Skip to content
Local AI Wizard
Your machine
Theme

Model data and methodology

How current the data is, how every estimate is computed, and what the site does not claim.

Data status

The only place where freshness is stated; app pages carry no status blocks.

2 of 4 sources use a saved copy. When a source fails, the Wizard keeps the last valid data and says so here.

Quality benchmark (LiveBench)
Using saved copy from 2026-10-18LiveBench 2026-06-25 · checked 2026-11-19 · valid as of 2026-10-18source
Model inventory (Hugging Face)
Currentchecked 2026-11-19 · valid as of 2026-11-19source
Hardware specifications
Currentchecked 2026-11-14 · valid as of 2026-11-14source
Prices and stock
Using saved copy from 2026-09-22checked 2026-11-19 · valid as of 2026-09-22 · saved copy: RTX 5090, DGX Sparksource

Memory formulas

Every memory number on the site is an estimate from one of these, or a publisher's documented figure.

Text models · Q4 planning estimate

Used for every memory column, chart and “runs on” list, in Models, Discover, Compare and the hardware profiles.

Decimal GB, rounded up
memory = ceil( parameters × 0.625 × 1.2 + reserve )
parameters
Billions of parameters, as the publisher states them or as counted from the repository's floating-point tensors. For mixture-of-experts models, all experts count; active parameters do not shrink the weights you store.
0.625
Bytes per parameter for a 4-bit file with its quantization scales (about 5 bits). A planning constant, not the size of any one file.
1.2
A 20% working allowance for activations and batch buffers.
reserve
25% of unified memory or system RAM for the system (operating system, apps), at least 2 GB and at most 8 GB: 8 GB → 2, 16 GB → 4, 24 GB → 6, 32 GB or more → 8. On a dedicated GPU, 2 GB per card. Without a chosen machine, Models, Discover and Compare show the 8 GB figure (a 32 GB or larger machine).
Fit tiers
  • Runs well: the estimate fits the fast memory (graphics card VRAM, or unified memory within the macOS default GPU share of about 75% plus the reserve).
  • Runs slowly: within the machine's memory but outside the fast part; a graphics card runs the layers that do not fit on the CPU from PC RAM, up to VRAM + RAM − the RAM's reserve (the listed RAM, or 32 GB assumed when a card has none listed).
  • Doesn't run: above both.
  • Hardware tables mark a model that runs well but leaves less than 10% of the fast memory (at least 1 GB) free beyond the reserve as a tight fit, and list it after the models with room to spare.
  • A tier is a planning estimate, never a speed.
Not included
The KV cache that grows with the context window, vision or audio components, and runtime overhead. The estimate is a floor for planning, not a measured requirement.
Example
27 × 0.625 × 1.2 + 8 = 28.25 rounded up: 29 GB · on 16 GB: 27 × 0.625 × 1.2 + 4 = 24.25 → 25 GBA 27B model on a 32 GB and a 16 GB machine

Q8 and FP16 · the custom-machine planner

Hardware › “Different hardware? Plan a custom machine”. It splits the estimate into parts you can adjust.

weights = parameters × bytes[Q4 0.625 | Q8 1.125 | FP16 2]
memory = ceil( weights + 20% × weights + cache + reserve )
bytes
Bytes per parameter by weight precision: Q4 0.625, Q8 1.125, FP16 2.
20% × weights
The same working allowance, now a separate line you can see.
cache
A budget in GB that you enter. It is not calculated from the context length, because the cache depends on the architecture.
reserve
Automatic: 25% of the planned unified or system memory (2 to 8 GB); 2 GB per dedicated GPU. A value you type overrides it.
Example
27B · Q4 29 GB · Q8 45 GB · FP16 73 GBUnified memory of 32 GB or more (8 GB reserve), no extra cache

Embedding models · 8-bit estimate with a small reserve

An encoder has no growing chat cache, so the chat reserve would overstate it (data v2).

memory = ceil( parameters × 1 × 1.2 + 2 )
1
Bytes per parameter: 8-bit weights, between the 16-bit small embedders and the 4-bit large ones that runtimes ship.
1.2
The same 20% working allowance.
2 GB reserve
Runtime process plus desktop headroom. A planning assumption: no embedder was benchmarked on hardware, and very long inputs (32K tokens) need more.
Publisher figure
When a publisher states its own memory (for example under 200 MB), it is shown separately and never replaces the estimate.
Example
0.74 × 1 × 1.2 + 2 = 2.89 rounded up: 3 GBA 0.74B encoder

Image, video and audio models · documented minimums, then a 16-bit estimate

First the lowest GPU memory a publisher documents, with its platform (CUDA or Apple MPS) and whether the path is direct or conditional (offload, quantization). A published requirement above a machine's memory moves the model to the next tier. A documented path on one platform is never carried to another.

When nothing is documented for the machine's platform, hardware pages show a planning estimate from the published parameter count, labelled “16-bit est.”. Resolution, frame count, text encoders and offload change the real requirement.

memory = ceil( parameters × 2 × 1.2 + 2 )

Evidence words

Four words, used the same way on every page.

Published
A number or fact a publisher or benchmark states, shown with its source and date: a LiveBench score, a parameter count, a context length, a license, a documented GPU memory.
Estimated
Computed here from published inputs with a formula above: memory and disk. A planning guide, never a test and never a compatibility promise.
Measured
A result from a named run: hardware, runtime, build, quantization, prompt and output length. Shown exactly as published, never normalized or ranked across setups. Today these are tokens per second from llama.cpp scoreboards and benchmark pages, the mlx-lm benchmark table and a few vendor posts. A machine can also show a “Same chip” reference row: a run on another machine with the same chip and the same memory size, named in the row and never presented as a run on this machine.
Not measured
Nothing is published for that checkpoint or hardware. It appears as a dash with the reason, never as zero and never filled in from a neighbor: a reference row comes only from the same chip with the same memory size, never from a different chip, memory size or model.

What we do not claim

That a model runs on your machine
“Fits” compares a memory estimate with a memory size. It is not a compatibility test and says nothing about speed.
Local speed from a score or a spec
Benchmark scores and memory bandwidth do not predict tokens per second. Speed appears only when someone measured it.
Quality of a quantized file
LiveBench scores the publisher's checkpoint, not the Q4 or Q8 build you would download.
A score where there is none
A model missing from a benchmark release has no score; it is not zero. Scores from different benchmarks or modalities are never merged into one index. LiveBench has scored no open model below about 27 B parameters since its 2025-04-25 table; older tables use six categories and another question set, so they are not mixed in. Models that fit 8–16 GB therefore have no LiveBench score (current table).Checked 2026-10-10 across every published LiveBench table
Current prices
A price appears only as an observed offer with its date, seller and condition, and goes stale after 7 days. Launch prices are labelled historical.
Advice
Nothing here is purchase, legal or licensing advice. Licenses belong to their publishers; read them before use.

Rights

Whose scores, pictures and weights appear here, and under what terms.

Attribution and reuse by source8
Artificial Analysis
Excluded: its free tier allows internal use only. No Artificial Analysis number appears anywhere; pages link to its site.checked 2026-09-22
LiveBench
Scores are attributed to LiveBench and its release; reuse follows its datasheet policy.
LMArena
CC BY 4.0, attributed on every chart and table that shows a rating. Ratings: Arena (lmarena-ai) text leaderboard, overall, in its default view that adjusts ratings for answer length and formatting, published 2026-10-22, from the Hugging Face dataset lmarena-ai/leaderboard-dataset, licensed under CC BY 4.0. Modified by Local AI Wizard: selected rows only, model names matched to our catalog, ratings rounded. Not endorsed by Arena. Ratings: Arena (lmarena-ai) vision leaderboard, overall, in its default view that adjusts ratings for answer length and formatting, published 2026-10-19, from the Hugging Face dataset lmarena-ai/leaderboard-dataset, licensed under CC BY 4.0. Modified by Local AI Wizard: selected rows only, model names matched to our catalog, ratings rounded. Not endorsed by Arena. Ratings: Arena (lmarena-ai) text-to-image leaderboard, overall, published 2026-10-07, from the Hugging Face dataset lmarena-ai/leaderboard-dataset, licensed under CC BY 4.0. Modified by Local AI Wizard: selected rows only, model names matched to our catalog, ratings rounded. Not endorsed by Arena. Ratings: Arena (lmarena-ai) text-to-video leaderboard, overall, published 2026-09-22, from the Hugging Face dataset lmarena-ai/leaderboard-dataset, licensed under CC BY 4.0. Modified by Local AI Wizard: selected rows only, model names matched to our catalog, ratings rounded. Not endorsed by Arena.
Open ASR Leaderboard
Apache-2.0: the word error rates come from the results file in the leaderboard's GitHub repository, which carries the repository's Apache-2.0 licence. The Hub results dataset and the Space declare no licence and are not read. Word error rates: Open ASR Leaderboard (Hugging Face Audio team and contributors), English short-form results file scripts/data/en_shortform.csv of github.com/huggingface/open_asr_leaderboard, version 16743c6 of 2026-05-22 (results stated as of 2026-03-26), licensed under Apache-2.0. Modified by Local AI Wizard: selected rows only, model names matched to our catalog, values rounded. Not endorsed by Hugging Face.
Music Arena
MIT leaderboard files computed from a CC BY 4.0 vote dataset. Rating: Music Arena (CMU G-CLef) instrumental leaderboard of 2026-07-31, github.com/gclef-cmu/music-arena (MIT; Copyright (c) 2025 Carnegie Mellon University, Music Arena Team, Chris Donahue), computed from the Music Arena Dataset (CC BY 4.0). Modified by Local AI Wizard: selected rows only, model names matched to our catalog. Not endorsed by Music Arena.
MTEB
Embedding scores are computed from the public results dataset, which is CC0.
Hardware pictures
Only permitted generated images and reviewed manufacturer reference images, each with its caption. A missing or failed image becomes a labelled illustration.
Models and weights
Each model keeps its publisher's license, shown with a link to the license text. The site republishes facts and links, not weights.

Sources and dates

Collection date, benchmark release and page publication are different dates; each is labelled.

Every source, with collection date13
LiveBench
Quality scores for text models, with seven category means.release 2026-06-25 · collected 2026-10-18
LMArena
Text leaderboard ratings (overall, in the default "style control" view, which adjusts ratings for answer length and formatting) for the open models matched by hand review, with interval and votes; read from the public Hugging Face dataset, never from the site. A rating is blind pairwise human preference on the hosted model, not a test of a quantized local file and not task success; it has its own scale, shown in its own lens and column, and is never merged with, converted to or ranked against LiveBench. Some models below 27 B have a rating; a model with no Arena entry has none, never a sibling's.release 2026-10-22 · collected 2026-10-25
LMArena · vision, image, video
One leaderboard per category, each read from the same CC BY 4.0 Hugging Face dataset: vision (style control, overall) for the image-understanding models on Models › Vision, text-to-image (overall) for image generation, and text-to-video (overall) for video. Identity comes only from a reviewed map per category with one exact name per repository or catalog entry: base, quantized, distilled, preset and effort variants never take a rating. Inside a category every compared model comes from the same board and release; boards are never merged with each other, with the text board or with publisher metrics. Video matches only two of our models, so its rating is a fact on those two pages and is not ranked. Wan's VBench numbers are publisher-reported and are not ranked against other models either.vision 2026-10-19 · image 2026-10-07 · video 2026-09-22 · collected 2026-10-25
Audio quality
Audio has no single score: speech to text, text to speech and music are separate quality slots, each judged by its own test on its own scale, never merged. Speech to text: the Open ASR Leaderboard word error rate (English short-form, macro-average of eight test sets; lower is better). Music: the Music Arena instrumental rating, a fact on the one model it covers. Text to speech: no board publishes its results under a licence, so no score.Open ASR 2026-05-22 · Music Arena 2026-07-31
Hugging Face Hub
Model inventory: tasks, licenses, parameter counts, files, downloads.bundled copy 2026-09-26 · live status above
MTEB
Embedding scores (Multilingual v2 mean over tasks).retrieved 2026-10-25
GenEval · VBench
Published image and video results for the few models they cover. VBench comes from the Wan authors' own report: publisher-reported, shown on those model pages and never ranked against other models.observed 2026-10-20
Apple and NVIDIA product pages
Memory specifications of the hardware profiles, plus each profile's own sourced fields.memory checked 2026-09-30 · profiles reviewed 2026-09-30
llama.cpp
Community llama-bench scoreboards (CUDA, ROCm, Apple Silicon, Vulkan) and the repository’s own benches/ pages: measured speeds, copied as published.scoreboards checked 2026-10-25 · benches/ commit 2280f0f · checked 2026-10-25
mlx-lm
The maintainers’ benchmark table (mlx_lm/BENCHMARKS.md, MIT license): generation tokens per second on an Apple M4 Max for each MLX precision.commit 4cfc647 · checked 2026-10-25
Steam Hardware Survey
Only the order of graphics cards in the hardware catalog. Not sales and not local-AI use.August 2026
Publisher-reported results
Benchmark numbers each publisher prints on its own model card (the card's model-index and evaluation tables), copied as printed with their benchmark variant, unit, shots, thinking mode and harness when stated, and a link to the card. They are a different kind of evidence from the independent scores above: LiveBench, LMArena and MTEB run the same test for every model they cover, while a publisher chooses its own prompts, settings and comparison models. So publisher-reported values are shown only as facts on the model page, under "Reported by the publisher"; they never fill a comparable-score column, chart or ranking, and are never compared across publishers. Sizes of one family stand side by side only when the publisher reports them in the same card section with the same settings. A card with no numbers is recorded as not published; a gated card stays unread.cards read 2026-10-26
Model cards
The reviewed catalog links each checkpoint to its official card.catalog reviewed 2026-09-30

Downloads and fields

This curated table selects 260 checkpoints with primary model cards. IDs are stable within this catalog. The dynamic Wizard may show newer D1 data; these static files remain dated until editorial review and publication.

Of 60 selected language checkpoints, 36 fall at or below 32 GB and 37 at or below 64 GB under the same generic Q4 planning formula. 5 has no defensible simplified estimate. These are counts of this curated sample, not successful local runs or a census of all open models. Reproduce by filtering the JSON on task = language and reasoning, then counting non-null memoryEstimateGB at each threshold.

Fields and units
id
Stable catalog slug, not the publisher repository identifier.
parametersB / parameterBasis
Billions of stated model parameters or repository tensor elements, as qualified by the basis field; null when not verified. Auxiliary components may be separate.
score / scoreStatus
LiveBench overall 0–100 for language and MTEB (Multilingual, v2) mean(task) 0–100 for embeddings, never mixed: scoreLabel names the scale, vector size and who measured it. Artificial Analysis media scores are excluded because its free tier is internal use only; scoreStatus explains the excluded values and scoreSource links to the original scale.
memoryEstimateGB
Generic Q4 planning estimate in decimal GB for selected language checkpoints; not measured VRAM. Null means unknown or an inapplicable simplified formula.
repositoryStorageGB
Full published repository size where verified. It may include multiple formats and is not the minimum download.
Reviewed checkpoints260
ID / modelModalityQuality evidenceSource
Qwen3.5 9B
qwen3.5-9b
text + image + video—model card
Qwen3.8 27B
qwen3.8-27b
text + image + video75.3 · LiveBench 2026-06-25 overallmodel card
Qwen3.5 35B A3B
qwen3.5-35b-a3b
text + image + video—model card
Qwen3.6 27B
qwen3.6-27b
text + image + video64 · LiveBench 2026-06-25 overallmodel card
Smaug Mini
smaug-mini
text + image76.9 · LiveBench 2026-06-25 overallmodel card
DeepSeek V4 Flash · 0731
deepseek-v4-flash-0731
text74.2 · LiveBench 2026-06-25 overallmodel card
GLM-5.3 Flash
glm-5.3-flash
text + image71.6 · LiveBench 2026-06-25 overallmodel card
GLM-5.3
glm-5.3
text76.1 · LiveBench 2026-06-25 overallmodel card
Kimi K2.6 · thinking
kimi-k2.6-thinking
text + image70.5 · LiveBench 2026-06-25 overallmodel card
DeepSeek V4.1 Flash · max
deepseek-v4.1-flash-max
text + image81.1 · LiveBench 2026-06-25 overallmodel card
DeepSeek V4 Flash
deepseek-v4-flash
text65.5 · LiveBench 2026-06-25 overallmodel card
DeepSeek V4 Pro
deepseek-v4-pro
text71.6 · LiveBench 2026-06-25 overallmodel card
GLM-5.2
glm-5.2
text73.2 · LiveBench 2026-06-25 overallmodel card
Kimi K2.7 Code
kimi-k2.7-code
text + image68.4 · LiveBench 2026-06-25 overallmodel card
MiniMax M3
minimax-m3
text + image + video67.3 · LiveBench 2026-06-25 overallmodel card
Kimi K3
kimi-k3
text + image79.2 · LiveBench 2026-06-25 overallmodel card
Inkling · xhigh
inkling-xhigh
text + image + audio71.9 · LiveBench 2026-06-25 overallmodel card
Qwen3.8 Max
qwen3.8-max
text78.5 · LiveBench 2026-06-25 overallmodel card
Smaug Agentic
smaug-agentic
text + image79.5 · LiveBench 2026-06-25 overallmodel card
DeepSeek V4 Pro · 0813
deepseek-v4-pro-0813
text77.4 · LiveBench 2026-06-25 overallmodel card
DeepSeek V4 Flash Vision · exp
deepseek-v4-flash-vision-exp
text + image76.8 · LiveBench 2026-06-25 overallmodel card
Qwen3.8 Flash Next
qwen3.8-flash-next
text + image + video76.2 · LiveBench 2026-06-25 overallmodel card
Nemotron 3 Ultra 550B A55B
nemotron-3-ultra-550b-a55b
text67.4 · LiveBench 2026-06-25 overallmodel card
Smaug Flash
smaug-flash
text77.4 · LiveBench 2026-06-25 overallmodel card
Nemotron 3 Super 120B A12B
nemotron-3-super-120b-a12b
text—model card
MiMo V2.6 Pro
mimo-v2.6-pro
text + image + video + audio—model card
MiMo V2.6 Flash
mimo-v2.6-flash
text + image + video + audio—model card
Ling 3.0 Flash
ling-3.0-flash
text—model card
FLUX.2 [dev]
flux-2-dev
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Qwen-Image
qwen-image
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Wan2.1 T2V 1.3B
wan2.1-t2v-1.3b
video—model card
Stable Audio Open Small
stable-audio-open-small
audio—model card
FLUX.1 [dev]
flux-1-dev
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FLUX.1 [schnell]
flux-1-schnell
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HunyuanImage-2.1
hunyuanimage-2.1
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Ideogram 4.0
ideogram-4
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Qwen-Image-2512
qwen-image-2512
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Stable Diffusion 3.5 Large
stable-diffusion-3.5-large
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Wan2.1 T2V 14B
wan2.1-t2v-14b
video—model card
Wan2.2 T2V-A14B
wan2.2-t2v-a14b
video—model card
CogVideoX-5B
cogvideox-5b
video—model card
HunyuanVideo
hunyuanvideo
video—model card
LTX-Video 2B · 0.9.8 distilled
ltx-video-2b-0.9.8-distilled
video—model card
Whisper large-v3
whisper-large-v3
audio—model card
Whisper large-v3-turbo
whisper-large-v3-turbo
audio—model card
Stable Audio Open 1.0
stable-audio-open-1.0
audio—model card
MusicGen Small
musicgen-small
audio—model card
Gemma 4 E2B
gemma-4-e2b
text + image + audio + video—model card
Gemma 4 E4B
gemma-4-e4b
text + image + audio + video—model card
Gemma 4 12B
gemma-4-12b
text + image + audio + video—model card
Gemma 4 26B A4B
gemma-4-26b-a4b
text + image—model card
Gemma 4 31B
gemma-4-31b
text + image—model card
EmbeddingGemma 2
embeddinggemma-2
text + image + video + audio61.4 · MTEB (Multilingual, v2) mean(task) · 768d · publisher-reportedmodel card
EmbeddingGemma (300M)
embeddinggemma-300m
text61.1 · MTEB (Multilingual, v2) mean(task) · 768d · independent runmodel card
Qwen3-Embedding 0.6B
qwen3-embedding-0.6b
text64.3 · MTEB (Multilingual, v2) mean(task) · 1024d · independent runmodel card
Qwen3-Embedding 4B
qwen3-embedding-4b
text69.5 · MTEB (Multilingual, v2) mean(task) · 2560d · independent runmodel card
Qwen3-Embedding 8B
qwen3-embedding-8b
text70.6 · MTEB (Multilingual, v2) mean(task) · 4096d · independent runmodel card
BGE-M3
bge-m3
text59.6 · MTEB (Multilingual, v2) mean(task) · 1024d · independent runmodel card
Nomic Embed Text v2 (MoE)
nomic-embed-text-v2-moe
text57.6 · MTEB (Multilingual, v2) mean(task) · 768d · independent runmodel card
Granite Embedding 278M Multilingual
granite-embedding-278m-multilingual
text53.7 · MTEB (Multilingual, v2) mean(task) · 768d · independent runmodel card
Granite Embedding 311M Multilingual R2
granite-embedding-311m-multilingual-r2
text56 · MTEB (Multilingual, v2) mean(task) · 768d · independent runmodel card
Multilingual E5 Large Instruct
multilingual-e5-large-instruct
text63.2 · MTEB (Multilingual, v2) mean(task) · 1024d · independent runmodel card
Multilingual E5 Base
multilingual-e5-base
text57 · MTEB (Multilingual, v2) mean(task) · 768d · independent runmodel card
Snowflake Arctic Embed L v2.0
snowflake-arctic-embed-l-v2.0
text57 · MTEB (Multilingual, v2) mean(task) · 1024d · independent runmodel card
BGE Small EN v1.5
bge-small-en-v1.5
text43.8 · MTEB (Multilingual, v2) mean(task) · 384d · independent runmodel card
BGE Base EN v1.5
bge-base-en-v1.5
text44.5 · MTEB (Multilingual, v2) mean(task) · 768d · independent runmodel card
BGE Large EN v1.5
bge-large-en-v1.5
text45.1 · MTEB (Multilingual, v2) mean(task) · 1024d · independent runmodel card
Granite Embedding Small English R2
granite-embedding-small-english-r2
textNo MTEB result with a source for this model.model card
Llama Nemotron Embed 1B v2
llama-nemotron-embed-1b-v2
textNo MTEB result with a source for this model.model card
BGE Multilingual Gemma2
bge-multilingual-gemma2
textNo MTEB result with a source for this model.model card
Harrier OSS v1 (0.6B)
harrier-oss-v1-0.6b
text69 · MTEB (Multilingual, v2) mean(task) · 1024d · independent runmodel card
Nemotron 3 Embed 1B
nemotron-3-embed-1b-bf16
textNo MTEB result with a source for this model.model card
Qwen3 1.7B
qwen3-1.7b
text—model card
Qwen3 4B Instruct 2507
qwen3-4b-instruct-2507
text—model card
Qwen3 8B
qwen3-8b
text—model card
Qwen3 14B
qwen3-14b
text—model card
Qwen3 30B A3B Instruct 2507
qwen3-30b-a3b-instruct-2507
text—model card
Qwen3 Coder 30B A3B
qwen3-coder-30b-a3b
text—model card
gpt-oss-20b
gpt-oss-20b
text—model card
Llama 3.1 8B Instruct
llama-3.1-8b-instruct
text—model card
Llama 3.2 3B Instruct
llama-3.2-3b-instruct
text—model card
Llama 3.2 1B Instruct
llama-3.2-1b-instruct
text—model card
Mistral Small 3.2 24B
mistral-small-3.2-24b
text + image—model card
Ministral 3 3B
ministral-3-3b
text + image—model card
Ministral 3 8B
ministral-3-8b
text + image—model card
Ministral 3 14B
ministral-3-14b
text + image—model card
Devstral Small 2 24B
devstral-small-2-24b
text + image—model card
Granite 4.2 3B
granite-4.2-3b
text—model card
Granite 4.2 8B
granite-4.2-8b
text—model card
Granite 4.2 30B
granite-4.2-30b
text—model card
Olmo 3 7B Instruct
olmo-3-7b-instruct
text—model card
Olmo 3 32B Think
olmo-3-32b-think
text—model card
Phi-4 mini
phi-4-mini
text—model card
Phi-4
phi-4
text—model card
SmolLM3 3B
smollm3-3b
text—model card
Nemotron 3 Nano 30B A3B
nemotron-3-nano-30b-a3b
text—model card
Nemotron 3.5 Lightning 30B A3B
nemotron-3.5-lightning-30b-a3b
text—model card
GLM-4.7 Flash
glm-4.7-flash
text—model card
DeepSeek R1 0528 Qwen3 8B
deepseek-r1-0528-qwen3-8b
text—model card
Stable Diffusion XL 1.0
stable-diffusion-xl-1
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Stable Diffusion 1.5
stable-diffusion-1-5
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Z-Image Turbo
z-image-turbo
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FLUX.2 [klein] Base 4B
flux-2-klein-base-4b
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FLUX.2 [klein] 4B
flux-2-klein-4b
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FLUX.2 [klein] 9B
flux-2-klein-9b
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FLUX.2 [klein] Base 9B
flux-2-klein-base-9b
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Stable Diffusion 3.5 Medium
stable-diffusion-3-5-medium
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HiDream-I1-Fast
hidream-i1-fast
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HunyuanImage 3.0 Instruct (Fal)
hunyuanimage-3-instruct
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FLUX.1 Krea [dev]
flux-1-krea-dev
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
LongCat Image
longcat-image
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Cosmos3-Super-Text2Image
cosmos3-super
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Stable Diffusion 3.5 Large Turbo
stable-diffusion-3-5-large-turbo
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HiDream-O1-Image
hidream-o1-image
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
ERNIE Image Turbo
ernie-image-turbo
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Bagel
bagel
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
GLM-Image
glm-image
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
OmniGen V2
omnigen-v2
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Stable Diffusion 3 Medium
stable-diffusion-3-medium
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
ERNIE Image
ernie-image
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HunyuanImage 3.0 (Fal)
hunyuanimage-3
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Lumina Image v2
lumina-image-v2
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HiDream-O1-Image-Dev
hidream-o1-image-dev
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
FIBO
fibo
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
SRPO
srpo
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
HiDream-I1-Dev
hidream-i1-dev
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Infinity 8B
infinity-8b
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Sana Sprint 1.6B
sana-sprint-1-6b
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Cosmos3-Super-Text2Image-4Step
cosmos3-super-4step
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Janus Pro
janus-pro
imageArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
MiniMax H3
minimax-h3
videoArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
LTX-2.5
ltx-2-5-pro
videoArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
LTX-2.3
ltx-2-3-pro
videoArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
LTX-2
ltx-2-pro
videoArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Kokoro 82M v1.0
kokoro-82m
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
VibeVoice 1.5B
vibevoice-1-5b
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Higgs Audio V3 TTS
higgs-audio-v3-tts
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Fish Audio S2 Pro
fish-audio-s2-pro
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Breeze TTS 2
breeze-tts-2
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Magpie-Multilingual 357M
magpie-multilingual-357m
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Maya1
maya1
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
OpenAudio S1 Mini
openaudio-s1-mini
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Voxtral TTS
voxtral-tts
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Chatterbox
chatterbox
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
Zonos-v0.1
zonos-v0-1
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
MetaVoice v1
metavoice-v1
audioArtificial Analysis score, rank and sample count excluded: its free tier is internal use only. Source checked 2026-09-22.model card
vit-base-patch16-224
vit-base-patch16-224
image + text—model card
Qwen3-VL-8B-Instruct
qwen3-vl-8b-instruct
image + text—model card
Qwen2.5-VL-7B-Instruct
qwen2.5-vl-7b-instruct
image + text—model card
Qwen3.6-35B-A3B
qwen3.6-35b-a3b
image + text—model card
Qwen3-VL-4B-Instruct
qwen3-vl-4b-instruct
image + text—model card
Qwen3-VL-2B-Instruct
qwen3-vl-2b-instruct
image + text—model card
DeepSeek-OCR
deepseek-ocr
image + text—model card
siglip2-giant-opt-patch16-384
siglip2-giant-opt-patch16-384
image + text—model card
GLM-OCR
glm-ocr
image + text—model card
owlv2-base-patch16-ensemble
owlv2-base-patch16-ensemble
image + text—model card
siglip2-base-patch16-224
siglip2-base-patch16-224
image + text—model card
chandra-ocr-2
chandra-ocr-2
image + text—model card
Unlimited-OCR
unlimited-ocr
image + text—model card
SmolVLM2-500M-Video-Instruct
smolvlm2-500m-video-instruct
image + text—model card
siglip-base-patch16-224
siglip-base-patch16-224
image + text—model card
siglip2-base-patch16-naflex
siglip2-base-patch16-naflex
image + text—model card
llava-1.5-7b-hf
llava-1.5-7b-hf
image + text—model card
siglip-so400m-patch14-384
siglip-so400m-patch14-384
image + text—model card
siglip2-so400m-patch14-384
siglip2-so400m-patch14-384
image + text—model card
Qwen2.5-VL-32B-Instruct
qwen2.5-vl-32b-instruct
image + text—model card
dots.ocr
dots.ocr
image + text—model card
Qwen2-VL-7B-Instruct
qwen2-vl-7b-instruct
image + text—model card
Phi-3.5-vision-instruct
phi-3.5-vision-instruct
image + text—model card
siglip2-base-patch16-256
siglip2-base-patch16-256
image + text—model card
LightOnOCR-2-1B
lightonocr-2-1b
image + text—model card
CLIP-ViT-H-14-laion2B-s32B-b79K
clip-vit-h-14-laion2b-s32b-b79k
image + text—model card
InternVL2-2B
internvl2-2b
image + text—model card
GOT-OCR2_0
got-ocr2-0
image + text—model card
gemma-4-31B
google-gemma-4-31b
image + text—model card
chandra
chandra
image + text—model card
Qwen3-VL-235B-A22B-Instruct
qwen3-vl-235b-a22b-instruct
image + text—model card
granite-docling-258M
granite-docling-258m
image + text—model card
SmolVLM-256M-Instruct
smolvlm-256m-instruct
image + text—model card
Qwen3-VL-30B-A3B-Instruct
qwen3-vl-30b-a3b-instruct
image + text—model card
MinerU2.5-Pro-2605-1.2B
mineru2.5-pro-2605-1.2b
image + text—model card
dots.mocr
dots.mocr
image + text—model card
granite-vision-3.3-2b
granite-vision-3.3-2b
image + text—model card
UI-TARS-1.5-7B
ui-tars-1.5-7b
image + text—model card
MiniCPM-V-4.6
minicpm-v-4.6
image + text—model card
granite-vision-4.1-4b
granite-vision-4.1-4b
image + text—model card
Llama-4-Scout-17B-16E-Instruct
llama-4-scout-17b-16e-instruct
image + text—model card
Phi-4-multimodal-instruct
phi-4-multimodal-instruct
image + text—model card
SmolVLM2-2.2B-Instruct
smolvlm2-2.2b-instruct
image + text—model card
SmolVLM-500M-Instruct
smolvlm-500m-instruct
image + text—model card
Cosmos-Reason2-8B
cosmos-reason2-8b
image + text—model card
GLM-4.6V-Flash
glm-4.6v-flash
image + text—model card
InternVL3-8B
internvl3-8b
image + text—model card
MiniCPM-V-4_5
minicpm-v-4-5
image + text—model card
olmOCR-7B-0225-preview
olmocr-7b-0225-preview
image + text—model card
olmOCR-2-7B-1025
olmocr-2-7b-1025
image + text—model card
Llama-3.1-Nemotron-Nano-VL-8B-V1
llama-3.1-nemotron-nano-vl-8b-v1
image + text—model card
gemma-4-12B
google-gemma-4-12b
image + text—model card
Idefics3-8B-Llama3
idefics3-8b-llama3
image + text—model card
granite-4.0-3b-vision
granite-4.0-3b-vision
image + text—model card
llava-onevision-qwen2-7b-ov-hf
llava-onevision-qwen2-7b-ov-hf
image + text—model card
Llama-3.2-11B-Vision-Instruct
llama-3.2-11b-vision-instruct
image + text—model card
Qianfan-OCR
qianfan-ocr
image + text—model card
InternVL3_5-8B
internvl3-5-8b
image + text—model card
SmolVLM2-256M-Video-Instruct
smolvlm2-256m-video-instruct
image + text—model card
Emu3-Chat-hf
emu3-chat-hf
image + text—model card
GLM-4.5V
glm-4.5v
image + text—model card
SmolDocling-256M-preview
smoldocling-256m-preview
image + text—model card
Holo2-4B
holo2-4b
image + text—model card
SmolVLM-Instruct
smolvlm-instruct
image + text—model card
Molmo-7B-D-0924
molmo-7b-d-0924
image + text—model card
NVIDIA-Nemotron-Nano-12B-v2-VL-BF16
nvidia-nemotron-nano-12b-v2-vl-bf16
image + text—model card
Step3-VL-10B
step3-vl-10b
image + text—model card
Keye-VL-1_5-8B
keye-vl-1-5-8b
image + text—model card
Llama-3.2-11B-Vision
llama-3.2-11b-vision
image + text—model card
kosmos-2.5
kosmos-2.5
image + text—model card
MinerU2.5-2509-1.2B
mineru2.5-2509-1.2b
image + text—model card
NVLM-D-72B
nvlm-d-72b
image + text—model card
Molmo-72B-0924
molmo-72b-0924
image + text—model card
Llama-4-Maverick-17B-128E-Instruct
llama-4-maverick-17b-128e-instruct
image + text—model card
Mage-VL
mage-vl
image + text—model card
deepseek-vl2
deepseek-vl2
image + text—model card
LLaVA-OneVision-1.5-8B-Instruct
llava-onevision-1.5-8b-instruct
image + text—model card
Falcon-OCR
falcon-ocr
image + text—model card
Fara1.5-9B
fara1.5-9b
image + text—model card
Fara-7B
fara-7b
image + text—model card
Phi-4-reasoning-vision-15B
phi-4-reasoning-vision-15b
image + text—model card
Llama-4-Scout-17B-16E
llama-4-scout-17b-16e
image + text—model card
Fara1.5-27B
fara1.5-27b
image + text—model card
Llama-Guard-3-11B-Vision
llama-guard-3-11b-vision
image + text—model card
Youtu-VL-4B-Instruct
youtu-vl-4b-instruct
image + text—model card
Fara1.5-4B
fara1.5-4b
image + text—model card
SmolVLM2-2.2B-Base
smolvlm2-2.2b-base
image + text—model card
MolmoMotion-4B-H1-F32
molmomotion-4b-h1-f32
image + text—model card
Youtu-Parsing
youtu-parsing
image + text—model card
tipsv1-g14-lowres
tipsv1-g14-lowres
image + text—model card
nemotron-labs-audio-visual-flamingo-hf
nemotron-labs-audio-visual-flamingo-hf
image + text—model card
Llama-3.2-90B-Vision-Instruct
llama-3.2-90b-vision-instruct
image + text—model card
Keye-VL-2.0-30B-A3B
keye-vl-2.0-30b-a3b
image + text—model card
tipsv1-l14
tipsv1-l14
image + text—model card
tipsv1-g14
tipsv1-g14
image + text—model card
tipsv1-so400m14
tipsv1-so400m14
image + text—model card
Aquila-VL-2B-llava-qwen
aquila-vl-2b-llava-qwen
image + text—model card
POINTS-Reader
points-reader
image + text—model card
Llama-4-Maverick-17B-128E
llama-4-maverick-17b-128e
image + text—model card
Perception-LM-8B
perception-lm-8b
image + text—model card
cogvlm2-llama3-caption
cogvlm2-llama3-caption
image + text—model card
Video-XL-2
video-xl-2
image + text—model card
tipsv1-s14
tipsv1-s14
image + text—model card
Bunny-v1_1-Llama-3-8B-V
bunny-v1-1-llama-3-8b-v
image + text—model card
tipsv1-b14
tipsv1-b14
image + text—model card
Recon2Reason-Reasoning-4B
recon2reason-reasoning-4b
image + text—model card
Llama-3.2-90B-Vision
llama-3.2-90b-vision
image + text—model card
LFM2-VL-1.6B
lfm2-vl-1.6b
image + text—model card
LFM2.5-VL-1.6B
lfm2.5-vl-1.6b
image + text—model card
ERNIE-4.5-VL-28B-A3B-PT
ernie-4.5-vl-28b-a3b-pt
image + text—model card
diffusiongemma-26B-A4B-it
diffusiongemma-26b-a4b-it
image + text—model card
North-Micro-Vision-Instruct
north-micro-vision-instruct
image + text—model card
Apriel-1.6-15b-Thinker
apriel-1.6-15b-thinker
image + text—model card
surya-ocr-2
surya-ocr-2
image + text—model card

Directory

Every reviewed page, each with its source, date and limits.

Models 260

Reviewed model profiles

All discovered models

All 260 reviewed models

Hardware 120

Hardware configurations

All hardware

All 120 configurations

Comparisons 32

Task and hardware comparisons

Open Compare

All 32 comparisons

Catalog profiles 38

Hub checkpoints with publisher evidence and a hardware answer

All discovered models

All 38 catalog profiles