GTX 1650 · 4 GB VRAM
NVIDIA · GTX 1650
15 text models run well · 128 GB/s bandwidth
Models on this hardware
Start here
File Q4_K_M · 0.8 GB · bartowski/Llama-3.2-1B-Instruct-GGUF · What is Q4?
Memory Needs ~1 GB · ~3 GB on GTX 1650 · 4 GB VRAM, including ~2 GB the system keeps
ollama run hf.co/bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M- Install Ollama
- Run the command above
- Ask:
Explain what a context window is, in two sentences.
Text
Text · LMArena
rating · higher is better
Source: LMArena 2026-10-22 · collected 2026-10-25 · dataset · CC BY 4.0 · Models without a rating are left out, not scored as zero.
| Llama 3.2 1B InstructQ4 · published params · source | 1,111 ± 8 | ~3 GB | Q4 · published params |
|---|---|---|---|
| Qwen3-0.6BQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
| SmolLM2-135MQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
| SmolLM2-135M-InstructQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
| Llama-3.2-1BQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
| Qwen3Guard-Gen-0.6BQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
| SmolLM2-360M-InstructQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
| SmolLM2-360MQ4 · tensor count · source | Catalog record: not reviewed for LMArena | ~3 GB | Q4 · tensor count |
No model within 4 GB has a LiveBench score, so the table has no LiveBench column. LMArena ratings in this table: LMArena 2026-10-22 · collected 2026-10-25 · dataset · CC BY 4.0
Tight fit: it runs, but leaves less than 1 GB free beyond the 2 GB the system keeps (the memory figure already includes that reserve).
Every reviewed text model here 35
Every one that runs here, well or slowly, scored or not (2 run well, 33 slowly); 17 of 35 have a LiveBench or LMArena result. Run well first, then tight fits, then slow; within each, by Hub downloads over 30 days, read 2026-10-25, not by quality.
| Llama 3.2 1B Instruct | Not in LiveBench 2026-06-25 | 1,111 | ~3 GB | 0.8 | Llama 3.2 | 3.5M |
|---|---|---|---|---|---|---|
| Qwen3 1.7B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~4 GBTight fit | 1.1 | Apache 2.0 | 4.1M |
| Qwen3 14B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~14 GBRuns slowly | 9.0 | Apache 2.0 | 10.4M |
| Gemma 4 31B | Not in LiveBench 2026-06-25 | 1,452 | ~26 GBRuns slowly | 18.3 | Apache 2.0 | 8.4M |
| Gemma 4 26B A4B | Not in LiveBench 2026-06-25 | 1,437 | ~21 GBRuns slowly | 16.9 | Apache 2.0 | 8.2M |
| Qwen3 8B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~9 GBRuns slowly | 5.0 | Apache 2.0 | 8.1M |
| Qwen3.8 27B | 75.3 | 1,437 | ~23 GBRuns slowly | 16.5 | Apache 2.0 | 7.8M |
| Qwen3.5 9B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~9 GBRuns slowly | 5.7 | Apache 2.0 | 6.1M |
| gpt-oss-20b | Not in LiveBench 2026-06-25 | 1,317 | ~18 GBRuns slowly | 11.6 | Apache 2.0 | 5.8M |
| Llama 3.1 8B Instruct | Not in LiveBench 2026-06-25 | 1,211 | ~8 GBRuns slowly | 4.9 | Llama 3.1 | 4.8M |
| Gemma 4 E4B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~8 GBRuns slowly | 5.0 | Apache 2.0 | 4M |
| Gemma 4 E2B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~6 GBRuns slowly | 3.1 | Apache 2.0 | 3M |
| Qwen3 4B Instruct 2507 | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~5 GBRuns slowly | 2.5 | Apache 2.0 | 2.6M |
| Qwen3.6 27B | 64.0 | Not on the LMArena text leaderboard of 2026-10-22 | ~23 GBRuns slowly | 16.8 | Apache 2.0 | 2.4M |
| Gemma 4 12B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~11 GBRuns slowly | 7.1 | Apache 2.0 | 1.7M |
| Llama 3.2 3B Instruct | Not in LiveBench 2026-06-25 | 1,167 | ~5 GBRuns slowly | 2.0 | Llama 3.2 | 1.1M |
| Qwen3 30B A3B Instruct 2507 | Not in LiveBench 2026-06-25 | 1,382 | ~25 GBRuns slowly | 18.6 | Apache 2.0 | 1.1M |
| GLM-4.7 Flash | Not in LiveBench 2026-06-25 | 1,365 | ~25 GBRuns slowly | 18.3 | MIT | 953.7K |
| Nemotron 3.5 Lightning 30B A3B | Not in LiveBench 2026-06-25 | Arena rated the NVFP4 build, NVIDIA-Nemotron-3.5-Lightning-30B-A3B (NVFP4); this entry is the BF16 reference, which NVIDIA reports with different scores | ~25 GBRuns slowly | 18.9 | OpenMDW 1.1 | 861.5K |
| Qwen3 Coder 30B A3B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~25 GBRuns slowly | 18.6 | Apache 2.0 | 745K |
| SmolLM3 3B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~5 GBRuns slowly | 1.9 | Apache 2.0 | 718.7K |
| Phi-4 | Not in LiveBench 2026-06-25 | 1,256 | ~13 GBRuns slowly | 8.0 | MIT | 368.3K |
| DeepSeek R1 0528 Qwen3 8B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~9 GBRuns slowly | 5.0 | MIT | 355.8K |
| Phi-4 mini | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~5 GBRuns slowly | 2.5 | MIT | 342.8K |
| Olmo 3 7B Instruct | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~8 GBRuns slowly | 4.5 | Apache 2.0 | 262.5K |
| Devstral Small 2 24B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~20 GBRuns slowly | 14.3 | Apache 2.0 | 195.6K |
| Mistral Small 3.2 24B | Not in LiveBench 2026-06-25 | 1,356 | ~21 GBRuns slowly | 14.3 | Apache 2.0 | 181.4K |
| Ministral 3 14B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~13 GBRuns slowly | 8.2 | Apache 2.0 | 126.5K |
| Ministral 3 3B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~5 GBRuns slowly | 2.1 | Apache 2.0 | 87.4K |
| Ministral 3 8B | Not in LiveBench 2026-06-25 | Not on the LMArena text leaderboard of 2026-10-22 | ~9 GBRuns slowly | 5.2 | Apache 2.0 | 79.4K |
| Granite 4.2 8B | Not in LiveBench 2026-06-25 | 1,290 | ~8 GBRuns slowly | 5.3 | Apache 2.0 | 78.3K |
| Granite 4.2 3B | Not in LiveBench 2026-06-25 | 1,290 | ~5 GBRuns slowly | 2.2 | Apache 2.0 | 39.8K |
| Granite 4.2 30B | Not in LiveBench 2026-06-25 | 1,340 | ~25 GBRuns slowly | 17.7 | Apache 2.0 | 39.3K |
| Olmo 3 32B Think | Not in LiveBench 2026-06-25 | 1,306 | ~26 GBRuns slowly | 19.5 | Apache 2.0 | 13.1K |
| Smaug Mini | 76.9 | Not on the LMArena text leaderboard of 2026-10-22 | ~23 GBRuns slowly | 16.5 | Apache 2.0 | 817 |
Also runs slowly
Above the 4 GB of VRAM, the layers that do not fit run on the CPU from PC RAM, up to 28 GB in all. 32 GB of PC RAM assumed: this configuration lists none. Planning estimate, not a speed figure. The 33 reviewed models that run slowly are in the table above.
- granite-4.1-3bibm-granite~5 GBruns well with 6 GB
- Llama-3.2-3Bmeta-llama~5 GBruns well with 6 GB
- granite-4.1-3b-baseibm-granite~5 GBruns well with 6 GB
- granite-3b-code-instruct-2kibm-granite~5 GBruns well with 6 GB
- granite-3b-code-instruct-128kibm-granite~5 GBruns well with 6 GB
- granite-3b-code-base-128kibm-granite~5 GBruns well with 6 GB
- Qwen3-4BQwen~6 GBruns well with 6 GB
- Qwen3-4B-Thinking-2507Qwen~6 GBruns well with 6 GB
- Qwen3Guard-Gen-4BQwen~6 GBruns well with 6 GB
- Riva-Translate-4B-Instruct-v2nvidia~6 GBruns well with 6 GB
- Riva-Translate-4B-Instruct-v1.1nvidia~6 GBruns well with 6 GB
- Riva-Translate-4B-Instructnvidia~6 GBruns well with 6 GB
- Mistral-7B-Instruct-v0.2mistralai~8 GBQ4 · tensor count
- Llama-2-7b-hfmeta-llama~8 GBQ4 · tensor count
- Mistral-7B-v0.1mistralai~8 GBQ4 · tensor count
- DeepSeek-R1-Distill-Qwen-7Bdeepseek-ai~8 GBQ4 · tensor count
- Llama-2-7b-chat-hfmeta-llama~8 GBQ4 · tensor count
- Mistral-7B-Instruct-v0.1mistralai~8 GBQ4 · tensor count
- Slerp-CM-mist-dpoabacusai~8 GBQ4 · tensor count
- CodeLlama-7b-Python-hfmeta-llama~8 GBQ4 · tensor count
- LlamaGuard-7bmeta-llama~8 GBQ4 · tensor count
- CodeLlama-7b-Instruct-hfmeta-llama~8 GBQ4 · tensor count
- HARC-Qwen2.5-7B-Instructmicrosoft~8 GBQ4 · tensor count
- SynLogic-7BMiniMaxAI~8 GBQ4 · tensor count
- BAR-7Ballenai~8 GBQ4 · tensor count
- japanese-stablelm-base-ja_vocab-beta-7bstabilityai~8 GBQ4 · tensor count
- japanese-stablelm-base-beta-7bstabilityai~8 GBQ4 · tensor count
- japanese-stablelm-instruct-beta-7bstabilityai~8 GBQ4 · tensor count
- japanese-stablelm-instruct-ja_vocab-beta-7bstabilityai~8 GBQ4 · tensor count
- japanese-stablelm-base-gamma-7bstabilityai~8 GBQ4 · tensor count
- StableBeluga-7Bstabilityai~8 GBQ4 · tensor count
- japanese-stablelm-instruct-gamma-7bstabilityai~8 GBQ4 · tensor count
- Infinity-Instruct-7M-Gen-mistral-7BBAAI~8 GBQ4 · tensor count
- Infinity-Instruct-3M-0613-Mistral-7BBAAI~8 GBQ4 · tensor count
- Infinity-Instruct-3M-0625-Qwen2-7BBAAI~8 GBQ4 · tensor count
- Infinity-Instruct-3M-0625-Mistral-7BBAAI~8 GBQ4 · tensor count
- Fewshot-Metamath-OrcaVicuna-Mistralabacusai~8 GBQ4 · tensor count
- Llama-3.1-8Bmeta-llama~9 GBQ4 · tensor count
- Meta-Llama-3-8B-Instructmeta-llama~9 GBQ4 · tensor count
- Meta-Llama-3-8Bmeta-llama~9 GBQ4 · tensor count
- DeepSeek-R1-Distill-Llama-8Bdeepseek-ai~9 GBQ4 · tensor count
- granite-4.1-8bibm-granite~9 GBQ4 · tensor count
- Qwen3Guard-Gen-8BQwen~9 GBQ4 · tensor count
- granite-guardian-4.1-8bibm-granite~9 GBQ4 · tensor count
- Llama-Guard-3-8Bmeta-llama~9 GBQ4 · tensor count
- Llama-3-Smaug-8Babacusai~9 GBQ4 · tensor count
- Infinity-Instruct-3M-0625-Llama3-8BBAAI~9 GBQ4 · tensor count
- Infinity-Instruct-3M-0625-Yi-1.5-9BBAAI~9 GBQ4 · tensor count
- granite-4.1-8b-baseibm-granite~9 GBQ4 · tensor count
- granite-8b-code-base-4kibm-granite~9 GBQ4 · tensor count
- Meta-Llama-Guard-2-8Bmeta-llama~9 GBQ4 · tensor count
- granite-8b-code-instruct-128kibm-granite~9 GBQ4 · tensor count
- WebWorld-8BQwen~9 GBQ4 · tensor count
- HARC-Llama-3.1-8B-Instructmicrosoft~9 GBQ4 · tensor count
- ContextPilot-8Btencent~9 GBQ4 · tensor count
- granite-8b-code-instruct-4kibm-granite~9 GBQ4 · tensor count
- gemma-7b-aps-itgoogle~9 GBQ4 · tensor count
- granite-8b-code-base-128kibm-granite~9 GBQ4 · tensor count
- LongWriter-llama3.1-8bzai-org~9 GBQ4 · tensor count
- txgemma-9b-predictgoogle~9 GBQ4 · tensor count
- txgemma-9b-chatgoogle~9 GBQ4 · tensor count
- KernelLLMfacebook~9 GBQ4 · tensor count
- Llama-2-13b-chat-hfmeta-llama~12 GBQ4 · tensor count
- Llama-2-13b-hfmeta-llama~12 GBQ4 · tensor count
- bigstral-12b-32kabacusai~12 GBQ4 · tensor count
- CodeLlama-13b-Instruct-hfmeta-llama~12 GBQ4 · tensor count
- StableBeluga-13Bstabilityai~12 GBQ4 · tensor count
- codellama13b_instruct_260k_synthesisstabilityai~12 GBQ4 · tensor count
- Giraffe-13b-32k-v3abacusai~12 GBQ4 · tensor count
- bigstral-12b-v0.2-32kabacusai~12 GBQ4 · tensor count
- CodeLlama-13b-hfmeta-llama~12 GBQ4 · tensor count
- CodeLlama-13b-Python-hfmeta-llama~12 GBQ4 · tensor count
- DeepSeek-R1-Distill-Qwen-14Bdeepseek-ai~14 GBQ4 · tensor count
- bigyi-15babacusai~14 GBQ4 · tensor count
- MagenticBrainmicrosoft~14 GBQ4 · tensor count
- ContextPilot-14Btencent~14 GBQ4 · tensor count
- WebWorld-14BQwen~14 GBQ4 · tensor count
- txgemma-27b-chatgoogle~23 GBQ4 · tensor count
- txgemma-27b-predictgoogle~23 GBQ4 · tensor count
- granite-4.1-30bibm-granite~24 GBQ4 · tensor count
- granite-4.1-30b-baseibm-granite~24 GBQ4 · tensor count
- Qwen3-30B-A3B-Thinking-2507Qwen~25 GBQ4 · tensor count
- DeepSeek-R1-Distill-Qwen-32Bdeepseek-ai~27 GBQ4 · tensor count
- SWE-Dev-32Bzai-org~27 GBQ4 · tensor count
- SynLogic-32BMiniMaxAI~27 GBQ4 · tensor count
- SynLogic-Mix-3-32BMiniMaxAI~27 GBQ4 · tensor count
- DRIVE-SFTtencent~27 GBQ4 · tensor count
- DRIVE-RLtencent~27 GBQ4 · tensor count
- Smaug-34B-v0.1abacusai~28 GBQ4 · tensor count
- CodeLlama-34b-hfmeta-llama~28 GBQ4 · tensor count
- CodeLlama-34b-Instruct-hfmeta-llama~28 GBQ4 · tensor count
- MetaMath-Bagel-DPO-34Babacusai~28 GBQ4 · tensor count
- MetaMath-bagel-34b-v0.2-c1500abacusai~28 GBQ4 · tensor count
- MM-OV-bagel-DPO-34b-c1000-250abacusai~28 GBQ4 · tensor count
- MM-Orc-Vic-bagel-34b-c1000abacusai~28 GBQ4 · tensor count
- CodeLlama-34b-Python-hfmeta-llama~28 GBQ4 · tensor count
Adjust the estimate: quantization, context, reserve
Vision
| Falcon-OCRQ4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
|---|---|---|
| Florence-2-baseQ4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
| Florence-2-largeQ4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
| GLM-OCRQ4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
| GOT-OCR2_0Q4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
| granite-docling-258MQ4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
| LightOnOCR-1B-1025Q4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
| LightOnOCR-2-1BQ4 · Hugging Face count · source | ~3 GB | Q4 · Hugging Face count |
Tight fit: it runs, but leaves less than 1 GB free beyond the 2 GB the system keeps (the memory figure already includes that reserve).
Also runs slowly
Above the 4 GB of VRAM, the layers that do not fit run on the CPU from PC RAM, up to 28 GB in all. 32 GB of PC RAM assumed: this configuration lists none. Planning estimate, not a speed figure.
- Gemma 4 E2BGoogle~6 GBruns well with 6 GB
- Gemma 4 E4BGoogle~8 GBQ4 · published params
- Qwen3.5 9BAlibaba~9 GBQ4 · published params
- Gemma 4 12BGoogle~11 GBQ4 · published params
- Gemma 4 26B A4BGoogle~21 GBQ4 · published params
- Qwen3.6 27BAlibaba~23 GBQ4 · published params
- Qwen3.8 27BAlibaba~23 GBQ4 · published params
- Smaug MiniAbacus AI~23 GBQ4 · published params
- Gemma 4 31BGoogle~26 GBQ4 · published params
- DeepSeek-OCRdeepseek-ai~5 GBruns well with 6 GB
- dots.mocrdots-studio~5 GBruns well with 6 GB
- dots.ocrdots-studio~5 GBruns well with 6 GB
- granite-4.0-3b-visionibm-granite~5 GBruns well with 6 GB
- granite-vision-3.3-2bibm-granite~5 GBruns well with 6 GB
- granite-vision-4.1-4bibm-granite~5 GBruns well with 6 GB
- Nanonets-OCR-snanonets~5 GBruns well with 6 GB
- Nanonets-OCR2-3Bnanonets~5 GBruns well with 6 GB
- paligemma2-3b-mix-224google~5 GBruns well with 6 GB
- Unlimited-OCRbaidu~5 GBruns well with 6 GB
- chandra-ocr-2datalab-to~6 GBruns well with 6 GB
- Fara1.5-4Bmicrosoft~6 GBruns well with 6 GB
- GELab-Zero-4B-preview-Sico-Evolutionmicrosoft~6 GBruns well with 6 GB
- gemma-3-4b-itgoogle~6 GBruns well with 6 GB
- Holo-3.1-4BHcompany~6 GBruns well with 6 GB
- Holo2-4BHcompany~6 GBruns well with 6 GB
- Mage-VLmicrosoft~6 GBruns well with 6 GB
- medgemma-4b-itgoogle~6 GBruns well with 6 GB
- Molmo2-4Ballenai~6 GBruns well with 6 GB
- MolmoMotion-4B-H1-F32allenai~6 GBruns well with 6 GB
- MolmoMotion-4B-H3-F30allenai~6 GBruns well with 6 GB
- Phi-3.5-vision-instructmicrosoft~6 GBruns well with 6 GB
- POINTS-Readertencent~6 GBruns well with 6 GB
- Qianfan-OCRbaidu~6 GBruns well with 6 GB
- Qwen3-VL-4B-InstructQwen~6 GBruns well with 6 GB
- Recon2Reason-Reasoning-4BBAAI~6 GBruns well with 6 GB
- translategemma-4b-itgoogle~6 GBruns well with 6 GB
- Youtu-VL-4B-Instructtencent~6 GBruns well with 6 GB
- Phi-4-multimodal-instructmicrosoft~7 GBQ4 · Hugging Face count
- gemma-3n-E4B-itgoogle~8 GBQ4 · Hugging Face count
- InternVL3-8BOpenGVLab~8 GBQ4 · Hugging Face count
- llava-1.5-7b-hfllava-hf~8 GBQ4 · Hugging Face count
- Molmo-7B-D-0924allenai~8 GBQ4 · Hugging Face count
- aya-vision-8bCohereLabs~9 GBQ4 · Hugging Face count
- Bunny-v1_1-Llama-3-8B-VBAAI~9 GBQ4 · Hugging Face count
- chandradatalab-to~9 GBQ4 · Hugging Face count
- Cosmos-Reason2-8Bnvidia~9 GBQ4 · Hugging Face count
- Eagle2.5-8Bnvidia~9 GBQ4 · Hugging Face count
- Emu3-Chat-hfBAAI~9 GBQ4 · Hugging Face count
- Fara-7Bmicrosoft~9 GBQ4 · Hugging Face count
- Idefics3-8B-Llama3HuggingFaceM4~9 GBQ4 · Hugging Face count
- InternVL3_5-8BOpenGVLab~9 GBQ4 · Hugging Face count
- Keye-VL-1_5-8BKwai-Keye~9 GBQ4 · Hugging Face count
- Llama-3.1-Nemotron-Nano-VL-8B-V1nvidia~9 GBQ4 · Hugging Face count
- LLaVA-OneVision-1.5-8B-Instructlmms-lab~9 GBQ4 · Hugging Face count
- llava-onevision-qwen2-7b-ov-hfllava-hf~9 GBQ4 · Hugging Face count
- MiniCPM-V-2_6openbmb~9 GBQ4 · Hugging Face count
- MiniCPM-V-4_5openbmb~9 GBQ4 · Hugging Face count
- Molmo2-8Ballenai~9 GBQ4 · Hugging Face count
- moondream3-previewmoondream~9 GBQ4 · Hugging Face count
- nemotron-labs-audio-visual-flamingo-hfnvidia~9 GBQ4 · Hugging Face count
- olmOCR-2-7B-1025allenai~9 GBQ4 · Hugging Face count
- olmOCR-7B-0225-previewallenai~9 GBQ4 · Hugging Face count
- POINTS-GUI-Gtencent~9 GBQ4 · Hugging Face count
- Qwen2-VL-7B-InstructQwen~9 GBQ4 · Hugging Face count
- Qwen2.5-VL-7B-InstructQwen~9 GBQ4 · Hugging Face count
- Qwen3-VL-8B-InstructQwen~9 GBQ4 · Hugging Face count
- UI-TARS-1.5-7BByteDance-Seed~9 GBQ4 · Hugging Face count
- Video-XL-2BAAI~9 GBQ4 · Hugging Face count
- Fara1.5-9Bmicrosoft~10 GBQ4 · Hugging Face count
- GLM-4.6V-Flashzai-org~10 GBQ4 · Hugging Face count
- Llama-3.2-11B-Visionmeta-llama~10 GBQ4 · Hugging Face count
- Perception-LM-8Bfacebook~10 GBQ4 · Hugging Face count
- Step3-VL-10Bstepfun-ai~10 GBQ4 · Hugging Face count
- gemma-4-12Bgoogle~11 GBQ4 · Hugging Face count
- Llama-3.2-11B-Vision-Instructmeta-llama~11 GBQ4 · Hugging Face count
- Llama-Guard-3-11B-Visionmeta-llama~11 GBQ4 · Hugging Face count
- Llama-Guard-4-12Bmeta-llama~11 GBQ4 · Hugging Face count
- cogvlm2-llama3-captionzai-org~12 GBQ4 · Hugging Face count
- NVIDIA-Nemotron-Nano-12B-v2-VL-BF16nvidia~12 GBQ4 · Hugging Face count
- Apriel-1.6-15b-ThinkerServiceNow-AI~14 GBQ4 · Hugging Face count
- Phi-4-reasoning-vision-15Bmicrosoft~14 GBQ4 · Hugging Face count
- diffusiongemma-26B-A4B-itgoogle~22 GBQ4 · Hugging Face count
- deepseek-vl2deepseek-ai~23 GBQ4 · Hugging Face count
- Fara1.5-27Bmicrosoft~23 GBQ4 · Hugging Face count
- gemma-3-27b-itgoogle~23 GBQ4 · Hugging Face count
- ERNIE-4.5-VL-28B-A3B-PTbaidu~25 GBQ4 · Hugging Face count
- ERNIE-4.5-VL-28B-A3B-Thinkingbaidu~25 GBQ4 · Hugging Face count
- Keye-VL-2.0-30B-A3BKwai-Keye~26 GBQ4 · Hugging Face count
- Qwen3-VL-30B-A3B-InstructQwen~26 GBQ4 · Hugging Face count
- gemma-4-31Bgoogle~27 GBQ4 · Hugging Face count
- Qwen2.5-VL-32B-InstructQwen~28 GBQ4 · Hugging Face count
With 6 GB you also run… 6
The next memory size on this ladder runs these well. Figures are the requirement on this hardware type.
- CLIP-ViT-H-14-laion2B-s32B-b79Klaion~5 GB16-bit est. · Hugging Face count
- siglip-so400m-patch14-384google~5 GB16-bit est. · Hugging Face count
- siglip2-so400m-patch14-384google~5 GB16-bit est. · Hugging Face count
- tipsv1-so400m14google~5 GB16-bit est. · Hugging Face count
- tipsv1-g14google~6 GB16-bit est. · Hugging Face count
- tipsv1-g14-lowresgoogle~6 GB16-bit est. · Hugging Face count
12 more vision models in the catalog have no usable memory figure or parameter count for this profile. Browse Vision models
Image
Image · Prompt alignment · GenEval
0–1 · higher is better
Source: GenEval · 2023 · reviewed 2026-10-05 · Published quality in the original benchmark preset; not measured on this hardware.
| Stable Diffusion XL 1.0Platform path documented · stable-diffusion.cpp · source | 0.55 | Not on the LMArena text-to-image leaderboard of 2026-10-07 | Documented path without a memory figure | Platform path documented |
|---|---|---|---|---|
| Stable Diffusion 1.5Platform path documented · stable-diffusion.cpp · source | 0.43 | Not on the LMArena text-to-image leaderboard of 2026-10-07 | Documented path without a memory figure | Platform path documented |
| Qwen-ImageConditional path · CUDA · source | No comparable published result | 1,057 ± 3 | ~4 GB | Conditional path |
| FLUX.2-small-decoder16-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~3 GB | 16-bit est. · params |
| OmniGen V2Conditional path · CUDA · source | No comparable published result | Not on the LMArena text-to-image leaderboard of 2026-10-07 | ~3 GB | Conditional path |
| shap-e-img2img16-bit est. · params · source | No comparable published result | Not reviewed for LMArena | ~3 GB | 16-bit est. · params |
| FLUX.1 [dev]Platform path documented · Diffusers / PyTorch · source | No comparable published result | Arena rated the FP8 build (flux-1-dev-fp8); this entry is the original FLUX.1 [dev] weights | Documented path without a memory figure | Platform path documented |
| FLUX.1 [schnell]Platform path documented · Diffusers / PyTorch · source | No comparable published result | Not on the LMArena text-to-image leaderboard of 2026-10-07 | Documented path without a memory figure | Platform path documented |
LMArena ratings in this table: LMArena 2026-10-07 · collected 2026-10-25 · dataset · CC BY 4.0
With 6 GB you also run… 7
The next memory size on this ladder. Figures are the requirement on this hardware type.
- stable-diffusion-v1-4CompVis~5 GB16-bit est. · params
- stable-fast-3dstabilityai~5 GB16-bit est. · params
- URSA-0.6B-IBQ1024BAAI~5 GB16-bit est. · params
- Sana Sprint 1.6BNVIDIA~6 GB16-bit est. · params
- Sana_1600M_1024pxEfficient-Large-Model~6 GB16-bit est. · params
- TripoSGVAST-AI~6 GB16-bit est. · params
- VGGT-1Bfacebook~6 GB16-bit est. · params
17 more image models in the catalog have no usable memory figure or parameter count for this profile. Browse Image models
Video
| animatediff-motion-adapter-v1-5-216-bit est. · params · source | ~4 GB | 16-bit est. · params |
|---|---|---|
| URSA-0.6B-FSQ32016-bit est. · params · source | ~4 GB | 16-bit est. · params |
| Wan2.1 T2V 14BPlatform path documented · Wan2.1 / PyTorch ≥2.4.0 · source | Documented path without a memory figure | Platform path documented |
How this list works Documented execution paths on this hardware first, then publisher minimums (CUDA or Apple MPS).
With 6 GB you also run… 6
The next memory size on this ladder. Figures are the requirement on this hardware type.
- CogVideoX-5BZ.ai / Tsinghua KEG~5 GBPublished requirement
- i2vgen-xlali-vilab~6 GB16-bit est. · params
- stable-video-diffusion-img2vidstabilityai~6 GB16-bit est. · params
- stable-video-diffusion-img2vid-xtstabilityai~6 GB16-bit est. · params
- stable-video-diffusion-img2vid-xt-1-1stabilityai~6 GB16-bit est. · params
- text-to-video-ms-1.7bali-vilab~6 GB16-bit est. · params
20 more video models in the catalog have no usable memory figure or parameter count for this profile. Browse Video models
Audio
Audio · Word error rate · Open ASR
% · lower is better
Source: Word error rate · Open ASR · reviewed 2026-10-20 · Published quality in the original benchmark preset; not measured on this hardware.
| Whisper large-v3-turboPlatform path documented · whisper.cpp / CUDA · source | 7.83 | Documented path without a memory figure | Platform path documented |
|---|---|---|---|
| Whisper large-v3Platform path documented · whisper.cpp / CUDA · source | 7.44 | Documented path without a memory figure | Platform path documented |
| canary-1b-flash16-bit est. · params · source | No comparable published result | ~4 GB | 16-bit est. · params |
| chatterbox16-bit est. · params · source | No comparable published result | ~4 GB | 16-bit est. · params |
| Chatterbox16-bit est. · params · source | No comparable published result | ~4 GB | 16-bit est. · params |
| CosyVoice2-0.5B16-bit est. · params · source | No comparable published result | ~4 GB | 16-bit est. · params |
| distil-large-v316-bit est. · params · source | No comparable published result | ~4 GB | 16-bit est. · params |
| distil-large-v3.516-bit est. · params · source | No comparable published result | ~4 GB | 16-bit est. · params |
With 6 GB you also run… 22
The next memory size on this ladder. Figures are the requirement on this hardware type.
- canary-1b-v2nvidia~5 GB16-bit est. · params
- indic-parler-ttsai4bharat~5 GB16-bit est. · params
- mms-1b-allfacebook~5 GB16-bit est. · params
- MOSS-Transcribe-DiarizeOpenMOSS-Team~5 GB16-bit est. · params
- omniASR-CTC-1Bfacebook~5 GB16-bit est. · params
- parakeet-ctc-1.1bnvidia~5 GB16-bit est. · params
- parler-tts-mini-v1parler-tts~5 GB16-bit est. · params
- Qwen3-ASR-0.6BQwen~5 GB16-bit est. · params
- Qwen3-TTS-12Hz-0.6B-CustomVoiceQwen~5 GB16-bit est. · params
- VibeVoice-Realtime-0.5Bmicrosoft~5 GB16-bit est. · params
- whisper-mediumopenai~5 GBPublisher minimum
- whisper-medium.enopenai~5 GBPublisher minimum
- Breeze-ASR-26MediaTek-Research~6 GB16-bit est. · params
- csm-1bsesame~6 GB16-bit est. · params
- Dia-1.6Bnari-labs~6 GB16-bit est. · params
- LFM2.5-Audio-1.5BLiquidAI~6 GB16-bit est. · params
- Ming-omni-tts-0.5BinclusionAI~6 GB16-bit est. · params
- musicgen-mediumfacebook~6 GB16-bit est. · params
- musicgen-melodyfacebook~6 GB16-bit est. · params
- VibeVoice 1.5BMicrosoft~6 GB16-bit est. · params
- Zonos-v0.1Zyphra~6 GBPublished requirement
- Zonos-v0.1-transformerZyphra~6 GB16-bit est. · params
33 more audio models in the catalog have no usable memory figure or parameter count for this profile. Browse Audio models
Embeddings
Embeddings · MTEB (Multilingual, v2) · mean(task)
0–100 · higher is better
- Color = publisher
| Harrier OSS v1 (0.6B)8-bit · published params · source | 69.0 | ~3 GB | 8-bit · published params |
|---|---|---|---|
| Qwen3-Embedding 0.6B8-bit · published params · source | 64.3 | ~3 GB | 8-bit · published params |
| Multilingual E5 Large Instruct8-bit · published params · source | 63.2 | ~3 GB | 8-bit · published params |
| EmbeddingGemma 28-bit · published params · source | 61.4 | ~3 GB | 8-bit · published params |
| EmbeddingGemma (300M)8-bit · published params · source | 61.1 | ~3 GB | 8-bit · published params |
| BGE-M38-bit · published params · source | 59.6 | ~3 GB | 8-bit · published params |
| Nomic Embed Text v2 (MoE)8-bit · published params · source | 57.6 | ~3 GB | 8-bit · published params |
| Multilingual E5 Base8-bit · published params · source | 57.0 | ~3 GB | 8-bit · published params |
With 6 GB you also run… 7
The next memory size on this ladder runs these well. Figures are the requirement on this hardware type.
- gme-Qwen2-VL-2B-InstructAlibaba-NLP~5 GB8-bit · Hub count
- granite-speech-4.1-2b-naribm-granite~5 GB8-bit · Hub count
- llama-nemotron-embed-vl-1b-v2nvidia~5 GB8-bit · Hub count
- Qwen3-VL-Embedding-2BQwen~5 GB8-bit · Hub count
- Qwen3-VL-Reranker-2BQwen~5 GB8-bit · Hub count
- Youtu-Embeddingtencent~5 GB8-bit · Hub count
- MiniCPM-Embeddingopenbmb~6 GB8-bit · Hub count
Runs well within 4 GB of VRAM · planning estimates, not measured speed
Measurements
Speed on this hardware
We have not yet included a verifiable measurement for this exact configuration. The memory estimates above do not indicate tokens per second or time to first token.Specifications
15 sourced fields; the source is named once per group, and a field from another source links to it. Official specifications
Marker: the lightest text model that does not run well here but runs well with 6 GB. No text model that runs well here has a LiveBench score. Q4 planning estimate, not a measured fit.
Reviewed · Manufacturer source
- Graphics processor
- NVIDIA GeForce GTX 1650
- GPU runtime status
- Not stated by the source
- Memory capacity
- 4 GB
- Memory kind
- Dedicated VRAM
- Power connectors
- 6-pin
- GPU architecture
- Turing
- CUDA cores
- 896
- GPU memory type
- GDDR5 · GDDR6
- GPU power
- 75 W
- Memory interface
- 128 bit
- Minimum power supply
- 300 W
- PCIe interface
- Gen 3
- Slot width
- 2
NVIDIA GeForce GTX 1650 with 4 GB VRAM. Runtime support and local inference speed are not verified.
Photo: ASUS partner-board photo of a GeForce GTX 1650 · other partner boards vary. · Manufacturer source
Memory bandwidth is not measured model speed; verify runtime and checkpoint compatibility. Hardware specification reviewed: 2026-10-13.
How estimates work
- Candidates
- Reviewed pages use published parameter counts; automatic text candidates use official repository tensor counts (packed quantized weights and unsupported architectures excluded).
- Planning
- Q4 weights + 20% allowance + reserve: 25% of unified or system memory (at least 2 GB, at most 8 GB; 2 GB here) or 2 GB per dedicated GPU. Adjustable in “Adjust the estimate”. No KV cache is calculated from context.
- Fit tiers
- Runs well: the estimate fits the fast memory (VRAM, or unified memory within the macOS GPU share). Runs slowly: within the machine's memory but outside the fast part; a graphics card offloads layers to PC RAM (the listed RAM, or 32 GB assumed). Doesn't run: above both. Tiers are planning estimates, never a speed.
- Vision
- Vision-language models use the Q4 planning estimate from the parameter count (published by the model page, or read from Hugging Face). Classic vision models (classify, detect, segment) use a 16-bit estimate. No vision benchmark is tracked, so rows carry no score.
- Media and embeddings
- Media rows use documented execution paths (exact hardware or platform) and publisher minimums; a published requirement above this memory moves a model to the next tier. Without either, a 16-bit estimate from the published parameter count (× 1.2 + 2 GB), labelled as such. Embeddings use 8-bit weights + 20% + 2 GB.
Local AI basics: parameters, Q4 and memory · Full model catalog · Data and methodology
Alternatives
Other machines on the same memory ladder (VRAM); each page lists the models that fit it.
