Skip to content
Local AI Wizard
Your machine
Theme

RTX 3070 · 8 GB VRAM

NVIDIA · RTX 3070

48 text models run well · 448 GB/s bandwidth

Models on this hardware

Start here

Text

Within 8 GB: 48 run well · 29 tight · 13 reviewed · 83 more run slowly (40 of them run well with 12 GB)

Text · LMArena

rating · higher is better

  • Color = publisher

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.

Rows run well first, then tight fits; within each, reviewed models, then Hub downloads over 30 days, not quality. Sort a score column to rank by that source alone. All 40 text models that would run well with 12 GB but not here already run here, slowly.
Text models that run well within 8 GB
Granite 4.2 3BIBM · 3B · LMArena 1,290Q4 · published params · source1,290 ± 12~5 GBQ4 · published paramssource
Granite 4.2 8BIBM · 8B · LMArena 1,290Q4 · published params · source1,290 ± 11~8 GBTight fit: Needs ~8 GB: runs well, nothing left beyond the system reserveQ4 · published paramssource
Llama 3.1 8B InstructMeta · 8B · LMArena 1,211Q4 · published params · source1,211 ± 4~8 GBTight fit: Needs ~8 GB: runs well, nothing left beyond the system reserveQ4 · published paramssource
Llama 3.2 3B InstructMeta · 3.21B · LMArena 1,167Q4 · published params · source1,167 ± 8~5 GBQ4 · published paramssource
Llama 3.2 1B InstructMeta · 1.23B · LMArena 1,111Q4 · published params · source1,111 ± 8~3 GBQ4 · published paramssource
Qwen3 1.7BAlibaba · 1.7BQ4 · published params · sourceNot on the LMArena text leaderboard of 2026-10-22~4 GBQ4 · published paramssource
Gemma 4 E2BGoogle · 5.1BQ4 · published params · sourceNot on the LMArena text leaderboard of 2026-10-22~6 GBQ4 · published paramssource
Qwen3 4B Instruct 2507Alibaba · 4BQ4 · published params · sourceNot on the LMArena text leaderboard of 2026-10-22~5 GBQ4 · published paramssource
8 of 77

No model within 8 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 37

Every one that runs here, well or slowly, scored or not (13 run well, 24 slowly); 19 of 37 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.

Reviewed text models that run on RTX 3070 · 8 GB VRAM
Qwen3 1.7BAlibaba · 1.7BNot in LiveBench 2026-06-25Not on the LMArena text leaderboard of 2026-10-22~4 GB1.1Apache 2.04.1M
Llama 3.2 1B InstructMeta · 1.2BNot in LiveBench 2026-06-251,111~3 GB0.8Llama 3.23.5M
Gemma 4 E2BGoogle · 5.1BNot in LiveBench 2026-06-25Not on the LMArena text leaderboard of 2026-10-22~6 GB3.1Apache 2.03M
Qwen3 4B Instruct 2507Alibaba · 4BNot in LiveBench 2026-06-25Not on the LMArena text leaderboard of 2026-10-22~5 GB2.5Apache 2.02.6M
Llama 3.2 3B InstructMeta · 3.2BNot in LiveBench 2026-06-251,167~5 GB2.0Llama 3.21.1M
SmolLM3 3BHugging Face · 3BNot in LiveBench 2026-06-25Not on the LMArena text leaderboard of 2026-10-22~5 GB1.9Apache 2.0718.7K
Phi-4 miniMicrosoft · 3.8BNot in LiveBench 2026-06-25Not on the LMArena text leaderboard of 2026-10-22~5 GB2.5MIT342.8K
Ministral 3 3BMistral · 3.8BNot in LiveBench 2026-06-25Not on the LMArena text leaderboard of 2026-10-22~5 GB2.1Apache 2.087.4K
8 of 37
Also runs slowly59 models

Above the 8 GB of VRAM, the layers that do not fit run on the CPU from PC RAM, up to 32 GB in all. 32 GB of PC RAM assumed: this configuration lists none. Planning estimate, not a speed figure. The 24 reviewed models that run slowly are in the table above.

6 of 59

Browse Text models

Adjust the estimate: quantization, context, reserveloads the live catalog

Vision

Within 8 GB: 79 run well · 5 tight · 2 reviewed · 58 more run slowly (39 of them run well with 12 GB)

Vision · LMArena

rating · higher is better

  • Color = publisher

Source: LMArena 2026-10-19 · collected 2026-10-25 · dataset · CC BY 4.0 · Models without a rating are left out, not scored as zero.

Rows run well first, then tight fits; within each, reviewed models, then Hub downloads over 30 days, not quality. Sort a score column to rank by that source alone. All 39 vision models that would run well with 12 GB but not here already run here, slowly.
Vision models that run well within 8 GB
Molmo-7B-D-0924allenai · 8B · Image to text · LMArena 994Q4 · Hugging Face count · source994 ± 14~8 GBTight fit: Needs ~8 GB: runs well, nothing left beyond the system reserveQ4 · Hugging Face countsource
Phi-3.5-vision-instructmicrosoft · 4.1B · Image to text · LMArena 919Q4 · Hugging Face count · source919 ± 16~6 GBQ4 · Hugging Face countsource
Gemma 4 E2BGoogle · 5.1B · Image to textQ4 · published params · sourceNot on the LMArena vision leaderboard of 2026-10-19~6 GBQ4 · published paramssource
Phi-4-multimodal-instructmicrosoft · 5.6B · Image to textQ4 · Hugging Face count · sourceNot on the LMArena vision leaderboard of 2026-10-19~7 GBQ4 · Hugging Face countsource
siglip2-giant-opt-patch16-384google · 1.9B · Classic vision16-bit est. · Hugging Face count · sourceNot on the LMArena vision leaderboard of 2026-10-19~7 GB16-bit est. · Hugging Face countsource
chandra-ocr-2datalab-to · 5.3B · Image to textQ4 · Hugging Face count · sourceNot on the LMArena vision leaderboard of 2026-10-19~6 GBQ4 · Hugging Face countsource
Fara1.5-4Bmicrosoft · 4.5B · Image to textQ4 · Hugging Face count · sourceNot on the LMArena vision leaderboard of 2026-10-19~6 GBQ4 · Hugging Face countsource
GELab-Zero-4B-preview-Sico-Evolutionmicrosoft · 4.4B · Image to textQ4 · Hugging Face count · sourceNot on the LMArena vision leaderboard of 2026-10-19~6 GBQ4 · Hugging Face countsource
8 of 84

LMArena ratings in this table: LMArena 2026-10-19 · 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).

Also runs slowly58 models

Above the 8 GB of VRAM, the layers that do not fit run on the CPU from PC RAM, up to 32 GB in all. 32 GB of PC RAM assumed: this configuration lists none. Planning estimate, not a speed figure.

6 of 58

12 more vision models in the catalog have no usable memory figure or parameter count for this profile. Browse Vision models

Image

21 within 8 GB · 6 more with 12 GB

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.

Image models within 8 GB
Stable Diffusion XL 1.0Stability AI · 2.6B · Text to imagePlatform path documented · stable-diffusion.cpp · source0.55Not on the LMArena text-to-image leaderboard of 2026-10-07Documented path without a memory figurePlatform path documentedstable-diffusion.cpp · source
Stable Diffusion 1.5Stability AI · 0.98B · Text to imagePlatform path documented · stable-diffusion.cpp · source0.43Not on the LMArena text-to-image leaderboard of 2026-10-07Documented path without a memory figurePlatform path documentedstable-diffusion.cpp · source
URSA-1.7B-IBQ1024BAAI · 2.3B · Text to image16-bit est. · params · sourceNo comparable published resultNot reviewed for LMArena~8 GB16-bit est. · paramssource
URSA-1.7B-IBQ512BAAI · 2.3B · Text to image16-bit est. · params · sourceNo comparable published resultNot reviewed for LMArena~8 GB16-bit est. · paramssource
Lumina Image v2OpenGVLab · 2B · Text to image16-bit est. · params · sourceNo comparable published resultNot on the LMArena text-to-image leaderboard of 2026-10-07~7 GB16-bit est. · paramssource
Stable Diffusion 3 MediumStability AI · 2B · Text to image16-bit est. · params · sourceNo comparable published resultNot on the LMArena text-to-image leaderboard of 2026-10-07~7 GB16-bit est. · paramssource
stable-point-aware-3dstabilityai · 2B · 3D16-bit est. · params · sourceNo comparable published resultNot reviewed for LMArena~7 GB16-bit est. · paramssource
Sana Sprint 1.6BNVIDIA · 1.6B · Text to image16-bit est. · params · sourceNo comparable published resultNot on the LMArena text-to-image leaderboard of 2026-10-07~6 GB16-bit est. · paramssource
8 of 21

LMArena ratings in this table: LMArena 2026-10-07 · collected 2026-10-25 · dataset · CC BY 4.0

With 12 GB you also run… 6

The next memory size on this ladder. Figures are the requirement on this hardware type.

17 more image models in the catalog have no usable memory figure or parameter count for this profile. Browse Image models

Video

19 within 8 GB · 4 more with 12 GB
Video models within 8 GB
VACE-Wan2.1-1.3B-Previewali-vilab · 2.2B · Image to video16-bit est. · params · source~8 GB16-bit est. · paramssource
Wan2.1-VACE-1.3BWan-AI · 2.2B · Image to video16-bit est. · params · source~8 GB16-bit est. · paramssource
Wan2.1-VACE-1.3B-diffusersWan-AI · 2.2B · Image to video16-bit est. · params · source~8 GB16-bit est. · paramssource
CogVideoX-2bzai-org · 1.7B · Text to video16-bit est. · params · source~7 GB16-bit est. · paramssource
Kandinsky-WM-1.0-I2V-5s-PHkandinskylab · 2B · Image to video16-bit est. · params · source~7 GB16-bit est. · paramssource
LTX-Video 2B · 0.9.8 distilledLightricks · 2B · Image to video16-bit est. · params · source~7 GB16-bit est. · paramssource
LTX-Video-0.9.5Lightricks · 1.9B · Text to video16-bit est. · params · source~7 GB16-bit est. · paramssource
Matrix-Game-2.0Skywork · 1.8B · Image to video16-bit est. · params · source~7 GB16-bit est. · paramssource
8 of 19

How this list works Documented execution paths on this hardware first, then publisher minimums (CUDA or Apple MPS).

With 12 GB you also run… 4

The next memory size on this ladder. Figures are the requirement on this hardware type.

20 more video models in the catalog have no usable memory figure or parameter count for this profile. Browse Video models

Audio

91 within 8 GB · 14 more with 12 GB

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.

Audio models within 8 GB
Whisper large-v3-turboOpenAI · 0.81B · Speech to textPlatform path documented · whisper.cpp / CUDA · source7.83Documented path without a memory figurePlatform path documentedwhisper.cpp / CUDA · source
Whisper large-v3OpenAI · 1.55B · Speech to textPlatform path documented · whisper.cpp / CUDA · source7.44Documented path without a memory figurePlatform path documentedwhisper.cpp / CUDA · source
cohere-transcribe-03-2026CohereLabs · 2.1B · Speech to text16-bit est. · params · sourceNo comparable published result~8 GB16-bit est. · paramssource
cohere-transcribe-arabic-07-2026CohereLabs · 2.1B · Speech to text16-bit est. · params · sourceNo comparable published result~8 GB16-bit est. · paramssource
dots.tts-soardots-studio · 2.2B · Text to speech16-bit est. · params · sourceNo comparable published result~8 GB16-bit est. · paramssource
GLM-ASR-Nano-2512zai-org · 2.3B · Speech to text16-bit est. · params · sourceNo comparable published result~8 GB16-bit est. · paramssource
granite-4.0-1b-speechibm-granite · 2.3B · Speech to text16-bit est. · params · sourceNo comparable published result~8 GB16-bit est. · paramssource
granite-speech-4.1-2bibm-granite · 2.3B · Speech to text16-bit est. · params · sourceNo comparable published result~8 GB16-bit est. · paramssource
8 of 91

With 12 GB you also run… 14

The next memory size on this ladder. Figures are the requirement on this hardware type.

6 of 14

33 more audio models in the catalog have no usable memory figure or parameter count for this profile. Browse Audio models

Embeddings

75 run well within 8 GB · 11 more with 12 GB

Embeddings · MTEB (Multilingual, v2) · mean(task)

0–100 · higher is better

  • Color = publisher

Source: MTEB results (curated rows, sourced per model)

Embeddings models that run well within 8 GB
Qwen3-Embedding 4BAlibaba (Qwen)8-bit · published params · source69.5~7 GB8-bit · published paramssource
Harrier OSS v1 (0.6B)Microsoft8-bit · published params · source69.0~3 GB8-bit · published paramssource
Qwen3-Embedding 0.6BAlibaba (Qwen)8-bit · published params · source64.3~3 GB8-bit · published paramssource
Multilingual E5 Large Instructintfloat8-bit · published params · source63.2~3 GB8-bit · published paramssource
EmbeddingGemma 2Google8-bit · published params · source61.4~3 GB8-bit · published paramssource
EmbeddingGemma (300M)Google8-bit · published params · source61.1~3 GB8-bit · published paramssource
BGE-M3BAAI8-bit · published params · source59.6~3 GB8-bit · published paramssource
Nomic Embed Text v2 (MoE)Nomic AI8-bit · published params · source57.6~3 GB8-bit · published paramssource
8 of 75

With 12 GB you also run… 11

The next memory size on this ladder runs these well. Figures are the requirement on this hardware type.

6 of 11

Browse Embeddings models

Runs well within 8 GB of VRAM · planning estimates, not measured speed bundled catalog copy · 1,777 repositories · checked

Measurements

Speed on this hardware

no published measurement yetWe 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

19 sourced fields; the source is named once per group, and a field from another source links to it. Official specifications

GPU VRAM8 GBDedicated · system RAM does not add · source
Memory bandwidth448 GB/sManufacturer spec · not model speed · source
Launch price (MSRP)USD 499.00US · launched · source · not a tracked offer

Marker: the lightest text model that does not run well here but runs well with 12 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 RTX 3070
GPU runtime status
Not stated by the source
Memory capacity
8 GB
Memory kind
Dedicated VRAM
Power connectors
1x PCIe 8-pin / (adapter to 1x 12-pin included)
GPU architecture
Ampere
GPU base clock
1,500 MHz
GPU boost clock
1,730 MHz
CUDA compute capability
8.6
CUDA cores
5,888
GPU memory type
GDDR6
GPU power
220 W
Card length
242 mm
Memory interface
256 bit
Minimum power supply
650 W
PCIe interface
Gen 4
Release date
2020-10-29
Slot width
2
Introduced
2020

NVIDIA GeForce RTX 3070 with 8 GB VRAM. Runtime support and local inference speed are not verified.

Photo: ASUS partner-board photo of a GeForce RTX 3070 · 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.