Pick the best local AI model for your machine
Open-source LLMs are now competitive with GPT-4 / Claude / Gemini — if you run the right one. Tell us your chip (Apple M1 / M2 / M3 / M4 / M5, NVIDIA, AMD, CPU), RAM, and use case. We surface 3 ranked picks with verified HuggingFace links, install commands, and quotes from credible engineers running them.
Codersera's free Local AI Model Picker recommends the best open-source LLM to run on your own machine based on your chip family (Apple Silicon M-series, NVIDIA / AMD GPU, or CPU-only), available RAM or VRAM, primary use case (general chat, coding, creative writing, vision, uncensored / role-play, or reasoning), and your preferred runtime (Ollama, LM Studio, llama.cpp, MLX, vLLM). The tool covers state-of-the-art June 2026 models — Llama 4, DeepSeek V4 + Coder, Qwen 3.5 + Coder, Gemma 4, Phi-4, Mistral Small 3, Kimi K2.6, vision models, and uncensored Dolphin / Hermes fine-tunes. Ranked top 3 with the Ollama install command included.
Find your model
How to read the recommendations
Four inputs, four design tradeoffs you might want to know about.
Pick your hardware
Apple Silicon (M1 / M2 / M3 / M4) uses unified memory, so RAM = VRAM. NVIDIA / AMD GPUs use discrete VRAM, so pick the VRAM size, not system RAM. CPU-only setups should stick to ≤8B models.
Pick the use case
General chat, coding, creative writing, reasoning, vision (image / screenshot understanding), or uncensored / role-play. Each model has per-use-case quality scores; the picker weights them with hardware fit.
Pick a runtime (or leave it on No preference)
Ollama is the easiest entry point. LM Studio is the polished GUI. llama.cpp is the lowest-level path. MLX is fastest on Apple Silicon. vLLM is the production server for NVIDIA boxes.
Read the top 3 picks
We rank by use-case score + memory fit + runtime fit, and surface 3 strong options so you can A/B them. Each card shows active params, working-set memory, license, and an Ollama install command.
Quick picks for buying a new Mac
Already know you want a Mac? Here are three citation-backed picks by budget. Sources are the gold-standard local-LLM benchmark posts (Sean Kim's M4 Max benchmarks, MacRumors / Wccftech on M3 Ultra running DeepSeek V3, PopularAI's Mac mini guide).
The M4 Pro Mac mini hits the M-Pro memory-bandwidth tier (273 GB/s) where 8B and 30B models run at conversational speed. The 48 GB upgrade fits Qwen 3.6 27B + Qwen 3.6 35B-A3B comfortably.
Per Jun Song's hardware shortlist (351 likes), the M5 Pro is one of three picks worth recommending for serious local LLM use. 307 GB/s bandwidth, 4× AI uplift on Apple's metric vs M4. Runs Qwen 3.6 27B at ~63 tok/s.
The only consumer machine that runs DeepSeek V3 671B in unified RAM under 200 W. ~17-18 tok/s. Per Jun Song: M3 Ultra at 546-819 GB/s bandwidth crushes DGX Spark's 273 GB/s — "people know what's up."
Models in the catalog — verified as of 2026-06-16
Every Hugging Face URL below has been verified to return 200. When a forward-projected name doesn't have a real HuggingFace page yet, we point to the closest currently-shipping model and label it honestly. We refresh the catalog whenever a new generation lands.
Runtimes — what to install
A runtime is the program that loads the model weights and actually generates tokens. You only need one to start.
Ollama
The default. Install with `curl -fsSL https://ollama.com/install.sh | sh` (Linux) or download the macOS / Windows app. One-line model pulls (`ollama pull llama4:8b`). Exposes an OpenAI-compatible REST API on localhost:11434 — point your IDE assistant or chatbot at it.
LM Studio
Polished desktop GUI. Drag-drop .gguf model files, browse Hugging Face from inside the app, switch between models instantly, chat in a built-in UI. Also exposes an OpenAI-compatible server. Best entry point for non-CLI users.
llama.cpp
The C++ inference engine that powers most of the others. Pick this if you want fine-grained control: custom KV cache size, speculative decoding, multi-GPU layer splitting, embedded into your own binary. No GUI; you build and call it directly.
MLX
Apple's native ML framework. Fastest inference on M-series Macs because it uses unified memory + Metal directly without copies. Use mlx-lm (`pip install mlx-lm`) to load Hugging Face models; expect 2-3× the throughput of llama.cpp on the same hardware.
vLLM
Production inference server for NVIDIA GPUs. PagedAttention scheduler, continuous batching, multi-GPU tensor parallelism. Pick this if you are serving many concurrent users — not for desktop chat.
How much RAM do I actually need?
Rule of thumb at Q4_K_M quantization: working-set memory ≈ 0.6 GB per billion parameters, plus 2-4 GB for KV cache + overhead. Add 4-8 GB for OS / apps when sizing.
| Available RAM / VRAM | Best models | Notes |
|---|---|---|
| 8 GB | Gemma 4 2B · Phi-4-mini 3.8B | Tight. Stick to ≤4B models. Quality drops fast. |
| 16 GB | Llama 4 8B · Qwen 3.5 7B · DeepSeek Coder V4 6.7B | Sweet spot for laptops. 7-9B models run comfortably. |
| 24-32 GB | Qwen 3.5 32B · Gemma 4 27B · Mistral Small 3 (24B) | Real workstation territory. Frontier-class single-machine performance. |
| 48-64 GB | Llama 4 70B · Kimi K2.6 70B · Hermes 4 70B | Mac Studio M3 Ultra / dual RTX 4090 territory. Approaches GPT-4-class quality. |
| 96 GB+ | Command R+ (104B) · DeepSeek V4 MoE 670B | Server / workstation. You can run almost any open model. |
Censored vs uncensored — what's the practical difference?
Base open models (Llama 4, DeepSeek V4, Qwen 3.5, Gemma 4, Phi-4) ship with safety / RLHF tuning that refuses certain prompts — graphic content, security-relevant content, role-play that crosses internal policies. For most users that's fine: it matches the ChatGPT / Claude UX.
Fine-tunes like Dolphin (Cognitive Computations) and Hermes (Nous Research) strip those refusals via additional DPO training. They're used by security researchers, fiction writers, role-play communities, and red-teamers. The picker exposes the Show uncensored models only toggle so you can route to those when you need them.
Same legal status as the base models — both Dolphin and Hermes are publicly hosted on Hugging Face under the base model's license. You are responsible for how you use the output.
Frequently asked questions
Why local AI instead of ChatGPT / Claude / Gemini?
Privacy (data never leaves your machine), zero per-token cost after the model is downloaded, works offline, no rate limits, and full control over the system prompt + sampling. Frontier-class open models like Llama 4 70B and DeepSeek V4 now approach proprietary quality.
What does Q4_K_M mean?
A 4-bit quantization scheme used by llama.cpp and Ollama. It reduces a 16-bit fp model to roughly 25% of its weight size with negligible quality loss for the majority of tasks. Our memory estimates assume Q4_K_M unless noted otherwise.
What is the difference between MoE and dense models?
Mixture-of-Experts models (e.g., DeepSeek V4: 670B total / 37B active) only route through a fraction of parameters per token. The full weights still need to live in memory, but inference speed is closer to the active parameter count — important when comparing throughput.
Which runtime should I install first?
Ollama. It is the lowest-friction path — one-line install, model registry with one-line pulls, OpenAI-compatible REST API on port 11434. Once you have a few models locally, look at LM Studio for the GUI or MLX for raw speed on Apple Silicon.
Are uncensored models illegal?
No. The Dolphin and Hermes fine-tunes are publicly hosted on Hugging Face under the base model license. They are widely used by security researchers, creative writers, and roleplay communities. Treat them like any other tool — you are responsible for how you use them.
What about API keys for OpenAI / Anthropic?
Not needed. This tool covers local-only models that run entirely on your hardware. You download the weights once and run them with Ollama / llama.cpp / MLX / vLLM. Zero API keys, zero per-token cost.
Does the tool send my hardware spec to a server?
No. The recommender runs 100% in your browser. Codersera does not log your chip, RAM, use case, or any picks. The model catalog is bundled into the page as a static dataset.
How often is the model catalog refreshed?
Whenever a major model lands (a new Llama / DeepSeek / Qwen / Gemma generation, or a notable fine-tune). Each model card carries a lastReviewed date so you can see how current the recommendation is.
