Claude Sonnet 4.7
AnthropicThe current default. Top-tier coding + tool use, 1M context.
vendor site ↗ experiment · 13 · model landscape
// madcool · lab · deep research
AI Model Landscape
Two views of the field. Below the fold: 25 open-weights models small enough to host yourself (full-precision footprint under 500 GB) with one-click links to source, weights, and a hosted-inference button on Featherless where available. Then: a regularly-updated multi-axial comparison of 11 cloud models scored across eight capability axes.
Cloud · Multi-Axial Comparison
Subjective curated scores. Click model chips to overlay up to four at once — the shape tells the story.
Claude Opus 4.7
AnthropicAnthropic's heavyweight. Deepest reasoning, premium tier.
vendor site ↗Claude Haiku 4
AnthropicCheap + fast Anthropic. Surprisingly capable for the price.
vendor site ↗GPT-5
OpenAIOpenAI's frontier. Unified reasoning + agentic stack.
vendor site ↗GPT-4o
OpenAIMultimodal workhorse. Solid all-rounder, native audio + vision.
vendor site ↗Gemini 2.5 Pro
GoogleGoogle's flagship. 2M context, video native, integrated thinking.
vendor site ↗Gemini 2.5 Flash
GoogleCheap Gemini. Astonishingly long context for the price.
vendor site ↗Grok 4
xAIxAI's frontier model. Real-time X data, large context, growing tool stack.
vendor site ↗DeepSeek V3.1
DeepSeekOpen-weights frontier. 671B MoE, hosted API at 1/10th the price.
vendor site ↗Mistral Large 2
Mistral AIEuropean frontier. EU-data-resident inference, strong multilingual.
vendor site ↗Command R+
CohereRAG-first cloud model. Citation-aware, tool-use solid.
vendor site ↗Local · Open Weights ≤ 500 GB
Filter by vendor or kind. Click any link to jump to source, weights, or a hosted runner on Featherless.
kind
vendor
DeepSeek V2.5
MoE generalist + coder fused. Activates 21B per token.
DeepSeek License
Mixtral 8x22B
MoE workhorse. Apache 2.0, fast at scale.
Apache 2.0
DBRX Instruct
MoE built for the data-platform crowd. Long context, good code.
Databricks Open Model
Mistral Large 2 (123B)
Mistral's flagship dense weights. Strong multilingual + function calling.
Mistral Research
Command R+ 104B
RAG + tool-use specialist. Non-commercial weights.
CC-BY-NC 4.0
Llama 3.2 90B Vision
Multimodal Llama. Image + text reasoning at frontier scale.
Llama 3.2 Community
Qwen 2.5 72B Instruct
Top open-weights generalist on LMSys. Tool-use solid.
Qwen License
Llama 3.3 70B Instruct
Meta's 2025 70B refresh. Closes most of the 405B gap with smaller footprint.
Llama 3 Community
Nemotron-70B Instruct
Llama 3.1 70B finetune. RewardBench leader at release.
Llama 3.1 Community
Hermes 3 70B
Steerable Llama 3 finetune. Sharper persona + tool-use control.
Llama 3.1 Community
Yi-1.5 34B
Apache-licensed bilingual model. Strong English + Chinese.
Apache 2.0
Qwen 2.5 32B Instruct
Two-card sweet spot. Apache-licensed Qwen sibling.
Apache 2.0
Qwen 2.5 Coder 32B
Best open-weights coder. Competitive with Sonnet on SWE-bench Verified.
Apache 2.0
QwQ 32B Preview
Open o1-style reasoning model. Long chain-of-thought, math heavy.
Apache 2.0
Gemma 2 27B
Google's open weights. Sliding-window attention, efficient.
Gemma Terms
Mistral Small 22B
Single-GPU dense model. Strong for size.
Mistral Research
DeepSeek Coder V2 Lite
Punchy MoE coder, single-GPU. 128K context, FIM training.
DeepSeek License
Phi-4 14B
Synthetic-data trained. Math + reasoning above its weight class.
MIT
OLMo 2 13B
Fully open: weights + data + training code. The open-science benchmark.
Apache 2.0
Llama 3.2 11B Vision
Compact vision-language model. Single-GPU territory.
Llama 3.2 Community
Gemma 2 9B
Small but mighty. Beats Llama 3 8B on most benchmarks.
Gemma Terms
Llama 3.1 8B Instruct
The workhorse small model. Fits on a single 24GB card.
Llama 3.1 Community
Granite 3 8B
IBM's enterprise-trained open model. Code + instruction tuned.
Apache 2.0
Llama 3.2 3B
Edge-deployment Llama. Pruned + distilled from 8B.
Llama 3.2 Community
SmolLM2 1.7B
On-device class. Runs on a Raspberry Pi 5 with room to spare.
Apache 2.0
Numbers are curated estimates — not a leaderboard. The shape of the
radar matters more than any single number. Refreshed monthly from
public/data/ai-models-cloud.json and
public/data/ai-models-local.json.