Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

247 — Top 50 Local AI Models for Privacy + Best Outputs (2026)

Status: research-complete Date: April 2, 2026 Topic: agents Type: infrastructure-research Last-validated: 2026-05-21 Original-query: Top 50 local AI models 2026 for ZAO - music generation (ACE-Step), TTS (Kokoro, Voxtral), embeddings (BGE-M3, Qwen3), coding (OmniCoder), image (FLUX.1), video (Wan 2.2) Tier: reference Source: @meta_alchemist on X — April 1, 2026 thread

Key Decisions / Recommendations

Decision Recommendation
Music generation UPGRADE to ACE-Step 1.5 XL (4B DiT) from current v1-3.5B — <2s/song on A100, <10s on RTX 3090, <4GB VRAM, LoRA training from a few songs, 50+ language lyrics, Apache 2.0. Already integrated at src/app/api/music/generate/route.ts via HuggingFace Inference API — update model ID from ACE-Step/ACE-Step-v1-3.5B to ACE-Step/ACE-Step-v1.5-XL
TTS for accessibility USE Kokoro-82M — 82M params, CPU-only, Apache 2.0, 50 voices, 6 languages, 2.2M+ downloads. Deploy via Kokoro-FastAPI container. Solves: read-aloud for proposals, audio versions of governance posts, podcast narration
TTS for artist voice cloning USE Voxtral TTS 4B (Mistral) — 3-second voice cloning, preferred over ElevenLabs by 62.8% in blind tests, 9 European languages, free self-hosted. For artist intros, personalized notifications
Embeddings for future RAG USE BGE-M3 (568M) as default — ~1.2GB VRAM, 1000+ languages, dense+sparse+multi-vector in one model, Apache 2.0. Scale to Qwen3-Embedding-8B (MTEB #1, 32K context, Matryoshka dimensions) when research library exceeds 500 docs. ZAO currently uses Supabase textSearch (full-text), not vector search
Local coding model USE OmniCoder-9B for lightweight local dev — 5.5GB VRAM (Q4), 262K context, trained on Claude Opus 4.6 traces, strong agentic behavior, Apache 2.0. Or Qwen3-Coder-Next 80B for heavy multi-file editing
Image generation USE FLUX.1 Schnell for community art — Apache 2.0 (commercial OK), 4-step generation, 12GB VRAM. For edits/brand consistency: FLUX.1 Kontext Dev (reference image + text prompt)
Video generation USE Wan 2.2 A14B (Apache 2.0) for local video — supplements Doc 209's cloud recommendations (Seedance, Neural Frames). 6GB VRAM minimum with 1.3B variant. VBench 84.7%+
Speech-to-text USE Whisper Large-v3 — 99 languages, Apache 2.0, massive tooling ecosystem. Already industry standard. For live transcription in Spaces, pair with existing Deepgram integration (Doc 216)
Edge/mobile model SKIP for now — ZAO is a web app (PWA), not a native mobile app. Revisit Qwen3.5-2B/4B and Phi-4-mini if ZAO ships a Capacitor native app
3D generation SKIP — not relevant to ZAO's music community use case
Lip sync SKIP for now — Doc 209 covers video tooling. Revisit MuseTalk/Wav2Lip when artist video content pipeline is built

Comparison of Options — ZAO-Relevant Models

Music Generation (Most Critical for ZAO)

Model Params VRAM License Duration Key Feature ZAO Fit
ACE-Step 1.5 XL 4B DiT <4GB Apache 2.0 10s–10min LoRA from few songs, 50+ lang lyrics, batch 8 songs BEST — already using v1 via HF API
HeartMuLa-oss-3B 3B ~2GB Open Variable Best lyrics controllability (HeartBeats Benchmark) Good alt for lyrics-heavy generation
YuE AI Variable Multi-GPU Open Up to 5min Full songs with synchronized vocals + instruments Too heavy for cloud API usage
Meta AudioCraft/MusicGen Variable 12GB+ MIT (research) Short clips PyTorch toolkit, sound effects + music Research only, ACE-Step surpasses it

Text-to-Speech

Model Params VRAM License Languages Voice Clone Best For
Kokoro-82M 82M CPU-only Apache 2.0 6 No Accessibility, narration, ultra-lightweight
Voxtral TTS 4B 4B 8GB+ Mistral Open 9 EU 3-second clone Artist voice features, notifications
Qwen3-TTS 0.6B/1.7B ~0.5-1.1GB Apache 2.0 10 Describe voice in text Creative voice design, CJK support
Higgs Audio V2 3B (Llama) ~2GB Open Multi Yes Emotional expressiveness
Bark (Suno) Transformer 4GB+ MIT Multi No Laughter, sighs, background sounds

Embeddings / RAG

Model Params VRAM License Context MTEB Score Best For
BGE-M3 568M ~1.2GB Apache 2.0 8,192 68.1 (EN) Default RAG, 1000+ languages, triple retrieval
Qwen3-Embedding-8B 8B ~5GB Apache 2.0 32,000 72.4 (#1 self-hosted) Scale-up option, code embeddings, Matryoshka dims
Jina Embeddings v4 3.8B ~3GB Apache 2.0 8,192 High Multimodal (text+images+PDFs in one space)

Coding / Agentic

Model Params VRAM License Context Key Feature Best For
OmniCoder-9B 9B ~5.5GB Apache 2.0 262K (1M+) Trained on Claude Opus 4.6 traces Lightweight local coding agent
Qwen3-Coder-Next 80B 80B MoE 32GB+ Apache 2.0 128K Multi-file repo editing Heavy local coding (Claude Code backend)
GLM-4.7-Flash 355B MoE (30B@24GB) 24GB MIT 200K Best 24GB agentic coding model Dedicated coding workstation
Qwen 2.5 Coder 14B 14B ~8.8GB Apache 2.0 128K FIM for IDE autocomplete, 85% HumanEval Solo vibe coding on 16GB VRAM

General Purpose (Reference)

Model Params (Active) VRAM License Context Intelligence Index Notable
GLM-5 744B (40B active) 80GB+ MIT 205K 50 (#1 open) Zero US hardware, SWE-bench 77.8%
MiniMax M2.5 230B (10B active) 128GB+ RAM Modified MIT 256K 46 Most-used on OpenRouter, interleaved thinking
Qwen3.5-27B 27B dense ~17GB Apache 2.0 262K Anti-slop, fits single 24GB GPU, 201 languages
Llama 4 Scout 109B (17B active) 48GB+ Llama 4 10M tokens Feed entire codebases
GPT-OSS 120B 117B MoE 80GB OpenAI Open Long Matches o4-mini OpenAI's first open-weight since GPT-2

ZAO OS Integration

Already Built — Upgrade Path

ZAO OS already integrates ACE-Step v1-3.5B for AI music generation:

  • API route: src/app/api/music/generate/route.ts — calls HuggingFace Inference API with ACE-Step/ACE-Step-v1-3.5B
  • Component: src/components/music/AiMusicGenerator.tsx — genre pills, prompt input, duration slider, audio playback
  • Upgrade: Change HF model ID to ACE-Step/ACE-Step-v1.5-XL for dramatically better output quality. ACE-Step 1.5 adds LoRA personalization, cover generation, vocal-to-BGM conversion, and 50+ language lyrics. No code changes needed beyond the model ID — same HuggingFace Inference API

Not Yet Built — Future Integration Points

Feature Model Integration Point Priority
Music recommendation RAG BGE-M3 src/lib/music/library.ts — replace textSearch with pgvector similarity P2 — needs Supabase pgvector extension enabled
Research library search BGE-M3 or Qwen3-Embedding 247+ research docs in research/ — semantic search over the knowledge base P3 — useful when library exceeds 300 docs
Proposal read-aloud Kokoro-82M src/components/governance/ — add audio playback for governance proposals P3 — accessibility feature
Artist voice intros Voxtral TTS 4B src/components/members/ — personalized audio greetings on profile pages P4 — nice-to-have
Live room transcription Whisper Large-v3 src/components/spaces/ — real-time captions. Doc 216 recommends Deepgram for live, Whisper for archival P2 — accessibility + searchability
Album art generation FLUX.1 Schnell New route src/app/api/music/generate-art/route.ts — pair with music gen P3 — artist tool

Current Search Architecture vs Vector Search

ZAO OS currently uses Supabase full-text search (PostgreSQL tsvector):

  • src/lib/music/library.ts:112query.textSearch('search_vector', search, { type: 'websearch' })
  • src/app/api/search/route.ts — unified search across members, tracks, proposals, casts

To add semantic/vector search, enable Supabase pgvector extension and embed content with BGE-M3 (568M, fits on any GPU with 4GB+). This would enable "find songs that sound like X" and "find research docs about Y concept" even when exact keywords don't match.

Hardware Reference

For ZAO's VPS (Hostinger KVM 2 at 31.97.148.88) and Zaal's local dev:

Tier Budget GPU Models That Fit
Minimal (CPU-only) $0 None Kokoro-82M (TTS), LFM-2.5-350M, Qwen3.5-2B
Consumer (8GB VRAM) $200-500 RTX 3060/4060 OmniCoder-9B, BGE-M3, Qwen 2.5 Coder 14B, Phi-4-mini
Prosumer (16GB VRAM) $500-1K RTX 4070Ti/4080 FLUX.1 Schnell, Wan 2.2 (1.3B), ACE-Step 1.5 local
Enthusiast (24GB VRAM) $1-2K RTX 3090/4090 Qwen3.5-27B, GLM-4.7-Flash, FLUX.2 Dev, Wan 2.2 14B
Mac Studio (64-128GB) $2-4K Apple Silicon MiniMax M2.5 (60+ t/s on M5 Max 128GB), Llama 4 Scout

For ZAO's HuggingFace Inference API approach (current): No local GPU needed. HF handles compute. Upgrade the model ID and you're done.

Key Benchmarks & Resources

Resource URL What It Tracks
Open LLM Leaderboard huggingface.co/spaces/open-llm-leaderboard General LLM rankings
Artificial Analysis artificialanalysis.ai/leaderboards/models Intelligence Index, speed, cost
MTEB Leaderboard huggingface.co/spaces/mteb/leaderboard Embedding model rankings
LiveCodeBench livecodebench.github.io Coding model benchmarks
SWE-bench swebench.com Real-world software engineering
VRAM Calculator apxml.com/tools/vram-calculator Check if a model fits your hardware
Ollama ollama.com One-command local model runner
LM Studio lmstudio.ai GUI for running local models
ComfyUI github.com/comfyanonymous/ComfyUI Image/video generation workflows

Cross-References to Existing Research

Doc Topic Relationship
Doc 209 AI Video Generation Tools Cloud-based video gen (Seedance, Neural Frames). This doc adds local alternatives (Wan 2.2, CogVideoX, Mochi 1)
Doc 216 AI Features for Live Rooms Recommends Deepgram for live transcription. Whisper Large-v3 from this doc is the local/archival alternative
Doc 42 Supabase Advanced Patterns Covers pgvector setup. BGE-M3/Qwen3-Embedding would plug into this
Doc 130 Next Music Integrations AI recs via pgvector mentioned. BGE-M3 is the embedding model to use
Doc 227 Agentic Workflows 2026 Local coding models (OmniCoder-9B, Qwen3-Coder) could serve as cheaper agent backends
Doc 161 Agent Harness Engineering LangChain/agent patterns that could use local models as backends

Article Quality Assessment

The @meta_alchemist article is high quality and mostly accurate as of April 2026:

Strengths:

  • Comprehensive coverage across 10 categories (50 models + extras)
  • Hardware-specific VRAM requirements for every model
  • License information included (critical for commercial use)
  • Benchmark sources cited
  • Practical "cheat code" workflow at the end

Minor Inaccuracies/Notes:

  • ACE-Step 1.5 article says "12GB VRAM" for RTX 3090 — actual RTX 3090 has 24GB VRAM. ACE-Step 1.5 needs <4GB VRAM per official docs
  • "Bark" link has a typo: suno/barkt instead of suno/bark
  • Some benchmark numbers are hard to independently verify (e.g., exact Intelligence Index scores)
  • MiniMax M2.5 "Modified MIT" license detail is important — requires displaying "MiniMax M2.5" in UI for commercial use
  • HeartMuLa-oss-3B "internal 7B version" is not publicly available — only the 3B is open

ZAO-Specific Gaps:

  • No coverage of Farcaster-specific AI models or agents
  • No mention of Supabase pgvector for embeddings (the typical indie stack approach)
  • Music gen section could have been deeper — ACE-Step 1.5 deserves more detail given its April 2, 2026 XL release

Sources