2 GB ~0.7 llama.cpp Why

Totals = quantised weights + f16 KV cache at 8K tokens + 0.6 GB runtime overhead, against 14.4 GB usable VRAM (16 GB less display and driver reserve). Tokens per second are estimated from 512 GB/s peak memory bandwidth at batch 1 — not measured. Read the method.


Same memory, different speed

10 of the 19 devices have the same 14.4 GB usable, so they run exactly the models above. Generation speed scales with memory bandwidth.

DeviceBandwidthSpeed vs yours
GeForce RTX 5080
Blackwell · 2025
960 GB/s 1.88×
GeForce RTX 5070 Ti
Blackwell · 2025
896 GB/s 1.75×
CMP 100-210
Volta · 2021
829 GB/s 1.62×
GeForce RTX 4080 Super
Ada Lovelace · 2024
736 GB/s 1.44×
Tesla P100 16 GB
Pascal · 2016
732 GB/s 1.43×
GeForce RTX 4080
Ada Lovelace · 2022
717 GB/s 1.40×
GeForce RTX 4070 Ti Super
Ada Lovelace · 2024
672 GB/s 1.31×
Radeon RX 9070 XT
RDNA 4 · 2025
645 GB/s 1.26×
Radeon RX 9070
RDNA 4 · 2025
645 GB/s 1.26×
Radeon RX 7800 XT
RDNA 3 · 2023
624 GB/s 1.22×

Best LLM for 16 GB VRAM, with every card in the class →

What more memory unlocks

At Q4_K_M and 8K context this device fits 38 of 80 models. Each step up, and the largest models it adds:

Every device ranked by what it can run →