Highest catalogue scores that fit
- LFM2.5-8B-A1B8.1/10 · min 8 GB
- Granite 4.2 (8B)8.0/10 · min 8 GB
- Qwen 3 (8B)8.0/10 · min 8 GB
LocalClawThe Local AI Compatibility Guide
Find the models, software and hardware that fit your machine.
Labs · Free browser AI playground
Try browser AI →Read about local AI. Then try it.
Five small model presets. Chat, play detective, remix text or shuffle a decision situation. Runs on your device, with no account or API key.
Your local AI workspace
My Machines →What can your machine run?
Create a free account, add your Mac, PC or NVIDIA workstation once, and LocalClaw keeps your compatible models and new releases ready.
Representative crawlable snapshot
Local LLM score leaders
Top 24 of 256 current local records, ordered only by the documented LocalClaw catalogue score. Community ratings remain a separate live signal; open the full catalogue for every record.
| Rank | Model / family | LocalClaw | Parameters | Min RAM | Licence | Released |
|---|---|---|---|---|---|---|
| 01 | DeepSeek V4 Flash Vision Expdeepseek-flash | 9.3/10 | 284B (13B active, multimodal MoE) | 128 GB | MIT | Aug 31, 2026 |
| 02 | DeepSeek V4 Flash 0731 (284B MoE)deepseek-flash | 9.3/10 | 284B (13B active) | 256 GB | MIT | Jul 31, 2026 |
| 03 | MiMo-V2.6-Flash-RLmimo | 9.2/10 | 309B (15B active, MoE) | 160 GB | MIT | Sep 21, 2026 |
| 04 | DeepSeek V4.1 Flashdeepseek-flash | 9.2/10 | 552B MoE (8B/16B active) | 256 GB | MIT | Sep 10, 2026 |
| 05 | GLM-5.3-Flashglm | 9.2/10 | 320B (18B active, MoE) | 160 GB | MIT | Aug 25, 2026 |
| 06 | Kimi K2.5 (1T MoE)kimi | 9.2/10 | 1T (32B active, MoE) | 1024 GB | Modified MIT | Jan 2026 |
| 07 | Qwen 3 (32B)qwen | 9.2/10 | 32B | 32 GB | Apache 2.0 | Apr 2025 |
| 08 | Qwen3-Coder-Nextqwen-coder | 9.1/10 | 80B (3B active, MoE) | 64 GB | Apache 2.0 | Feb 2026 |
| 09 | MiniMax M3 (428B/23B active)minimax | 9.0/10 | 428B (23B active) | 2048 GB | MiniMax Community License | Jun 2026 |
| 10 | Trinity Large Preview (398B MoE)trinity | 9.0/10 | 398B total (13B active, MoE) | 320 GB | OpenMDW 1.1 | Feb 2026 |
| 11 | DeepSeek V3.2 (37B/685B MoE)deepseek-v3 | 9.0/10 | 685B total (37B active, MoE) | 448 GB | MIT | Dec 2025 |
| 12 | Kimi K2 Instruct (1T MoE)kimi | 9.0/10 | 1T (32B active, 384 experts) | 1024 GB | See upstream model page | Jul 2025 |
| 13 | Qwen 3 MoE (235B/22B active)qwen-moe | 9.0/10 | 235B (22B active) | 192 GB | Apache 2.0 | Apr 2025 |
| 14 | Bonsai 2 27Bbonsai | 9.0/10 | 27.36B (ternary) | 16 GB | Apache 2.0 | Sep 17, 2026 |
| 15 | Xing4.0-29B-A4Bxing | 9.0/10 | 29B (4B active, MoE) | 32 GB | Apache 2.0 | Sep 17, 2026 |
| 16 | Qwen 3.6 35B-A3Bqwen | 9.0/10 | 35B (3B active, MoE) | 32 GB | Apache 2.0 | Apr 2026 |
| 17 | Qwen3.8-27Bqwen | 8.9/10 | 27B | 32 GB | Apache 2.0 | Aug 5, 2026 |
| 18 | Qwen 3.5 MoE (122B/10B active)qwen | 8.9/10 | 122B (10B active) | 80 GB | Apache 2.0 | Aug 2025 |
| 19 | GLM-5.2 (744B MoE)glm | 8.9/10 | 744B (40B active) | 256 GB | MIT | Jun 2026 |
| 20 | Kimi K2.7 Code (1T MoE)kimi | 8.9/10 | 1T (32B active) | 1024 GB | Modified MIT | Jun 2026 |
| 21 | DeepSeek V4 Pro (1.6T MoE)deepseek | 8.9/10 | 1.6T (49B active) | 1024 GB | See upstream model page | May 2026 |
| 22 | GLM-5.1glm | 8.9/10 | 754B MoE | 640 GB | See upstream model page | May 2026 |
| 23 | Kimi K2 Thinking (1T MoE)kimi | 8.9/10 | 1T (32B active, 384 experts) | 1024 GB | See upstream model page | Nov 2025 |
| 24 | DeepSeek V3.2 Exp (671B MoE)deepseek | 8.9/10 | 671B (37B active) | 512 GB | See upstream model page | Sep 2025 |
Representative crawlable snapshot
Local speech score leaders
Top 12 of 120 records, ordered by the homepage Audio score: 68% quality and 32% speed. Community stars remain independent; open the full catalogue for every record.
- 01Parakeet TDT 0.6B v2ASR · Audio 9.6/10 · CC-BY-4.0
- 02Parakeet TDT 0.6B v3ASR · Audio 9.6/10 · CC-BY-4.0
- 03Granite Speech 5.0 470M TurboCTCASR · Audio 9.5/10 · Apache 2.0
- 04Kyutai STT 2.6B EnglishASR · Audio 9.4/10 · CC-BY-4.0
- 05F5-TTS v1 BaseTTS · Audio 9.4/10 · CC-BY-NC-4.0 weights; MIT code
- 06Kokoro TTSTTS · Audio 9.4/10 · Apache 2.0
- 07Voxtral 4B TTS 2603TTS · Audio 9.4/10 · CC-BY-NC 4.0
- 08Nemotron 3.5 ASR Streaming 0.6BASR · Audio 9.4/10 · OpenMDW 1.1
- 09Qwen3-ASRASR · Audio 9.3/10 · Apache 2.0
- 10X2Streaming-ASR 4B 1009ASR · Audio 9.3/10 · Apache 2.0
- 11Phonon-2ASR · Audio 9.3/10 · CC-BY-4.0 weights; Apache 2.0 code
- 12F5-TTSTTS · Audio 9.3/10 · CC-BY-NC-4.0 weights; MIT code
Representative crawlable snapshot
Image, video, 3D, music and vision
Up to 4 verified records per family are shown here. JavaScript adds the complete interactive index with search plus system, compute, RAM and VRAM filters; each dedicated catalogue contains every record.
4 of 5 verified records
Browse the full image catalogue →Local Image
FLUX.2 Klein 4B16 GB RAM · 8 GB VRAMFast open image generation and multi-reference editing for consumer NVIDIA GPUs.FLUX.1 Schnell32 GB RAM · 16 GB VRAMFour-step open text-to-image model with Diffusers, ComfyUI and local inference support.Stable Diffusion 3.5 Medium24 GB RAM · 12 GB VRAMGeneral text-to-image foundation model with official local inference and broad tooling support.Qwen Image 251232 GB RAM · 24 GB VRAMOpen image generation model focused on complex text rendering, realism and precise editing.
4 of 67 verified records
Browse the full video catalogue →Local Video
NeoVerse128 GB RAM · 40 GB VRAMApache-licensed 4D world model for turning a monocular video or image into novel-trajectory MP4 video.SeedVR232 GB RAM · 8 GB VRAMApache-licensed one-step diffusion model for local video restoration, super-resolution and upscaling.Vimoo 1.0 Lite32 GB RAM · 16 GB VRAMApache-licensed local text-to-video and image-to-video family with Diffusers, ComfyUI and Apple Silicon MLX paths.RAVEN64 GB RAM · 24 GB VRAMNon-commercial causal autoregressive video model for streaming local text-to-video extrapolation on Wan 2.1 or MiniMax H3.
4 of 74 verified records
Browse the full 3d catalogue →Local 3D
SprongMotion-14M4 GB RAMCompact Apache-2.0 text-to-motion model that generates 3D human skeleton animation and BVH files locally on a CPU.BuildRome64 GB RAM · 24 GB VRAMNon-commercial single-image scene reconstructor that builds metric 3D scene meshes and exports GLB assets.RecGen64 GB RAM · 24 GB VRAMNon-commercial RGB-D object reconstruction model that exports textured meshes, GLB assets and Gaussian splats.Depth Anything 3 Base16 GB RAM · 8 GB VRAMCompact any-view visual geometry model for depth, camera pose and 3D reconstruction exports.
4 of 5 verified records
Browse the full music catalogue →Local Music
ACE-Step 1.516 GB RAM · 8 GB VRAMLocal music generation for songs, remixes, repainting and lyric-conditioned workflows.YuE32 GB RAM · 8 GB VRAMOpen lyrics-to-song foundation model for multi-minute songs with vocals and accompaniment.MusicGen Small16 GB RAM · 8 GB VRAMCompact text-to-music model from AudioCraft for short instrumental generations.MusicGen Medium24 GB RAM · 16 GB VRAMHigher-quality text-to-music and melody-conditioned generation through AudioCraft.
4 of 5 verified records
Browse the full vision catalogue →Local Vision
PaddleOCR-VL 1.616 GB RAM · 8 GB VRAMLocal document parsing, OCR, layout recognition and structured extraction across many hardware backends.Florence-2 Base8 GB RAM · 4 GB VRAMCompact general vision model for captioning, OCR, detection, segmentation and grounding.Moondream 28 GB RAM · 4 GB VRAMSmall vision-language model designed for local image understanding and edge deployment.Qwen2.5-VL 3B16 GB RAM · 8 GB VRAMCompact multimodal model for images, documents, charts and video understanding.
Progressive enhancement: JavaScript adds the complete interactive index with live community vote counts, machine-fit filtering, search, sorting and comparison across 532 local records. This verified sample remains useful if those enhancements are unavailable.
// RECENTLY VERIFIED
Fresh local AI worth checking
New open-weight releases are screened for a real local install path before they enter LocalClaw.
Nov 4, 2026 · Workstation coding
Ripple 65B Preview
40.43GB AD-Q4_K_M GGUF; 48GB memory recommended for moderate context. Requires Ripple-enabled llama.cpp.
Oct 27, 2026 · Local text agents
LFM3 1.7B
Official 1.10GB Q4_K_M GGUF. Plan for 8GB RAM at short context; requires LFM3-enabled llama.cpp.
Oct 27, 2026 · Local image understanding
LFM3 VL 2B
Official GGUF plus vision projector totals 1.43GB. Requires LFM3-enabled llama.cpp; no video input yet.
Verified index guide
Choose a local model with transparent signals
Catalogue snapshot updated . Counts, score formulas and links below are generated from the same repository data as the interactive directory.
Free tool · Transparent calculation
Calculate LLM memory →A model file is only part of the memory bill.
Estimate weights, context cache and system headroom. Compare workloads and share every assumption.
Quick answers by installed RAM
These are catalogue-score leaders whose minimum-RAM field fits each tier, not universal benchmark winners.
Highest catalogue scores that fit
- Bonsai 2 27B9.0/10 · min 16 GB
- Qwen 3 (14B)8.6/10 · min 16 GB
- Mellum2.1 12B-A2.5B Thinking8.5/10 · min 16 GB
Highest catalogue scores that fit
- Qwen 3 (32B)9.2/10 · min 32 GB
- Bonsai 2 27B9.0/10 · min 16 GB
- Xing4.0-29B-A4B9.0/10 · min 32 GB
Highest catalogue scores that fit
- Qwen 3 (32B)9.2/10 · min 32 GB
- Qwen3-Coder-Next9.1/10 · min 64 GB
- Bonsai 2 27B9.0/10 · min 16 GB
Four signals, never blended
Each signal answers a different question. Community votes, editorial ratings, machine fit and external benchmarks stay separate.
Community ★ /5
Raw signed-in member average plus vote count. The confidence sort uses a transparent 3.5/5 prior over five votes; fewer than five votes is marked EARLY.
Live public aggregates →LocalClaw /10
Editorial catalogue rubric: 38% quality, 24% coding, 24% reasoning and 14% speed, using only maintained repository fields. It is not a standardized third-party benchmark.
Hardware fit
Compares installed RAM with the catalogue minimum. Context, runtime overhead, GPU offload and other applications can require more headroom.
Independent benchmarks
Artificial Analysis adds third-party performance context for covered models. It never replaces LocalClaw's local-runtime, licence, RAM or VRAM checks.
Choose the right benchmark ↓Independent benchmark reality check
Use the external signal that matches the output, then come back to LocalClaw for local compatibility. Coverage is not universal, so absence is never scored as failure.
LLMOpen-weight intelligenceCompare intelligence, openness, model size and context. Local RAM fit remains a LocalClaw check.Open independent LLM data ↗
ImageVisual preference arenaCompare generated-image preference. Installation, quantization and VRAM are verified here.Open the image arena ↗
VoiceSpeech preference leaderboardUse the Open weights filter for local candidates. Voice quality does not prove a local runtime exists.Open the speech leaderboard ↗
VideoVideo preference arenaCompare text-to-video and image-to-video output. LocalClaw supplies the practical RAM and VRAM path.Open the video arena ↗
Source boundary: LocalClaw links to the official Artificial Analysis surfaces and does not scrape, copy or blend their rankings into the LocalClaw score. Live score ingestion can be added later through their official API with server-side caching and attribution.
Local model index FAQ
Short, source-aligned answers for people and answer engines.
What is the Local AI Compatibility Guide?
It is LocalClaw's hardware-aware guide combining 532 local AI records with computer, RAM/GPU and software guidance. The catalogue includes 256 indexable LLM pages, 120 speech records and 156 image, video, 3D, music and vision models.
How is the LocalClaw score calculated?
The homepage LocalClaw catalogue score is a separate 0-10 method calculated from repository fields: 38% quality, 24% coding, 24% reasoning and 14% speed. It is not a community rating or an external benchmark leaderboard.
Are community stars mixed into the LocalClaw score?
No. Community stars are the independent 1-5 average from signed-in LocalClaw members, always shown with the number of votes. Community confidence uses only stars and vote count for ordering; EARLY means fewer than five votes. It never changes the LocalClaw score.
Does an Artificial Analysis result prove that a model runs locally?
No. Artificial Analysis provides independent performance context for covered models. LocalClaw separately verifies the downloadable model, local runtime path, RAM or VRAM floor and licence. A model missing from an external leaderboard is treated as not covered, not as a poor performer.
How should I choose a local model for 8, 16, 32 or 64 GB RAM?
Use the machine RAM selector or a RAM guide. A model is considered compatible when its catalogue minimum-RAM value is at or below the selected memory; context length, runtime overhead and other applications can require additional headroom.
Does the homepage include cloud-only or API-only models?
No hosted-only or API-only records appear in the ranked homepage lists. The LLM list contains 215 records with a public GGUF repository, plus 44 public and 4 gated model-card records without a verified GGUF file; those pages do not claim a one-click desktop install. The 6 exact-repository-unavailable LLM routes, online Edge TTS, API-only OCTAVE 2 and unverified XTTS v3 route are excluded.
What speech models are covered?
The local speech index contains 120 records: 82 TTS models, 31 ASR models and 5 local speech application. Each entry keeps its Audio score separate from community stars.
How fresh is the LocalClaw catalogue?
This homepage snapshot was generated from the repository catalogue updated November 6, 2026. Release dates, RAM requirements, licences and links come from the maintained LocalClaw records and linked model pages.