Open-Source AI Model Comparison

Open-Source AI Model Comparison

"Every AI tool that you're using right now, you don't actually own. You're just renting them." — Kacper Rutkiewicz (Source: Which Open Source AI Model Should You Use, YouTube, 2026-06-20)

China is shipping open-weight models at the frontier — GLM-5.2, DeepSeek V4, Kimmy, Qwen, MiniMax — that match or approach Opus 4.8 quality at 5-10x lower cost. The West (Meta Llama, Mistral, Google Gemma) is 6-12 months behind. This page compares eight open-source models across capability, cost, and local-run feasibility, based on side-by-side testing in the Claude Code harness.


The Landscape

Two camps are shipping open-source AI:

Camp Key Players Status
China Z.ai (GLM), DeepSeek, Alibaba (Qwen), Moonshot (Kimmy), MiniMax Frontier-quality, aggressive open-weight releases
West Meta (Llama), Mistral (France), Google (Gemma) 6-12 months behind; restrictive licenses; smaller models

A common myth: Chinese models are "API only." This is false. DeepSeek put a 1.6-trillion-parameter model on Hugging Face. Z.ai's GLM-5.2 has been quantized to 2-bit by Unsloth and runs on a Mac Mini. Both sides ship open-weight models — the West is simply not competitive at the top tier right now.

"Open-source" does not mean "runs on your laptop." Flagship models range from 230B to 1.6T parameters. You either invest in serious hardware or rent them via API providers like OpenRouter. Smaller distilled versions can run locally.

Model Rankings

1. GLM-5.2 — Best Daily Driver

Attribute Detail
Maker Z.ai (Tsinghua University, China)
Size 753B parameters
Best at Agentic coding and planning — plans, builds, tests, fixes itself
Cost vs Opus ~5x cheaper ($0.20 vs $1.60 in side-by-side habit-tracker test)
Run locally? No (flagship). Yes via Unsloth distillation (82% accuracy on Mac Mini)

The model the creator uses as his daily driver alongside Opus 4.8. Same result, 5x fewer tokens.

2. DeepSeek V4 — Deep Reasoning

Attribute Detail
Maker DeepSeek (China)
Size 1.6T total (49B active — MoE architecture)
Best at Reasoning across massive context (1M window); multi-document conflict analysis
Cost vs Opus ~5x cheaper
Run locally? No (flagship). Distilled versions available via Unsloth/Hugging Face

The model that proved China would open-weight their biggest models. Highly efficient for its size.

3. Kimmy K 2.7 — Long Agentic Work

Attribute Detail
Maker Moonshot AI (China)
Size ~1T parameters
Best at Multi-stage tasks where weaker models forget steps; follows plans faithfully
Cost vs Opus ~7x cheaper ($1.33 vs $7.52)
Run locally? No (flagship)

Completed a 6-stage CLI build in 4:21 vs Opus's 6:16. Same functional result — differences were taste, not capability.

4. Qwen — Reliable Workhorse

Attribute Detail
Maker Alibaba (China)
Size Varies (flagship API-only; open models available down to 4-bit 8B)
Best at Clean, readable code; widest size range in the list
Cost vs Opus ~4x cheaper ($2.00 vs $8.73)
Run locally? Yes — Qwen 3 Coder Next 4-bit 8B runs on 8GB VRAM

The only model on this list the creator can confidently run locally and fully private. Qwen 3.7 Max is API-only, but the family has truly open models. Always check per-model, not per-brand.

5. MiniMax 3 — High-Volume King

Attribute Detail
Maker MiniMax (China)
Size Large MoE
Best at High-volume repetitive tasks — email triage, bulk responses, agents that never sleep
Cost $0.15/M input, $1.20/M output ($1.00 vs $9.83 Opus in 50-message test)
Run locally? Partially (needs good hardware)

Nearly 10x cheaper for the same result. If you're running AI agents at scale, this is the cost play.

6. Llama 4 — Largest Community, Disappointing Performance

Attribute Detail
Maker Meta (West)
Size Varies
Best at Multimodal (text + image); largest ecosystem of tools and tutorials
Cost vs Opus — (did not pass parity tests)
Run locally? Yes (multiple sizes available)
License Restrictive — not fully open

In testing, Llama struggled in the Claude Code harness: gave instructions instead of executing, couldn't use vision through the harness (text-only pipeline). Restrictive license is a red flag for true ownership.

7. Mistral Dev — True Ownership on One GPU

Attribute Detail
Maker Mistral (France)
Size Dev-sized (runs on single GPU)
Best at Coding on self-hosted hardware; fully open license
Cost vs Opus — (nowhere near Opus/Sonnet level)
Run locally? Yes — single GPU

Not competitive with frontier models, but genuinely ownable. If you want a completely private coding agent on your hardware, this is the best open option from the West.

8. Gemma — Ultralight Local

Attribute Detail
Maker Google (West)
Size Small
Best at Running on phones or 8GB VRAM; ultralight local tasks
Cost vs Opus — (not competitive)
Run locally? Yes — consumer hardware, even phones

Small and efficient, but underwhelming in testing. Only useful if your constraint is hardware, not capability.


Multimodal Support

Capability Models
Text + Image + Video MiniMax, Gemma
Text + Image Kimmy K 2.7, Llama 4
Text only GLM-5.2, DeepSeek V4, Qwen Max, Mistral Dev

Decision Guide

I want to... Use this model
Replace Opus as my daily coding driver GLM-5.2 — 5x cheaper, near-Opus agentic quality
Reason across huge documents or codebases DeepSeek V4 — 1M context, efficient MoE
Run a multi-stage agentic task that must not forget steps Kimmy K 2.7 — stays on plan
Run an AI model locally and privately on consumer hardware Qwen 3 Coder Next (4-bit 8B) — runs on 8GB VRAM
Process thousands of emails/support tickets per day at scale MiniMax 3 — $0.15/M input, 10x cheaper than Opus
Own a fully open coding model on a single GPU Mistral Dev — not frontier, but truly yours
Run a model on a phone or ultra-low-end hardware Gemma — ultralight, minimum capability

Cost Reality

All eight open-source models were tested in a single video session for under $2 total. Equivalent Opus 4.8 runs would have cost $40-50+. The takeaway is not that open-source models are "better" — it's that you don't need the most expensive model for every task.

"We don't need Albert Einstein to mop the floor." — Jack Roberts


Architecture Note

All tests were run through Claude Code + OpenRouter — same harness, different brains. This reinforces a key point: the harness matters as much as the model. The Western models (Llama, Gemma) underperformed specifically in the Claude Code tool-use pattern, not necessarily in raw intelligence. Different harnesses (Pi Agent, Codex, Hermes) might surface different strengths.


See Also