Click LLM
Local open-weight LLM picks by hardware tier
This post is the maintained recommendation source for Click LLM's local model picker. The app inspects the local device, reads the machine-readable guide below, and chooses the smallest model that fits the user's RAM and free disk while still matching the workload profile.
The practical constraint is simple: model weights are only the beginning. The host also needs space for the download, memory for the runtime and context window, and enough headroom that the desktop remains usable. Click LLM keeps a disk buffer and avoids recommending a model that is technically impressive but likely to fail during ollama pull.
Quick Picks
For an 8 GB machine, start with llama3.2:3b or phi4-mini:3.8b. These are small enough to be useful on constrained hardware and still expose tool-oriented capabilities in Ollama's catalog.
For a 16 GB laptop, qwen3:4b is the safest default when free disk is tight; qwen3:8b becomes the better general-purpose choice once there is at least 6 GB free. For local reasoning experiments, use deepseek-r1:8b.
For 32 GB unified memory or a modest NVIDIA GPU, qwen3:14b and gemma3:12b are the first larger step. gemma3:12b is useful when vision input matters; qwen3:14b is the default text assistant pick.
For 64 GB+ workstations, mistral-small:22b, magistral:24b, and qwen3.6:27b become reasonable. For very large local boxes, qwen3.6:35b-a3b and llama4:scout are catalog options, but Click LLM should only recommend them when memory, disk, and acceleration are clearly present.
Source Notes
This guide is tuned for Ollama because Click LLM's current installer/status/prompt path is Ollama-first. The sizes and context windows below are based on Ollama model tag pages checked on 2026-06-28:
Machine-Readable Guide
Click LLM parses this block. Keep the schema stable when updating the prose.
{
"schema": "click-llm.local-model-guide.v1",
"updated": "2026-06-28",
"source_post": "click-run-server/blog/click-llm/local-open-models-hardware-guide.md",
"policy": {
"runtime": "ollama",
"disk_buffer_gb": 0.75,
"notes": "Prefer the smallest model that fits RAM and disk. Do not auto-install; only install after an explicit user action."
},
"models": [
{
"id": "llama3.2:3b",
"label": "Llama 3.2 3B",
"runtime": "ollama",
"min_ram_gb": 8,
"min_vram_gb": 0,
"approx_size_gb": 2.0,
"context_window": "128K",
"profile": "small",
"capabilities": ["tools", "text"],
"source_url": "https://ollama.com/library/llama3.2/tags",
"best_for": "fast local chat, summarization, and simple coding help on constrained machines"
},
{
"id": "phi4-mini:3.8b",
"label": "Phi-4 Mini 3.8B",
"runtime": "ollama",
"min_ram_gb": 8,
"min_vram_gb": 0,
"approx_size_gb": 2.5,
"context_window": "128K",
"profile": "small",
"capabilities": ["tools", "text", "math"],
"source_url": "https://ollama.com/library/phi4-mini/tags",
"best_for": "small local assistant work where function calling and math are more important than broad knowledge"
},
{
"id": "qwen3:4b",
"label": "Qwen3 4B",
"runtime": "ollama",
"min_ram_gb": 12,
"min_vram_gb": 0,
"approx_size_gb": 2.5,
"context_window": "256K",
"profile": "balanced",
"capabilities": ["tools", "thinking", "text"],
"source_url": "https://ollama.com/library/qwen3/tags",
"best_for": "general assistant work on 16 GB laptops with limited free disk"
},
{
"id": "qwen3:8b",
"label": "Qwen3 8B",
"runtime": "ollama",
"min_ram_gb": 16,
"min_vram_gb": 0,
"approx_size_gb": 5.2,
"context_window": "40K",
"profile": "balanced",
"capabilities": ["tools", "thinking", "text"],
"source_url": "https://ollama.com/library/qwen3/tags",
"best_for": "general assistant work, coding, and tool-oriented prompts on 16 GB+ machines"
},
{
"id": "gemma3:4b",
"label": "Gemma 3 4B",
"runtime": "ollama",
"min_ram_gb": 16,
"min_vram_gb": 0,
"approx_size_gb": 3.3,
"context_window": "128K",
"profile": "multimodal",
"capabilities": ["vision", "text"],
"source_url": "https://ollama.com/library/gemma3/tags",
"best_for": "small multimodal experiments when image input matters"
},
{
"id": "deepseek-r1:8b",
"label": "DeepSeek-R1 8B",
"runtime": "ollama",
"min_ram_gb": 16,
"min_vram_gb": 0,
"approx_size_gb": 5.2,
"context_window": "128K",
"profile": "reasoning",
"capabilities": ["tools", "thinking", "text"],
"source_url": "https://ollama.com/library/deepseek-r1/tags",
"best_for": "small local reasoning experiments"
},
{
"id": "gemma3:12b",
"label": "Gemma 3 12B",
"runtime": "ollama",
"min_ram_gb": 24,
"min_vram_gb": 0,
"approx_size_gb": 8.1,
"context_window": "128K",
"profile": "multimodal",
"capabilities": ["vision", "text"],
"source_url": "https://ollama.com/library/gemma3/tags",
"best_for": "larger local multimodal work on 24 GB+ memory machines"
},
{
"id": "qwen3:14b",
"label": "Qwen3 14B",
"runtime": "ollama",
"min_ram_gb": 32,
"min_vram_gb": 12,
"approx_size_gb": 9.3,
"context_window": "40K",
"profile": "larger",
"capabilities": ["tools", "thinking", "text"],
"source_url": "https://ollama.com/library/qwen3/tags",
"best_for": "higher-quality local chat and coding on 32 GB+ machines"
},
{
"id": "mistral-small:22b",
"label": "Mistral Small 22B",
"runtime": "ollama",
"min_ram_gb": 48,
"min_vram_gb": 16,
"approx_size_gb": 13.0,
"context_window": "128K",
"profile": "larger",
"capabilities": ["tools", "text"],
"source_url": "https://ollama.com/library/mistral-small/tags",
"best_for": "larger instruction-following workloads below the 70B class"
},
{
"id": "magistral:24b",
"label": "Magistral 24B",
"runtime": "ollama",
"min_ram_gb": 48,
"min_vram_gb": 16,
"approx_size_gb": 14.0,
"context_window": "39K",
"profile": "reasoning",
"capabilities": ["tools", "thinking", "text"],
"source_url": "https://ollama.com/library/magistral/tags",
"best_for": "larger local reasoning workloads"
},
{
"id": "qwen3.6:27b",
"label": "Qwen3.6 27B",
"runtime": "ollama",
"min_ram_gb": 64,
"min_vram_gb": 24,
"approx_size_gb": 17.0,
"context_window": "256K",
"profile": "larger",
"capabilities": ["vision", "tools", "thinking", "text"],
"source_url": "https://ollama.com/library/qwen3.6/tags",
"best_for": "high-end local agentic coding and long-context work"
},
{
"id": "qwen3.6:35b-a3b",
"label": "Qwen3.6 35B-A3B",
"runtime": "ollama",
"min_ram_gb": 64,
"min_vram_gb": 24,
"approx_size_gb": 24.0,
"context_window": "256K",
"profile": "larger",
"capabilities": ["vision", "tools", "thinking", "text"],
"source_url": "https://ollama.com/library/qwen3.6/tags",
"best_for": "high-end local work when disk and accelerator headroom are available"
},
{
"id": "llama4:scout",
"label": "Llama 4 Scout",
"runtime": "ollama",
"min_ram_gb": 128,
"min_vram_gb": 48,
"approx_size_gb": 67.0,
"context_window": "10M",
"profile": "frontier",
"capabilities": ["vision", "tools", "text"],
"source_url": "https://ollama.com/library/llama4/tags",
"best_for": "very large local multimodal context experiments"
}
]
}Update Rule
When this post is updated, update both the prose and the JSON block. Click LLM expects schema, updated, policy.disk_buffer_gb, and models[] to remain present. Each model should include id, label, runtime, min_ram_gb, min_vram_gb, approx_size_gb, profile, source_url, and best_for.