Which Ollama models fit in 8 GB of VRAM?

Qwen3 0.6B, Qwen3.5 0.8B, Gemma 3 1B, Llama 3.2 1B Instruct, Qwen3 1.7B, Qwen3.5 2B, Llama 3.2 3B Instruct, Granite 4.1 3B, Phi-4-mini 3.8B, Ministral 3 3B, Qwen3 4B, Gemma 3 4B, Qwen3.5 4B, Gemma 4 E2B, Mistral 7B Instruct v0.3, DeepSeek-R1-Distill-Qwen 7B, Qwen2.5-Coder 7B, Gemma 4 E4B, LFM2.5 8B-A1B run with headroom on an 8 GB card at Ollama's default Q4_K_M with an 8K context.

Which Ollama models fit in 12 GB of VRAM?

Gemma 3 12B, Mistral NeMo 12B need a 12 GB card; everything that fits 8 GB fits too.

Which Ollama models fit in 24 GB of VRAM?

Gemma 4 26B-A4B, Gemma 3 27B, Qwen3.8 27B, Qwen3.6 27B, Granite 4.1 30B, Muse Glimmer 30B, Qwen3 30B-A3B, Qwen3 Coder 30B-A3B, GLM-4.7-Flash 30B-A3B, Nemotron 3.5 Lightning 30B-A3B need a 24 GB card, plus everything that fits a smaller one.

What quantisation does Ollama download by default?

Q4_K_M for library tags without a suffix, which is about 4.8 bits per weight: a 9B model is roughly 5.8 GB on disk. Tags such as :q8_0 pick a different quantisation; gpt-oss ships its native MXFP4 instead.

Ollama models

Ollama models: which ones fit your GPU?

Updated 8 Oct 2026 1 new this month

62 of the 80 models in the catalogue are in the Ollama library. Each row is the tag to pull, the size of the weights Ollama downloads by default (its library tags are Q4_K_M unless the model ships a native format), and the smallest card that runs it with headroom at an 8K context. Pick a device to see the verdict and estimated speed on it.

Ollama's default context is 4K tokens, not the model's maximum. Set OLLAMA_CONTEXT_LENGTH (every model page here prints the line with the size that fits) or the model silently forgets the start of long prompts. The KV cache for that context is the part of the memory bill the download size does not show — the KV cache calculator prices it.

On this device

Any RTX 4090 RTX 3060 12 GB RX 7900 XTX M3 Max · 64 GB

any other device →

Model	Pull	Params	Download	Max context	Needs	On Radeon RX 7900 XTX
Qwen3 0.6B Useful mostly as a speculative-decoding draft model for its larger siblings.	ollama pull qwen3:0.6b	0.6B	0.3 GB	32K	8 GB	1723 tok/s
Qwen3.5 0.8B Draft model for speculative decoding, or a classifier that fits in 1 GB.	ollama pull qwen3.5:0.8b	0.87B	0.5 GB	256K	8 GB	1188 tok/s
Gemma 3 1B Text-only. Single KV head makes its cache almost free at long context.	ollama pull gemma3:1b	1.0B	0.6 GB	32K	8 GB	1034 tok/s
Llama 3.2 1B Instruct The smallest Llama worth running. Fits anywhere, including phones and 4 GB cards.	ollama pull llama3.2:1b	1.24B	0.7 GB	128K	8 GB	834 tok/s
Qwen3 1.7B Punches above its size on structured tasks, with optional thinking mode.	ollama pull qwen3:1.7b	1.72B	1.0 GB	32K	8 GB	601 tok/s
Qwen3.5 2B Phone-class, and multimodal. Replaces Llama 3.2 3B as the "it runs on anything" answer.	ollama pull qwen3.5:2b	2.27B	1.3 GB	256K	8 GB	455 tok/s
Llama 3.2 3B Instruct The 2024 "it just runs" model for 8 GB laptops. Qwen3.5 4B does the same job better now.	ollama pull llama3.2:3b	3.21B	1.8 GB	128K	8 GB	322 tok/s
Granite 4.1 3B Dense, small, enterprise-flavoured: tool calling and instruction following, no thinking mode.	ollama pull granite4.1:3b	3.4B	1.9 GB	128K	8 GB	304 tok/s
Phi-4-mini 3.8B MIT-licensed, dense, and unusually strong on instruction following for its size.	ollama pull phi4-mini:3.8b	3.84B	2.2 GB	128K	8 GB	269 tok/s
Ministral 3 3B Edge model with a vision encoder and a 256K window. Apache 2.0.	ollama pull ministral-3:3b	3.85B	2.2 GB	256K	8 GB	268 tok/s
Qwen3 4B The 2025 sweet spot for 8 GB cards with reasoning traces. Qwen3.5 4B adds vision and 8× the context.	ollama pull qwen3:4b	4.02B	2.3 GB	32K	8 GB	257 tok/s
Gemma 3 4B Vision-capable at 4B. Superseded by Gemma 4 E4B, still everywhere.	ollama pull gemma3:4b	4.3B	2.4 GB	128K	8 GB	240 tok/s
Qwen3.5 4B The 8 GB coding agent. Q4 lands near 3.4 GB, leaving room for a real context window.	ollama pull qwen3.5:4b	4.66B	2.6 GB	256K	8 GB	222 tok/s
Gemma 4 E2B "E2B" is 2.3B effective, but the file holds 5B because of per-layer embeddings — size it as 5B. Text, image and audio in.	ollama pull gemma4:e2b	5.1B	2.9 GB	128K	8 GB	203 tok/s
Mistral 7B Instruct v0.3 Old but extremely well behaved, and permissively licensed for commercial use.	ollama pull mistral:7b	7.25B	4.1 GB	32K	8 GB	143 tok/s
Olmo 3 7B Instruct Fully open — training data and code included. Full multi-head attention, so its cache is 4× a GQA 7B at the same context.	ollama pull olmo-3:7b	7.3B	4.1 GB	64K	10 GB	142 tok/s
DeepSeek-R1-Distill-Qwen 7B Reasoning traces on a 7B budget. Expect long outputs — budget context accordingly.	ollama pull deepseek-r1:7b	7.62B	4.3 GB	128K	8 GB	136 tok/s
Qwen2.5-Coder 7B The standard local autocomplete model — small enough to keep resident all day, and still the best FIM model under 8B.	ollama pull qwen2.5-coder:7b	7.62B	4.3 GB	128K	8 GB	136 tok/s
Gemma 4 E4B The laptop Gemma. 4.5B effective, 8B on disk; a single KV head per window layer keeps its cache tiny.	ollama pull gemma4:e4b	8.0B	4.5 GB	128K	8 GB	129 tok/s
Llama 3.1 8B Instruct Still the most widely deployed local model, with the largest fine-tune ecosystem. Not the strongest 8B any more.	ollama pull llama3.1:8b	8.03B	4.5 GB	128K	10 GB	129 tok/s
Qwen3 8B Apache-2.0 alternative to Llama 3.1 8B, with a switchable thinking mode.	ollama pull qwen3:8b	8.19B	4.6 GB	128K	10 GB	126 tok/s
LFM2.5 8B-A1B An 8B MoE with ~1.5B active, aimed at laptops without a GPU. Licence is permissive below $10M revenue.	ollama pull lfm2.5:8b	8.47B MoE	4.8 GB	125K	8 GB	265 tok/s
Granite 4.1 8B Matches the old Granite 4.0 32B MoE at a quarter of the size. Fast, Apache 2.0, no reasoning traces.	ollama pull granite4.1:8b	8.79B	4.9 GB	128K	10 GB	118 tok/s
Ministral 3 8B Mistral's 8B with images in. Plain GQA, so budget more KV cache than Qwen3.5 9B at the same context.	ollama pull ministral-3:8b	8.92B	5.0 GB	256K	10 GB	116 tok/s
Ornith 1.5 9B The small Ornith: a coding-agent reasoning build on the Qwen3.5 9B architecture. Same VRAM as its base, thinks before every answer.	ollama pull ornith-1.5:9b	9.41B	5.3 GB	256K	10 GB	110 tok/s
Qwen3.5 9B The default for 8–12 GB cards in 2026: beats every older 8B on every published benchmark, with vision.	ollama pull qwen3.5:9b	9.65B	5.4 GB	256K	10 GB	107 tok/s
Gemma 4 12B The "unified" Gemma 4: text, image and audio in one 12B that fits a 12 GB card at Q4. 140+ languages.	ollama pull gemma4:12b	12B	6.7 GB	256K	11 GB	86 tok/s
Gemma 3 12B Strong multilingual chat with images, sized for 12–16 GB cards. Gemma 4 12B is the same size and better.	ollama pull gemma3:12b	12.2B	6.9 GB	128K	12 GB	85 tok/s
Mistral NeMo 12B Multilingual 12B with a 128K window, built with NVIDIA. A roleplay and fiction favourite that refuses to die.	ollama pull mistral-nemo:12b	12.2B	6.9 GB	128K	12 GB	85 tok/s
Ministral 3 14B The largest Ministral. A 12 GB card runs it at Q4 with a few gigabytes to spare.	ollama pull ministral-3:14b	13.9B	7.8 GB	256K	16 GB	74 tok/s
Phi-4 14B Trained heavily on synthetic reasoning data. Short 16K window is its main limitation.	ollama pull phi4:14b	14.7B	8.3 GB	16K	16 GB	70 tok/s
Qwen3 14B The largest Qwen3 that fits a 12 GB card at Q4 with room for context.	ollama pull qwen3:14b	14.8B	8.3 GB	128K	16 GB	70 tok/s
DeepSeek-R1-Distill-Qwen 14B The 2025 reasoning-per-gigabyte pick for a 12 GB card. Qwen3.5 9B in thinking mode has since overtaken it.	ollama pull deepseek-r1:14b	14.8B	8.3 GB	128K	16 GB	70 tok/s
Qwen2.5-Coder 14B Noticeably better at whole-file edits than the 7B, still comfortable on 12 GB.	ollama pull qwen2.5-coder:14b	14.8B	8.3 GB	128K	16 GB	70 tok/s
gpt-oss 20B Ships natively in MXFP4, so the 4-bit weights are the reference weights, not a lossy copy. Fits 16 GB.	ollama pull gpt-oss:20b	20.9B MoE	10.8 GB	128K	16 GB	120 tok/s
Mistral Small 3.2 24B Apache-2.0, vision-capable, and the most 24 GB-friendly of the 2025 generalists.	ollama pull mistral-small:24b	23.6B	13.3 GB	128K	20 GB	44 tok/s
Gemma 4 26B-A4B Mixture of experts with 3.8B active. Slower to think than Qwen3.6 35B-A3B, faster to answer, and it sees images.	ollama pull gemma4:26b	26.5B MoE	14.9 GB	256K	24 GB	105 tok/s
Gemma 3 27B The 2025 single-GPU generalist with vision. Its Gemma-licence terms are the reason to prefer Gemma 4 now.	ollama pull gemma3:27b	27.4B	15.4 GB	128K	24 GB	38 tok/s
Qwen3.8 27B The current default local Qwen: dense 27B, text + image + video, 262K context. Only 16 of its 64 blocks keep a KV cache, so long context is cheap.	ollama pull qwen3.8:27b	27.8B	15.6 GB	256K	24 GB	37 tok/s
Qwen3.6 27B The 24 GB coding pick of spring 2026 (77.2 SWE-bench Verified). Same shape as 3.8, one generation behind.	ollama pull qwen3.6:27b	27.8B	15.6 GB	256K	24 GB	37 tok/s
Granite 4.1 30B The largest Granite. Dense 29B at Q4 is a comfortable 24 GB fit.	ollama pull granite4.1:30b	28.9B	16.3 GB	128K	24 GB	36 tok/s
Muse Glimmer 30B Meta's first open weights since Llama 4: a dense 30B distilled from Muse Spark for always-on local agents. Two KV heads keep the cache small.	ollama pull muse-glimmer:30b	29.8B	16.8 GB	128K	24 GB	35 tok/s
Qwen3 30B-A3B The MoE that made "3B active" a category. Qwen3.6 35B-A3B is its direct replacement.	ollama pull qwen3:30b-a3b	30.5B MoE	17.1 GB	128K	24 GB	120 tok/s
Qwen3 Coder 30B-A3B Agentic coding MoE with a 256K native window. Still the most-downloaded local code model.	ollama pull qwen3-coder:30b	30.5B MoE	17.1 GB	256K	24 GB	120 tok/s
GLM-4.7-Flash 30B-A3B MIT-licensed 30B-A3B tuned for agentic coding, with a DeepSeek-style latent KV cache. 60–80 tok/s reported on a 4090.	ollama pull glm-4.7-flash:latest	31.2B MoE	17.5 GB	198K	24 GB	133 tok/s
Gemma 4 31B The dense flagship: strongest maths of the 24–32 GB class (89% AIME), clean prose, vision. Q4 is a tight 24 GB fit.	ollama pull gemma4:31b	31.3B	17.6 GB	256K	32 GB	33 tok/s
Nemotron 3.5 Lightning 30B-A3B Mamba-2 + MoE hybrid built for the execution layer of agents: only 6 attention blocks, so the KV cache is almost free. Weights, data and recipe all open.	ollama pull nemotron-3.5-lightning:30b-a3b	31.6B MoE	17.8 GB	256K	24 GB	124 tok/s
Olmo 3.1 32B Instruct The largest fully open model you can audit end to end. Q4 fits 24 GB, tightly.	ollama pull olmo-3:32b	32.2B	18.1 GB	64K	32 GB	32 tok/s
Qwen3 32B The classic 24 GB target, and still the strongest local translator under 70B. Qwen3.8 27B is smaller and better at everything else.	ollama pull qwen3:32b	32.8B	18.4 GB	128K	32 GB	32 tok/s
DeepSeek-R1-Distill-Qwen 32B MIT-licensed and close to the 70B distill on maths. A 24 GB card handles it at Q4.	ollama pull deepseek-r1:32b	32.8B	18.4 GB	128K	32 GB	32 tok/s
Qwen2.5-Coder 32B The first local code model that felt competitive with hosted assistants. Qwen3.8 27B is smaller and far ahead on agentic work.	ollama pull qwen2.5-coder:32b	32.8B	18.4 GB	128K	32 GB	32 tok/s
Qwen3.6 35B-A3B Mixture of experts with ~3B active: the fastest serious model a 24 GB card runs, and the best MoE under 40B on agentic coding.	ollama pull qwen3.6:35b-a3b	35.9B MoE	20.2 GB	256K	32 GB	120 tok/s
Ornith 1.5 35B-A3B A reasoning-first MIT build on the Qwen3.6 35B-A3B architecture (thinks before every answer). Same VRAM as its base.	ollama pull ornith-1.5:35b	35.9B MoE	20.2 GB	256K	32 GB	120 tok/s
Llama 3.3 70B Instruct Still the creative-writing favourite: consistent voice, takes direction. Needs 48 GB to sit comfortably on GPU at Q4.	ollama pull llama3.3:70b	70.6B	39.7 GB	128K	64 GB	offload
DeepSeek-R1-Distill-Llama 70B The strongest of the R1 distills, and the one that most needs 48 GB or more.	ollama pull deepseek-r1:70b	70.6B	39.7 GB	128K	64 GB	offload
Llama 4 Scout 109B-A17B A 10M-token window on paper, chunked attention in practice (8K chunks on 3 of 4 layers). Needs 64 GB+ at Q4.	ollama pull llama4:scout	109B MoE	61.3 GB	1024K	80 GB	won't fit
gpt-oss 120B Designed to land on one 80 GB card. Only ~5B parameters are active per token.	ollama pull gpt-oss:120b	117B MoE	60.5 GB	128K	80 GB	won't fit
Qwen3.5 122B-A10B The 96–128 GB unified-memory model: 122B of knowledge at 10B-active speed.	ollama pull qwen3.5:122b-a10b	125B MoE	70.3 GB	256K	141 GB	won't fit
Qwen3.8-Flash-Next 180B-A6B The open preview of the Qwen4 architecture: a 125B-A6B hybrid (Gated DeltaNet + sparse attention, KV cache on 12 of 48 blocks) plus a 51B n-gram embedding and a 4B draft head — 180B on disk, 6B active. The hosted "Qwen3.8-Flash" is this model with a 1M window.	ollama pull qwen3.8-flash-next:125b-a6b-nvfp4	180B MoE	101.2 GB	256K	141 GB	won't fit
Qwen3 235B-A22B Workstation class. Realistically a 192 GB unified-memory or multi-GPU model.	ollama pull qwen3:235b-a22b	235B MoE	132.1 GB	128K	192 GB	won't fit
Ornith 1.5 397B-A17B The flagship Ornith on the Qwen3.5 397B-A17B architecture, MIT-licensed. A 256 GB Mac Studio at Q4, and it is in the Ollama library.	ollama pull ornith-1.5:397b	397B MoE	223.2 GB	256K	multi-GPU	won't fit
DeepSeek-R1 671B The January 2025 moment. Multi-head latent attention keeps its KV cache tiny; the weights do not.	ollama pull deepseek-r1:671b	671B MoE	377.3 GB	128K	multi-GPU	won't fit

Models that are not in the Ollama library

18 catalogue models have no library tag — mostly the newest or largest releases. Ollama can still run most of them straight from Hugging Face with ollama run hf.co/ORG/MODEL-GGUF:Q4_K_M once a GGUF conversion exists; the models index links each one's model page, which sizes it and shows the llama.cpp command that always works.

Which quant is the tag?

A bare tag like qwen3.5:9b is Q4_K_M. Suffixes such as :9b-q8_0 pick another quantisation; what they mean.

Not sure what fits?

Best local LLM by VRAM class picks one model per job for 8 to 128 GB.

Ollama vs the rest

Five runtimes compared, and what LM Studio needs for the same models.