comind lab
Which model
for which job?
Online or local, big or small, paid or open. A live leaderboard by task, prices and context for every hosted model, and a calculator for what your own machine can run — with an honest tokens-per-second estimate. New models appear here automatically.
367
models rated by human votes
340
hosted models with prices
199
open-weights in the ranking
272
downloadable models sized
◆ ratings: LMArena, 2026-11-08 · prices: OpenRouter · weights: Hugging Face
question / A
Online or offline?
- Online (API)
- the strongest models, zero setup, pay per token. Default for real work.
- Offline (local)
- private data never leaves the machine, works without internet, free after the hardware. You trade quality and speed for it.
- Rule of thumb
- start online; go local when privacy, cost at volume, or offline use actually matter.
question / B
Pick by the task, not by the hype
- Code & agents
- sort the leaderboard by Coding, Web dev or Agentic — overall rank hides big gaps.
- Writing, chat
- Writing and Instructions columns; most top-20 models are fine here.
- Math, reasoning
- Math and Hard prompts; prefer models with a reasoning mode.
- Ukrainian & others
- Non-English column — some strong English models drop sharply.
question / C
Count the money
- Price is per 1M tokens
- 1M tokens ≈ 750k words. Input (what you send) and output (what it writes) are priced separately; output usually costs 3–5×.
- Agents eat input
- a coding agent re-sends the codebase context every step — input price and caching matter more than output.
- Cheap ≠ worse
- a flash/mini model one tier down often costs 10× less for 90% of the quality. Try it first.
question / D
How fast is fast enough?
- ~5–10 tok/s
- human reading speed. OK for chat, painful for anything long.
- 20–40 tok/s
- comfortable interactive use, code completion.
- 50+ tok/s
- agents and long generations — they produce far more text than you read.
- Locally
- speed is set by memory bandwidth, not by GPU compute. More GB/s = more tok/s.
| # | model | score | weights | $ in / out · 1M | context | features |
|---|
just released · not rated yet
Hosted in the last 60 days, waiting for enough votes. Prices are real; quality is unknown — try them on your own task.
quality
LMArena leaderboard: Elo-style ratings from millions of blind side-by-side human votes, per task. Agentic is a success score ×1000. When a model ships several effort levels we show its best one. A ~20-point gap is often within noise.
price, context, new models
OpenRouter catalog: the cheapest listed price per 1M tokens and max context. Direct provider prices can differ. Parameter counts and licences come from Hugging Face.
local estimate
Memory = weights (params × bits ÷ 8) + KV cache for the context + ~0.8 GB runtime. Speed ≈ 65% of memory bandwidth ÷ bytes read per token (active params for MoE, some estimated from config). Expect ±30%; prompt processing is not included. Data refreshes weekly.