DeskGPU
English

AMD Radeon RX 7900 XT: which LLMs can it run?

This page was written and translated automatically by AI and checked by software against its sources, with no human editing. Specifications come from the sources listed; memory requirements and speeds are estimates calculated with published formulas. How we use AI

AMD Radeon RX 7900 XT has 20 GB of memory (manufacturer figure). Below, every model in the database is checked against it with llama.cpp at a context length of 8,192 tokens (or the model’s maximum if shorter).

Values updated Oct 19, 2026

Specifications

Memory (manufacturer figure)
20 GB
Memory type
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
Memory bus width
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
Memory bandwidth
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
Board power (TDP)
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
FP16 tensor, dense
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
BF16 tensor, dense
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
FP8 tensor, dense
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
FP4 tensor, dense
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
INT8 tensor, dense
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.
Interface
Not shown: open sources do not give two independent, agreeing values with a manufacturer reference.

Each value lists every source with its own figure; units have been normalized. A manufacturer’s “GB” of GPU memory is counted as GiB (2³⁰ bytes), which is also how AMD’s ROCm table states it. How specifications are checked

Run profile

Run profile: llama.cpp, ROCm back end on Linux (AMD discrete GPU), all layers on the GPU, Flash Attention on, micro-batch (-ub) 512, F16 KV cache, context (-c) as shown; “Fits” is judged with llama-server’s default 4 parallel sequences (-np auto, unified KV cache) and “Doesn’t fit” with 1 sequence; idle GPU with no other programs using its memory.

Memory counted as available for “Fits” (C_low): 18.63 GiB per card = 20 GiB nominal − 144 MiB (the largest gap between nominal and reported total measured on consumer and workstation GPUs in this profile, used because no public log reports this GPU’s own total) − 1,216 MiB (the largest start-up and runtime use seen in this profile’s public logs), rounded down to a multiple of 64 MiB.

These results assume the card’s default configuration. Settings that reduce usable memory (for example ECC, where the card supports it) are not covered.

Memory used for “Doesn’t fit” (C_high): the full 20 GiB per card. Run profiles and their measurements

Which models fit

“Fits” means llama.cpp's default --fit load check would place all layers on the GPU: our estimate of llama.cpp's buffers, plus its error margin, plus llama.cpp's 1 GiB per-card reserve, is within the memory counted as available. It does not promise that peak memory while generating fits. How results are decided

No speed ceiling: the bandwidth could not be confirmed by two open sources.

Context 8,192 tokens (or the model’s maximum if shorter). The last columns show the longest context at that quantization and the bandwidth-based speed ceiling.
Model Parameters Highest quantization that fits Available memory needed for “Fits” Longest context Speed ceiling Result
AI21-Jamba2-Mini 51.6B (11.8B active) IQ2_M 17.3 GiB 81,920 — Fits
Kimi-Linear-48B-A3B-Base 49.1B (3.1B active) IQ2_M 17.2 GiB 148,224 — Fits
Kimi-Linear-48B-A3B-Instruct 49.1B (3.1B active) IQ2_M 16.9 GiB 181,760 — Fits
Hermes-4.3-36B 36.2B IQ3_XXS 16.9 GiB 14,592 — Fits
Kimi-Linear-REAP-35B-A3B-Instruct 35.1B (3.1B active) Q3_K_M 17.9 GiB 78,336 — Fits
Holo-3.1-35B-A3B 34.7B (2.9B active) Q3_K_M 18.1 GiB 29,696 — Fits
Holo3-35B-A3B 34.7B (2.9B active) Q3_K_M 18.1 GiB 29,696 — Fits
Holo4-35B-A3B 34.7B (2.9B active) Q3_K_M 18.4 GiB 18,176 — Fits
Ising-Calibration-1-35B-A3B 34.7B (2.9B active) Q3_K_M 18.1 GiB 29,696 — Fits
KAT-Coder-V2.5-Dev 34.7B (2.9B active) Q3_K_M 17.6 GiB 53,504 — Fits
Qwen3.5-35B-A3B 34.7B (2.9B active) Q3_K_M 17.6 GiB 56,576 — Fits
Qwen3.5-35B-A3B-Base 34.7B (2.9B active) Q3_K_M 18.1 GiB 29,952 — Fits
Qwen3.6-35B-A3B 34.7B (2.9B active) Q3_K_M 17.6 GiB 53,760 — Fits
Laguna-XS-2.1 33.4B (2.8B active) Q3_K_M 17.1 GiB 44,288 — Fits
DeepSeek-R1-Distill-Qwen-32B 32.8B Q2_K 15.2 GiB 21,248 — Fits
Qwen2.5-32B-Instruct 32.8B Q2_K 15.2 GiB 21,248 — Fits
Qwen2.5-Coder-32B-Instruct 32.8B Q2_K 15.2 GiB 21,248 — Fits
AesCode-32B 32.8B IQ3_XXS 16.4 GiB 16,640 — Fits
Cosmos-Reason2-32B 32.8B IQ3_XXS 15.7 GiB 19,456 — Fits
FrogBoss-32B-2510 32.8B IQ3_XXS 15.7 GiB 19,456 — Fits
Nemotron-Terminal-32B 32.8B IQ3_XXS 15.7 GiB 19,456 — Fits
Qwen3-VL-32B-Instruct 32.8B Q2_K 15.3 GiB 20,992 — Fits
Qwen3-VL-32B-Thinking 32.8B Q2_K 15.3 GiB 20,992 — Fits
SERA-32B 32.8B IQ3_XXS 15.7 GiB 19,456 — Fits
WebWorld-32B 32.8B IQ3_XXS 15.7 GiB 19,456 — Fits
Olmo-3.1-32B-Instruct 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
Olmo-3.1-32B-Instruct-DPO 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
Olmo-3.1-32B-Think 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
Olmo-3-1125-32B 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
Olmo-3-32B-Think 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
Olmo-3-32B-Think-DPO 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
Olmo-3-32B-Think-SFT 32.2B Q3_K_M 18.5 GiB 8,704 — Fits
gemma-4-31B 30.7B Q2_K 17.6 GiB 20,736 — Fits
Ising-Calibration-1.5-31B-BF16 30.7B IQ3_XXS 18.6 GiB 8,192 — Fits
Holo2-30B-A3B 30.5B (3.0B active) IQ4_XS 18.2 GiB 12,544 — Fits
Qwen3-30B-A3B-Instruct-2507 30.5B (3.0B active) IQ4_XS 18.0 GiB 14,592 — Fits
Qwen3-Coder-30B-A3B-Instruct 30.5B (3.0B active) IQ4_XS 18.0 GiB 14,592 — Fits
Qwen3-VL-30B-A3B-Instruct 30.5B (3.0B active) IQ4_XS 18.0 GiB 14,592 — Fits
UI-Venus-1.5-30B-A3B 30.5B (3.0B active) IQ4_XS 18.2 GiB 12,544 — Fits
North-Mini-Code-1.0 30.5B (3.3B active) IQ4_XS 18.6 GiB 8,192 — Fits
GLM-4.7-Flash 29.9B (3.6B active) Q4_0 18.5 GiB 11,008 — Fits
granite-4.2-30b 29.3B Q3_K_M 17.1 GiB 14,080 — Fits
granite-4.1-30b 28.9B Q3_K_M 17.1 GiB 14,080 — Fits
Muse-Glimmer-30B 27.9B Q4_K_M 18.1 GiB 43,008 — Fits
gemma-2-27b-it 27.2B Q3_K_M 17.7 GiB 8,192 — Fits
gemma-3-27b-it 27.0B IQ4_XS 18.6 GiB 8,192 — Fits
translategemma-27b-it 27.0B Q3_K_M 17.3 GiB 24,064 — Fits
Fara1.5-27B 26.9B Q4_0 17.7 GiB 21,504 — Fits
Holo4-27B 26.9B Q4_K_M 18.2 GiB 15,104 — Fits
Qwen3.5-27B 26.9B Q4_K_M 18.4 GiB 12,032 — Fits
Qwen3.6-27B 26.9B Q4_K_M 18.4 GiB 11,008 — Fits
Qwen3.8-27B 26.9B Q4_0 17.7 GiB 21,504 — Fits
Qwen3.8-3.6-27B-blend 26.9B Q4_0 17.7 GiB 21,504 — Fits
UI-Mate-27B 26.9B Q4_0 17.7 GiB 21,504 — Fits
UI-Mate-democua-27B 26.9B Q4_K_M 18.2 GiB 15,104 — Fits
Trinity-Mini 26.1B (3.1B active) Q4_K_M 17.4 GiB 77,056 — Fits
Trinity-Mini-Base 26.1B (3.1B active) Q4_K_M 17.2 GiB 83,200 — Fits
Trinity-Mini-Base-Pre-Anneal 26.1B (3.1B active) Q4_K_M 17.0 GiB 4,096 — Fits
gemma-4-26B-A4B 25.2B (3.8B active) IQ4_XS 16.4 GiB 108,288 — Fits
gemma-4-26B-A4B-it 25.2B (3.8B active) Q4_0 17.1 GiB 77,312 — Fits
Qwen3-Coder-REAP-25B-A3B 24.9B (3.0B active) Q4_K_M 16.8 GiB 26,368 — Fits
LFM2.5-24B-A2B 24.0B (2.5B active) Q5_K_M 18.1 GiB 30,464 — Fits
LFM2-24B-A2B 23.8B (2.3B active) Q5_K_M 18.0 GiB 35,072 — Fits
Devstral-Small-2-24B-Instruct-2512 23.6B Q5_K_M 18.6 GiB 8,448 — Fits
Mistral-Small-3.2-24B-Instruct-2506 23.6B Q5_K_M 18.6 GiB 8,448 — Fits
GLM-4.7-Flash-REAP-23B-A3B 23.0B (3.6B active) Q5_K_M 17.7 GiB 24,832 — Fits
EuroLLM-22B-2512 22.6B Q5_K_M 18.2 GiB 9,728 — Fits
EuroLLM-22B-Instruct-2512 22.6B Q5_K_M 18.2 GiB 9,728 — Fits
gpt-oss-20b 20.9B (3.6B active) Q8_0 12.9 GiB 131,072 — Fits
OptiMind-SFT 20.9B (3.6B active) Q5_K_M 17.8 GiB 40,960 — Fits
DeepSeek-R1-Distill-Qwen-14B 14.8B Q8_0 17.5 GiB 14,080 — Fits
Qwen2.5-14B-Instruct 14.8B Q8_0 17.5 GiB 14,080 — Fits
ContextPilot-14B 14.8B Q8_0 17.2 GiB 16,640 — Fits
FrogMini-14B-2510 14.8B Q6_K 13.9 GiB 37,120 — Fits
JanusCoder-14B 14.8B Q8_0 17.2 GiB 16,640 — Fits
Nemotron-Cascade-14B-Thinking 14.8B Q8_0 17.2 GiB 16,640 — Fits
Nemotron-Terminal-14B 14.8B Q6_K 13.9 GiB 37,120 — Fits
NousCoder-14B 14.8B Q8_0 17.2 GiB 16,640 — Fits
Qwen3-14B 14.8B Q8_0 17.2 GiB 16,640 — Fits
Qwen3-Nemotron-14B-BRRM 14.8B Q6_K 13.9 GiB 37,120 — Fits
WebWorld-14B 14.8B Q6_K 13.9 GiB 37,120 — Fits
phi-4 14.7B Q8_0 17.7 GiB 12,800 — Fits
Apriel-1.6-15b-Thinker 14.4B Q8_0 17.2 GiB 15,360 — Fits
Ministral-3-14B-Instruct-2512 13.5B Q8_0 16.0 GiB 24,320 — Fits
Ministral-3-14B-Reasoning-2512 13.5B Q8_0 16.0 GiB 24,320 — Fits
Mistral-Nemo-Instruct-2407 12.2B Q8_0 14.7 GiB 32,512 — Fits
Mellum2.1-12B-A2.5B-Thinking 12.1B (2.2B active) Q8_0 14.0 GiB 131,072 — Fits
Mellum2-12B-A2.5B-Base 12.1B (2.2B active) Q6_K 12.1 GiB 131,072 — Fits
Mellum2-12B-A2.5B-Instruct 12.1B (2.2B active) Q8_0 14.0 GiB 131,072 — Fits
Mellum2-12B-A2.5B-Instruct-SFT 12.1B (2.2B active) Q6_K 12.1 GiB 131,072 — Fits
Mellum2-12B-A2.5B-Thinking 12.1B (2.2B active) Q8_0 14.0 GiB 131,072 — Fits
Mellum2-12B-A2.5B-Thinking-SFT 12.1B (2.2B active) Q6_K 12.1 GiB 131,072 — Fits
gemma-4-12B 11.9B Q8_0 15.6 GiB 177,152 — Fits
gemma-4-12B-it 11.9B Q8_0 15.6 GiB 177,152 — Fits
gemma-3-12b-it 11.8B Q8_0 15.8 GiB 50,432 — Fits
translategemma-12b-it 11.8B Q8_0 15.8 GiB 50,432 — Fits
Bielik-11B-v3.0-Instruct 11.2B Q8_0 14.4 GiB 28,672 — Fits
AutoGLM-Phone-9B 9.4B Q8_0 10.8 GiB 65,536 — Fits
AutoGLM-Phone-9B-Multilingual 9.4B Q6_K 9.2 GiB 65,536 — Fits
GLM-4.6V-Flash 9.4B Q8_0 10.8 GiB 131,072 — Fits
gemma-2-9b-it 9.2B Q8_0 14.8 GiB 8,192 — Fits
EuroLLM-9B-Instruct-2512 9.2B Q8_0 11.6 GiB 32,768 — Fits
Fara1.5-9B 9.0B BF16 17.5 GiB 39,680 — Fits
Holo-3.1-9B 9.0B F16 17.5 GiB 39,680 — Fits
MiMo-V2.6-Distill-Qwen-9B 9.0B BF16 17.5 GiB 39,680 — Fits
Qwen3.5-9B 9.0B BF16 17.5 GiB 39,680 — Fits
Qwen3.5-9B-Base 9.0B F16 17.5 GiB 39,680 — Fits
UI-Mate-9B 9.0B BF16 17.5 GiB 39,680 — Fits
UI-Venus-2-9B 9.0B BF16 17.5 GiB 39,680 — Fits
NVIDIA-Nemotron-Nano-9B-v2 8.9B BF16 18.3 GiB 28,160 — Fits
granite-4.1-8b 8.8B Q8_0 11.3 GiB 52,992 — Fits
granite-4.2-8b 8.8B Q8_0 11.3 GiB 52,992 — Fits
Ministral-3-8B-Instruct-2512 8.5B BF16 18.0 GiB 12,544 — Fits
Ministral-3-8B-Reasoning-2512 8.5B BF16 18.0 GiB 12,544 — Fits
LFM2.5-8B-A1B 8.5B (1.7B active) BF16 18.0 GiB 53,504 — Fits
granite-guardian-4.1-8b 8.4B Q8_0 11.3 GiB 52,992 — Fits
LFM2-8B-A1B 8.3B (1.6B active) F16 17.7 GiB 71,424 — Fits
rnj-1-instruct 8.3B Q8_0 11.0 GiB 32,768 — Fits
AesCode-8B 8.2B BF16 17.3 GiB 16,896 — Fits
ArmorOCR 8.2B Q8_0 10.4 GiB 64,256 — Fits
AstaBrief_8B 8.2B Q6_K 8.6 GiB 40,960 — Fits
ContextPilot-8B 8.2B Q6_K 8.6 GiB 40,960 — Fits
Cosmos-Reason2-8B 8.2B Q6_K 8.6 GiB 76,544 — Fits
DeepSeek-R1-0528-Qwen3-8B 8.2B BF16 17.3 GiB 16,896 — Fits
Holo2-8B 8.2B Q6_K 8.6 GiB 76,544 — Fits
intent-aware-lfqa-qwen3-8b-baseline 8.2B F16 17.3 GiB 16,896 — Fits
intent-aware-lfqa-qwen3-8b-intent-explicit 8.2B F16 17.3 GiB 16,896 — Fits
intent-aware-lfqa-qwen3-8b-intent-implicit 8.2B F16 17.3 GiB 16,896 — Fits
intent-aware-lfqa-qwen3-8b-multiview 8.2B F16 17.3 GiB 16,896 — Fits
JanusCoder-8B 8.2B BF16 17.3 GiB 16,896 — Fits
MathForm-8B 8.2B F16 17.3 GiB 16,896 — Fits
Nemotron-Cascade-8B 8.2B BF16 17.3 GiB 16,896 — Fits
Nemotron-Cascade-8B-Thinking 8.2B BF16 17.3 GiB 16,896 — Fits
Nemotron-Orchestrator-8B 8.2B BF16 17.3 GiB 16,896 — Fits
Nemotron-Terminal-8B 8.2B Q6_K 8.6 GiB 40,960 — Fits
Qwen3-8B 8.2B Q8_0 10.4 GiB 40,960 — Fits
Qwen3-Nemotron-8B-BRRM 8.2B Q6_K 8.6 GiB 40,960 — Fits
Qwen3-VL-8B-Instruct 8.2B BF16 17.3 GiB 16,896 — Fits
Qwen3-VL-8B-Thinking 8.2B BF16 17.3 GiB 16,896 — Fits
RLVR-8B-0926 8.2B Q6_K 8.6 GiB 76,544 — Fits
Step3-VL-10B-Base 8.2B Q6_K 8.6 GiB 40,960 — Fits
UI-Venus-1.5-8B 8.2B Q6_K 8.6 GiB 76,544 — Fits
WebWorld-8B 8.2B F16 17.3 GiB 16,896 — Fits
ZwZ-8B 8.2B Q6_K 8.6 GiB 76,544 — Fits
MiniCPM-o-4_5 8.2B F16 17.3 GiB 16,896 — Fits
granite-3.3-8b-instruct 8.2B F16 18.6 GiB 8,192 — Fits
granite-guardian-3.2-8b-factuality-detection 8.2B F16 18.6 GiB 8,192 — Fits
intent-aware-lfqa-llama3-8b-baseline 8.0B F16 17.0 GiB 20,224 — Fits
intent-aware-lfqa-llama3-8b-intent-explicit 8.0B F16 17.0 GiB 20,224 — Fits
intent-aware-lfqa-llama3-8b-intent-implicit 8.0B F16 17.0 GiB 20,224 — Fits
intent-aware-lfqa-llama3-8b-multiview 8.0B F16 17.0 GiB 20,224 — Fits
DeepSeek-R1-Distill-Llama-8B 8.0B BF16 17.0 GiB 20,224 — Fits
HARC-Llama-3.1-8B-Instruct 8.0B Q6_K 8.4 GiB 86,784 — Fits
Llama-3.1-8B-Instruct 8.0B BF16 17.0 GiB 20,224 — Fits
AprielGuard 7.9B Q6_K 7.9 GiB 117,248 — Fits
MiMo-Embodied-7B 7.6B Q6_K 8.1 GiB 79,616 — Fits
DeepSeek-R1-Distill-Qwen-7B 7.6B F16 15.6 GiB 59,136 — Fits
Fara-7B 7.6B BF16 15.6 GiB 59,136 — Fits
HARC-Qwen2.5-7B-Instruct 7.6B Q6_K 7.5 GiB 32,768 — Fits
JanusCoderV-7B 7.6B BF16 15.6 GiB 59,136 — Fits
olmOCR-2-7B-1025 7.6B BF16 15.6 GiB 59,136 — Fits
Qwen2.5-7B-Instruct 7.6B Q8_0 9.1 GiB 32,768 — Fits
Qwen2.5-Coder-7B-Instruct 7.6B Q8_0 9.1 GiB 32,768 — Fits
Qwen2.5-VL-7B-Instruct 7.6B BF16 15.6 GiB 59,136 — Fits
Qwen2.5-VL-7B-Surg-CholecT50 7.6B F16 15.6 GiB 59,136 — Fits
Spatial-SSRL-7B 7.6B Q6_K 7.5 GiB 128,000 — Fits
X-Reasoner-7B 7.6B Q6_K 7.5 GiB 128,000 — Fits
ZwZ-7B 7.6B Q6_K 7.5 GiB 128,000 — Fits
gemma-4-E4B 7.5B F16 11.0 GiB 131,072 — Fits
gemma-4-E4B-it 7.5B BF16 11.0 GiB 131,072 — Fits
Bielik-Minitron-7B-v3.0-Instruct 7.5B Q8_0 10.2 GiB 32,768 — Fits
ContextPilot-E4B 7.5B F16 10.9 GiB 131,072 — Fits
Olmo-3.1-7B-RL-Zero-Code 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3.1-7B-RL-Zero-Math 7.3B Q8_0 12.7 GiB 27,648 — Fits
Olmo-3-7B-Instruct 7.3B Q8_0 12.7 GiB 27,648 — Fits
Olmo-3-7B-Instruct-DPO 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-Instruct-SFT 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-RL-Zero-Code 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-RL-Zero-General 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-RL-Zero-IF 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-RL-Zero-Math 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-RL-Zero-Mix 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-Think 7.3B Q8_0 12.7 GiB 27,648 — Fits
Olmo-3-7B-Think-DPO 7.3B Q6_K 11.0 GiB 40,192 — Fits
Olmo-3-7B-Think-SFT 7.3B Q6_K 11.0 GiB 40,192 — Fits
Mistral-7B-Instruct-v0.2 7.2B Q6_K 8.1 GiB 32,768 — Fits
Trinity-Nano-Base 6.1B (819M active) Q6_K 6.7 GiB 131,072 — Fits
Trinity-Nano-Base-Pre-Anneal 6.1B (819M active) Q6_K 6.4 GiB 4,096 — Fits
Trinity-Nano-Preview 6.1B (819M active) BF16 13.5 GiB 131,072 — Fits
Youtu-VL-4B-Instruct 4.9B Q6_K 5.9 GiB 32,768 — Fits
Holo-3.1-4B 4.8B F16 10.2 GiB 253,952 — Fits
gemma-4-E2B 4.6B Q6_K 3.5 GiB 131,072 — Fits
gemma-4-E2B-it 4.6B BF16 6.1 GiB 131,072 — Fits
AgentCPM-Explore 4.4B Q6_K 5.7 GiB 96,000 — Fits
CapRL-Qwen3VL-4B 4.4B F16 10.4 GiB 64,256 — Fits
Holo2-4B 4.4B F16 10.4 GiB 64,256 — Fits
intent-aware-lfqa-qwen3-4b-baseline 4.4B F16 10.4 GiB 40,960 — Fits
intent-aware-lfqa-qwen3-4b-intent-explicit 4.4B F16 10.4 GiB 40,960 — Fits
intent-aware-lfqa-qwen3-4b-intent-implicit 4.4B F16 10.4 GiB 40,960 — Fits
intent-aware-lfqa-qwen3-4b-multiview 4.4B F16 10.4 GiB 40,960 — Fits
Spatial-SSRL-Qwen3VL-4B 4.4B F16 10.4 GiB 64,256 — Fits
ZwZ-4B 4.4B Q6_K 5.7 GiB 96,000 — Fits
Fara1.5-4B 4.2B BF16 10.2 GiB 253,952 — Fits
FrogNano-4B-2609 4.2B BF16 10.2 GiB 253,952 — Fits
Qwen3.5-4B 4.2B BF16 10.2 GiB 253,952 — Fits
Qwen3.5-4B-Base 4.2B F16 10.2 GiB 253,952 — Fits
UserRepGen-4B 4.2B BF16 10.2 GiB 253,952 — Fits
Riva-Translate-4B-Instruct-v1.1 4.2B Q6_K 6.0 GiB 8,192 — Fits
GELab-Zero-4B-preview 4.0B BF16 10.4 GiB 64,256 — Fits
Qwen3-4B 4.0B Q8_0 6.7 GiB 40,960 — Fits
Qwen3-4B-Instruct-2507 4.0B F16 10.4 GiB 64,256 — Fits
Qwen3-VL-4B-Instruct 4.0B BF16 10.4 GiB 64,256 — Fits
Qwen3-VL-4B-Thinking 4.0B BF16 10.4 GiB 64,256 — Fits
NVIDIA-Nemotron-3-Nano-4B-BF16 4.0B BF16 8.7 GiB 262,144 — Fits
gemma-3-4b-it 3.9B BF16 9.9 GiB 131,072 — Fits
medgemma-1.5-4b-it 3.9B BF16 9.9 GiB 131,072 — Fits
Nemotron-3-Content-Safety 3.9B Q6_K 5.4 GiB 131,072 — Fits
translategemma-4b-it 3.9B F16 9.9 GiB 131,072 — Fits
Phi-4-mini-instruct 3.8B BF16 10.0 GiB 74,240 — Fits
Apertus-v1.1-4B 3.8B Q6_K 4.4 GiB 4,096 — Fits
Phi-3.5-mini-instruct 3.8B Q8_0 8.3 GiB 34,560 — Fits
granite-4.2-3b 3.7B BF16 8.6 GiB 129,792 — Fits
BitCPM-CANN-3B-unquantized 3.6B Q6_K 4.3 GiB 32,768 — Fits
kanana-2-3b-instruct 3.5B F16 8.6 GiB 32,768 — Fits
Ministral-3-3B-Instruct-2512 3.4B BF16 8.9 GiB 99,840 — Fits
Ministral-3-3B-Reasoning-2512 3.4B BF16 8.9 GiB 99,840 — Fits
granite-vision-4.1-4b 3.4B BF16 8.6 GiB 129,792 — Fits
granite-4.1-3b 3.4B BF16 8.6 GiB 129,792 — Fits
granite-guardian-4.0-3b-toxicity-ja 3.4B Q6_K 4.7 GiB 131,072 — Fits
Qwen2.5-3B-Instruct 3.4B Q8_0 4.8 GiB 32,768 — Fits
tiny-aya-earth 3.3B BF16 8.7 GiB ≥ 262,144 (longer contexts not checked against measured logs) — Fits
tiny-aya-fire 3.3B BF16 8.7 GiB ≥ 262,144 (longer contexts not checked against measured logs) — Fits
tiny-aya-global 3.3B BF16 8.7 GiB ≥ 262,144 (longer contexts not checked against measured logs) — Fits
tiny-aya-water 3.3B BF16 8.7 GiB ≥ 262,144 (longer contexts not checked against measured logs) — Fits
Llama-3.2-3B-Instruct 3.2B BF16 8.5 GiB 96,512 — Fits
AI21-Jamba-Reasoning-3B 3.2B BF16 7.3 GiB 262,144 — Fits
AI21-Jamba2-3B 3.2B BF16 7.3 GiB 262,144 — Fits
kanana-2-3b-base 3.2B Q6_K 4.9 GiB 32,768 — Fits
SmolLM3-3B 3.1B BF16 7.9 GiB 65,536 — Fits
SmolLM3-3B-GSM8K-SFT 3.1B Q6_K 4.3 GiB 65,536 — Fits
LFM2-2.6B-Exp 2.7B F16 6.5 GiB 128,000 — Fits
d1-3B 2.7B BF16 6.7 GiB 32,768 — Fits
LFM2.5-2.6B 2.7B BF16 6.7 GiB 131,072 — Fits
LFM2.5-2.6B-Base 2.7B Q6_K 3.6 GiB 128,000 — Fits
LFM2.5-VL-3B 2.7B BF16 6.7 GiB 32,768 — Fits
EuroMoE-2.6B-A0.6B-Instruct-2512 2.6B (498M active) Q6_K 3.6 GiB 32,768 — Fits
LFM2-2.6B-Longevity 2.6B F16 6.5 GiB 128,000 — Fits
LFM2-2.6B-Transcript 2.6B BF16 6.5 GiB 128,000 — Fits
LFM2-VL-3B 2.6B BF16 6.5 GiB 128,000 — Fits
MiniCPM5-2B 2.5B BF16 6.0 GiB 131,072 — Fits
MiniCPM5-2B-Base 2.5B Q6_K 3.4 GiB ≥ 262,144 (longer contexts not checked against measured logs) — Fits
CapRL-Qwen3VL-2B 2.0B F16 5.6 GiB 121,856 — Fits
HiPO-1.7B 2.0B Q6_K 3.6 GiB 40,960 — Fits
UI-Venus-1.5-2B 2.0B F16 5.6 GiB 121,856 — Fits
ZwZ-2B 2.0B F16 5.6 GiB 121,856 — Fits
Youtu-LLM-2B 2.0B Q6_K 3.2 GiB 131,072 — Fits
Youtu-LLM-2B-Base 2.0B Q6_K 3.2 GiB 131,072 — Fits
Qwen3.5-2B 1.9B BF16 5.4 GiB 262,144 — Fits
Qwen3.5-2B-Base 1.9B Q6_K 3.2 GiB 262,144 — Fits
Qwen3-1.7B 1.7B BF16 5.6 GiB 40,960 — Fits
Qwen3-VL-2B-Instruct 1.7B BF16 5.6 GiB 121,856 — Fits
Qwen3-VL-2B-Thinking 1.7B BF16 5.6 GiB 121,856 — Fits
SingGuard-2b 1.7B F16 5.6 GiB 121,856 — Fits
granite-4.0-1b 1.6B BF16 5.1 GiB 131,072 — Fits
Apertus-v1.1-1.5B 1.5B F16 3.8 GiB 4,096 — Fits
Apertus-v1.1-1.5B-Instruct 1.5B F16 3.8 GiB 4,096 — Fits
granite-4.0-h-1b 1.5B BF16 4.4 GiB 131,072 — Fits
Llama-3.2-1B-Instruct 1.2B BF16 4.0 GiB 131,072 — Fits
LFM2.5-1.2B-Base 1.2B BF16 3.7 GiB 128,000 — Fits
LFM2.5-1.2B-Instruct 1.2B BF16 3.7 GiB 128,000 — Fits
LFM2.5-1.2B-JP 1.2B F16 3.7 GiB 128,000 — Fits
LFM2.5-1.2B-JP-202606 1.2B F16 3.7 GiB 128,000 — Fits
LFM2.5-1.2B-Thinking 1.2B BF16 3.7 GiB 128,000 — Fits
LFM2.5-VL-1.6B 1.2B BF16 3.7 GiB 128,000 — Fits
LFM2.5-VL-1.6B-Extract 1.2B F16 3.7 GiB 128,000 — Fits
MiniCPM5-1B 1.1B Q6_K 2.2 GiB 131,072 — Fits
MiniCPM5-1B-Base 1.1B Q6_K 2.2 GiB 131,072 — Fits
MiniCPM5-1B-SFT 1.1B Q6_K 2.2 GiB 131,072 — Fits
Holo-3.1-0.8B 1.0B Q6_K 2.3 GiB 262,144 — Fits
MiniCPM-V-4.7-1B 1.0B F16 3.2 GiB 262,144 — Fits
gemma-3-1b-it 1.0B BF16 3.6 GiB 32,768 — Fits
Qwen3.5-0.8B 752M BF16 3.2 GiB 262,144 — Fits
Qwen3.5-0.8B-Base 752M Q6_K 2.3 GiB 262,144 — Fits
MiniCPM-V-4.6 752M F16 3.2 GiB 262,144 — Fits
MiniCPM-V-4.6-Thinking 752M F16 3.2 GiB 262,144 — Fits
Qwen3-0.6B 752M BF16 3.4 GiB 40,960 — Fits
HunyuanOCR 539M F16 3.1 GiB 32,768 — Fits
LFM2.5-350M 354M BF16 2.1 GiB 128,000 — Fits
LFM2.5-Encoder-350M-Diffusion 354M Q6_K 1.7 GiB 128,000 — Fits
LFM2.5-VL-450M 354M BF16 2.1 GiB 128,000 — Fits
LFM2.5-VL-450M-Extract 354M F16 2.1 GiB 128,000 — Fits
granite-4.0-350m 352M BF16 2.2 GiB 32,768 — Fits
granite-4.0-h-350m 340M BF16 2.1 GiB 32,768 — Fits
functiongemma-270m-it 268M BF16 2.1 GiB 32,768 — Fits
LFM2.5-230M 230M BF16 1.8 GiB 128,000 — Fits

Borderline models

Within the estimate error range: it may fit or it may not. Try a shorter context or a lower quantization.

Context 8,192 tokens. No quantization of these models fits; the one shown is the highest that is borderline.
Model Parameters Highest borderline quantization Available memory needed for “Fits” Result
Qwen3-Coder-Next-Base 79.7B (3.6B active) IQ2_XXS 22.3 GiB Tight
Llama-3.1-70B-Instruct 70.6B IQ2_XXS 22.3 GiB Tight
MiniCPM-V-4.7-35B-A3B 34.7B (2.9B active) Q4_K_M 22.4 GiB Tight
QwQ-32B 32.8B Q4_K_M 22.5 GiB Tight
Qwen3-32B 32.8B Q4_K_M 22.4 GiB Tight
EXAONE-4.5-33B 32.5B Q4_K_M 22.3 GiB Tight
Holotron4-30B-A3B 31.6B (3.2B active) Q4_0 19.0 GiB Tight
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 31.6B (3.2B active) Q4_K_M 23.3 GiB Tight
Nemotron-Cascade-2-30B-A3B 31.6B (3.2B active) Q4_0 19.2 GiB Tight
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16 31.6B (3.2B active) Q4_0 19.0 GiB Tight
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 31.6B (3.2B active) Q4_0 19.3 GiB Tight
NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 31.6B (3.2B active) Q4_0 19.2 GiB Tight
NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-BF16 31.6B (3.2B active) Q4_0 19.0 GiB Tight
gemma-4-31B-it 30.7B Q4_K_M 23.9 GiB Tight
Qwen3-30B-A3B 30.5B (3.0B active) Q5_K_M 23.2 GiB Tight

Models that don’t fit on one AMD Radeon RX 7900 XT

Context 8,192 tokens. Even the lowest quantization does not fit on one card; the requirement shown is for that quantization.
Model Parameters Smallest quantization Available memory needed for “Fits” Several identical cards Result
KAT-Dev-72B-Exp 72.7B IQ2_XXS 28.5 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Solar-Open-100B 103B (11.2B active) IQ2_XXS 29.4 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
DeepSeek-R1-Distill-Llama-70B 70.6B Q2_K 29.4 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Llama-3.3-70B-Instruct 70.6B Q2_K 29.4 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Qwen3-Next-80B-A3B-Instruct 79.7B (3.6B active) Q2_K 30.3 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Qwen2.5-72B-Instruct 73.0B Q2_K 30.3 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Qwen3-Coder-Next 79.7B (3.6B active) Q2_K 30.4 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
GLM-4.5-Air-REAP-82B-A12B 81.9B (12.8B active) IQ2_XXS 31.1 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Step-3.5-Flash-REAP-121B-A11B 121B (11.4B active) IQ2_XXS 32.9 GiB 2 identical cards (default even split) are expected to fit. Doesn't fit
Mistral-Small-4-119B-2603 119B (6.1B active) IQ2_XXS 34.3 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Qwen3.5-122B-A10B 122B (9.0B active) IQ2_XXS 35.0 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Qwen-AgentWorld-35B-A3B 34.7B (2.9B active) Q8_0 37.6 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiniMax-M2.5-REAP-139B-A10B 139B (10.4B active) IQ2_XXS 39.2 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
INTELLECT-3 107B (12.8B active) IQ2_XXS 39.6 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Step-3.5-Flash-REAP-149B-A11B 149B (11.4B active) IQ2_XXS 40.1 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Llama-4-Scout-17B-16E-Instruct 108B (16.1B active) Q2_K 41.5 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
GLM-4.5-Air 107B (12.8B active) Q2_K 45.6 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
GLM-4.6V 107B (12.8B active) Q2_K 45.6 GiB 3 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Step-3.5-Flash 197B (11.5B active) IQ2_XXS 48.9 GiB 3 identical cards (default even split) are expected to fit. Doesn't fit
Devstral-2-123B-Instruct-2512 125B Q2_K 49.3 GiB 3 identical cards (default even split) are expected to fit. Doesn't fit
Mistral-Medium-3.5-128B 125B Q2_K 49.3 GiB 3 identical cards (default even split) are expected to fit. Doesn't fit
NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16 121B (12.2B active) IQ2_XXS 49.8 GiB 3 identical cards (default even split) are expected to fit. Doesn't fit
NVIDIA-Nemotron-3-Super-120B-A12B-BF16 121B (12.2B active) IQ2_XXS 51.0 GiB 3 identical cards (default even split) are expected to fit. Doesn't fit
gpt-oss-120b 117B (5.1B active) Q2_K 62.8 GiB 4 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiniMax-M2.7 229B (10.4B active) IQ2_XXS 62.9 GiB 4 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Baichuan-M3-235B 235B (21.6B active) IQ2_XXS 63.2 GiB 4 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Holo2-235B-A22B 235B (21.6B active) IQ2_XXS 63.2 GiB 4 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Qwen3-Nemotron-235B-A22B-GenRM 235B (21.6B active) IQ2_XXS 63.2 GiB 4 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
command-a-plus-05-2026-bf16 218B (25.0B active) IQ2_XXS 63.6 GiB 5 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 121B (12.2B active) Q4_0 71.8 GiB 5 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
GLM-4.7-REAP-268B-A32B 269B (32.8B active) IQ2_XXS 73.0 GiB 5 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Step-3.7-Flash 197B (11.5B active) IQ3_XXS 76.1 GiB 5 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Hy3 295B (20.2B active) IQ2_XXS 82.2 GiB 5 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiMo-V2.5 309B (14.8B active) IQ2_XXS 82.4 GiB 6 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
GLM-4.7-REAP-218B-A32B 218B (32.8B active) Q2_K 82.9 GiB 5 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiniMax-M2 229B (10.4B active) Q2_K 84.9 GiB 6 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiniMax-M2.1 229B (10.4B active) Q2_K 84.9 GiB 6 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiniMax-M2.5 229B (10.4B active) Q2_K 84.9 GiB 6 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Qwen3-235B-A22B-Instruct-2507 235B (21.6B active) Q2_K 86.7 GiB 6 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Qwen3-VL-235B-A22B-Instruct 235B (21.6B active) Q2_K 86.7 GiB 6 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
GLM-5.3-Flash-BF16 313B (16.7B active) IQ2_XXS 92.0 GiB 7 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Laguna-S-2.1 118B (8.1B active) Q4_K_M 95.9 GiB Borderline with 8 cards. Doesn't fit
Trinity-Large-Base 399B (12.8B active) IQ2_XXS 96.4 GiB 7 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Trinity-Large-TrueBase 399B (12.8B active) IQ2_XXS 96.4 GiB 7 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
GLM-4.6-REAP-268B-A32B 269B (32.8B active) Q2_K 100.9 GiB 7 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Trinity-Large-Thinking 399B (12.8B active) IQ2_XXS 107.1 GiB 8 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiMo-V2-Flash 309B (14.8B active) Q2_K 111.6 GiB 8 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Intern-S2-397B 396B (16.3B active) IQ2_XXS 114.3 GiB 8 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiMo-V2.6-Flash-MOPD 309B (14.8B active) Q2_K 124.9 GiB 8 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
MiMo-V2.6-Flash-RL 309B (14.8B active) Q2_K 124.9 GiB 8 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
K-EXAONE-236B-A23B 237B (22.7B active) IQ4_XS 126.3 GiB Borderline with 8 cards. Doesn't fit
GLM-4.5 353B (32.9B active) Q2_K 130.7 GiB Borderline with 8 cards. Doesn't fit
GLM-4.7 353B (32.9B active) Q2_K 130.7 GiB Borderline with 8 cards. Doesn't fit
Qwen3-Coder-REAP-363B-A35B 363B (34.5B active) Q2_K 132.7 GiB 8 identical cards (default even split) are expected to fit. Fewer cards are borderline. Doesn't fit
Qwen3-235B-A22B 235B (21.6B active) Q4_K_M 141.9 GiB Borderline with 8 cards. Doesn't fit
Qwen3-Coder-480B-A35B-Instruct 480B (34.5B active) Q2_K 174.3 GiB Borderline with 8 cards. Doesn't fit
Qwen3.5-397B-A17B 396B (16.3B active) Q3_K_M 175.6 GiB Borderline with 8 cards. Doesn't fit
GLM-5.1 744B (40.8B active) IQ2_XXS 197.6 GiB Does not fit even on 8 cards. Doesn't fit
DeepSeek-V3.2 671B (36.6B active) Q2_K 242.1 GiB Does not fit even on 8 cards. Doesn't fit
cogito-671b-v2.1 671B (36.6B active) Q2_K 242.4 GiB Does not fit even on 8 cards. Doesn't fit
DeepSeek-V3.1 671B (36.6B active) Q2_K 242.4 GiB Does not fit even on 8 cards. Doesn't fit
Mistral-Large-3-675B-Instruct-2512 673B (39.0B active) Q2_K 243.8 GiB Does not fit even on 8 cards. Doesn't fit
DeepSeek-V3 671B (36.6B active) Q2_K 257.8 GiB Does not fit even on 8 cards. Doesn't fit
Kimi-K2.6 1.03T (31.7B active) IQ2_XXS 267.6 GiB Does not fit even on 8 cards. Doesn't fit
GLM-5 744B (40.8B active) Q2_K 268.9 GiB Does not fit even on 8 cards. Doesn't fit
MiMo-V2.5-Pro 1.02T (41.0B active) IQ2_XXS 270.4 GiB Does not fit even on 8 cards. Doesn't fit
DeepSeek-R1 671B (36.6B active) Q2_K 280.3 GiB Does not fit even on 8 cards. Doesn't fit
Muse-Spark-Open-816B-A42B 820B (44.4B active) Q2_K 301.0 GiB Does not fit even on 8 cards. Doesn't fit
GLM-5.3-Flash 313B (16.7B active) Q8_0 327.9 GiB Does not fit even on 8 cards. Doesn't fit
Ling-1T 1.00T (49.8B active) Q2_K 360.6 GiB Does not fit even on 8 cards. Doesn't fit
Kimi-K2-Instruct 1.03T (31.7B active) Q2_K 367.4 GiB Does not fit even on 8 cards. Doesn't fit
Kimi-K2.5 1.03T (31.7B active) Q2_K 368.0 GiB Does not fit even on 8 cards. Doesn't fit
Kimi-K2-Thinking 1.03T (31.7B active) Q2_K 368.3 GiB Does not fit even on 8 cards. Doesn't fit
MiMo-V2.6-Pro-MOPD 1.02T (41.0B active) Q2_K 411.6 GiB Does not fit even on 8 cards. Doesn't fit
MiMo-V2.6-Pro-RL 1.02T (41.0B active) Q2_K 411.6 GiB Does not fit even on 8 cards. Doesn't fit
Beam-1.0-501B-A23B 501B (22.7B active) Q8_0 523.6 GiB Does not fit even on 8 cards. Doesn't fit
GLM-5.2 744B (40.8B active) Q8_0 776.4 GiB Does not fit even on 8 cards. Doesn't fit
GLM-5.3 744B (40.8B active) Q8_0 776.4 GiB Does not fit even on 8 cards. Doesn't fit
Qwen3.8-2.4T-A95B 2.42T (93.3B active) Q8_0 2,524.1 GiB Does not fit even on 8 cards. Doesn't fit

The speed ceiling is memory bandwidth divided by the bytes read per generated token. It is a theoretical upper bound; public llama.cpp benchmarks typically reach about half to four-fifths of it, and small models and mixture-of-experts models often less. Public benchmarks

Detailed pages for this GPU