Qalarc / Local Model Atlas

Every local AI model we actually run — measured, not marketed

By Alexei Brown, Qalarc (Sydney) · hardware: GMKtec EVO-X2 AI (Ryzen AI Max+ 395, 128GB) + 8060S workstation · verified 2026-09-10 · llms.txt
These are the models installed on our studio machines right now — 19 local LLMs, the full ComfyUI image/video stack, and the audio studio — with real sizes, measured speeds and honest use-notes from daily work. If a model is not earning its disk space, it is not on this page.

How to read this atlas

Every model below is installed and in daily studio use — this is not a listicle. Sizes are real disk footprints from our machines; speed notes are measured, not vendor claims. Reference hardware: GMKtec EVO-X2 AI (AMD Ryzen AI Max+ 395, 128GB unified memory) plus an 8060S-class iGPU workstation. The whole stack runs through AI-OS + ComfyUI + Ollama, fully offline.

Language models (Ollama)

ModelSizeUse it forMeasured notes
GLM 4.7 Flash (30B-A3B MoE)19GBDaily driver: chat, analysis, agents38 tok/s on 128GB unified; :32k variant for long context
Qwen3.8 (27B)17GBGeneral + coding~40 tok/s class on Strix Halo
Ornith 1.5 (35B-A3B MoE / 9B)22GB / 6.7GBStrong MoE alternative; 9B fits smaller machinesNew addition, 32k context variants
Qwen2.5-Coder (32B/14B/7B)19/9/4.7GBSpecialized code generationThe 7B is the best coding-per-GB in our set
DeepSeek-R1 7B4.7GBReasoning chainsSlow but deliberate
Llama 3.2 3B / Gemma2 2B2GB/1.6GBEdge tasks, classificationInstant load on anything
nomic-embed-text274MBEmbeddings / RAGPowers local search

Image & video models (ComfyUI)

ModelSizeUse it forMeasured notes
z-Image Turbo12GBFast image generation — our house generator~26s per 1216x832 on the mini PC; generated every thumbnail on qalarc.com
FLUX.1 Krea-dev (fp8)12GBHighest-quality creative imagesSlower than turbo, noticeably better composition
Wan 2.2 ti2v 5B9.4GBFast text-to-videoThe practical entry point for local video
Wan 2.2 Animate 14B17GBCharacter animation from reference64GB+ recommended
Wan 2.2 S2V 14B16GBSpeech-driven videoAudio-to-talking-head pipeline
Wan 2.1 flf2v 720p 14B16GBFirst/last-frame video interpolationKeyframe-driven clips
LTX 2.3 22B (+distilled LoRA)VAEs 2GB + LoRA 7GBVideo with audio trackOne-pass video+audio generation
SAM 2.1 Hiera~400MBSegmentationVideo object masking
RIFE v4.26smallFrame interpolationSmooths generated video

LoRA stack worth knowing: cyberrealistic-z-Image (photoreal), unstableRevolution V3 (SD-class creative), famegrid (house style). Text encoders that matter: Qwen3-4B and Qwen2.5-VL-7B (multimodal understanding), T5-XXL / UMT5-XXL (classic conditioning).

Audio models (the audio studio)

ModelUse it forNotes
ACE-StepFull music tracks from text promptsGenre-accurate, runs locally on GPU
Stable Audio 3General audio synthesis, SFXFast iteration
F5-TTSVoice cloning from short samplesBest quality-per-ease in our set
GPT-SoVITSMultilingual voiceCross-lingual cloning
MelBand RoformerStem separationState-of-the-art demixing
VocosNeural vocoder (mel to audio)The speed layer under TTS pipelines
Wav2Vec2 (large-xlsr)Speech recognition featuresFeeds audio-driven pipelines

The honest caveats

"Local" still means downloading these weights once — the full stack above is roughly 200GB of models. Speed notes are from AMD unified-memory hardware; NVIDIA discrete VRAM behaves differently (faster compute, tighter capacity). Model quality moves fast — this atlas is dated and re-verified against what is actually installed, not what is hyped.

Why this stack (the strategy)

Unified memory changed the math: a 128GB mini PC runs 30B MoE language models and 14B video models that would need an expensive discrete-GPU rig otherwise. The pattern we would recommend to anyone: one Strix-Halo-class machine, Ollama for text, ComfyUI for vision, a dedicated audio stack — everything offline, zero marginal cost per generation.

AI-OS — the workstation stack Local media guide The mini PC writeup

When NOT to go local

Occasional use (a few images a month) is cheaper on cloud services. Frontier-class quality still lives behind APIs. Local wins on volume, privacy, custom characters, and cost-per-generation at scale.