AI Agent Store/Model comparisonsBuild an AI worker ↗THE MODEL SELECTION SERIES
Open weights are downloadable files containing a trained model, available under its license. That can create more deployment options, but it does not automatically make a model free to run, fully open source or unrestricted for commercial use.
Data reviewed · LiveBench question set: 25 June 2026 · View every source ↗ · Download the data ↓
You can download model files to your own hardware or pay a service to run them. The prices below cover hosted services. Running a model yourself adds hardware, electricity and maintenance costs. Compressed files use less memory, but can change answer quality.
Before using a model in a business, read its license. Some downloads are for research only; others restrict commercial hosting. Record the exact version and file, since a smaller compressed download may behave differently from the original.
A model may be able to request a tool, such as a search, while your local software still needs to connect and run it. Check that answers follow your required format, tool calls work and errors are handled. Use the same software and request sizes you plan to use in practice.
If you run a model yourself, include hardware, electricity and time spent maintaining it, even when it sits idle. Measure how many useful tasks your computer completes in a typical day. Compare that total cost with paying a hosted service for the same work.
Check three things before downloading: free disk space, memory while the model is running, and software that supports the exact file. A “24GB model” can need much more than 24GB on disk.
The publisher lists a 70.6 GiB download, 19.92 GiB of GPU weights and about 21.1 GiB while running its 64K-context example. A large lookup table stays on a fast drive instead of the graphics card.
That example targets a 24GB graphics card. Longer requests and different settings can need more memory. Standard llama.cpp cannot load the file; use the compatible publisher build.
Check the files, setup and license ↗Flora TempSeg is an add-on, called an adapter. It also needs the specified Qwen3.5-9B model. The adapter’s file size alone therefore understates what you need to download and run.
It labels action boundaries in videos for non-commercial robotics research. Check the intended task and license as well as the memory requirement.
Check the required base and research limits ↗These are publisher specifications, not our hardware measurements. Try your longest expected request and the number of requests you need to run at once before choosing hardware.
The comparison starts with open-weight models. Prices represent hosted API endpoints; check the model license before self-hosting.
YOUR WORKLOAD. YOUR PRIORITIES.
Tokens are the small pieces of text a model reads or produces. Include charged thinking tokens in output.
This estimate uses your token counts and standard text API rates. It excludes cache discounts, tool fees, retries and provider premiums. If your input and output exceed a model’s listed capacity, we leave its cost and Value rating blank. A large capacity does not guarantee that it will use every detail correctly.
Shaded zone: 73.4+ points and $10.13 or less. Numbers mark models with strong price–quality tradeoffs in or near this zone. On the dotted line, no other shown model is cheaper with an equal or higher score, or scores higher at the same price. Your chosen measure, usage and filters change this shortlist.
Practical Value · General agent
70% Agent Fit + 30% affordability. Rankings use your workload. Filtering the list does not change how scores are calculated.
THE EVIDENCE, SIDE BY SIDE
DeepSeek · max
Moonshot AI · source default
Alibaba · source default
DeepSeek · experimental · source default
Z.ai · source default
Alibaba · source default
DeepSeek · source default
Mistral · high
DeepSeek · source default
Thinking Machines · xhigh
Z.ai · source default
Z.ai · source default
| Compare | Input / outputUSD / 1M tokens | ||||||||
|---|---|---|---|---|---|---|---|---|---|
01 DeepSeek V4.1 Flash ↗DeepSeek · maxOpen weights | 78.9 | 83.2 | — | 81.1 | — | $1.45 | 1.05M | $0.045 / $0.5OpenRouter catalog | |
02 Kimi K3 ↗Moonshot AI · source defaultOpen weights | 77.4 | 65.4 | — | 79.2 | — | $33.92 | 1.05M | $0.192 / $16OpenRouter catalog | |
03 Qwen3.8 Flash Next ↗Alibaba · source defaultOpen weights | 76.2 | — | — | 76.2 | — | — | Not verified | Not verified | |
04 DeepSeek V4 Flash Vision Exp ↗DeepSeek · experimental · source defaultOpen weightsPreview | 75.1 | 79.0 | — | 76.8 | — | $3.45 | 1.05M | $0.2156 / $0.6468OpenRouter catalog | |
05 GLM-5.3 ↗Z.ai · source defaultOpen weights | 73.7 | 77.5 | — | 76.1 | — | $3.5 | 1.05M | $0.0699 / $1.4OpenRouter catalog | |
06 Qwen3.8 27B ↗Alibaba · source defaultOpen weights | 73.5 | 75.9 | — | 75.3 | — | $5.25 | 1.02M | $0.15 / $1.88OpenRouter catalog | |
07 DeepSeek V4 Pro 0813 ↗DeepSeek · source defaultOpen weights | 73.3 | 73.2 | — | 77.4 | — | $10.35 | 1.05M | $0.115 / $4.6OpenRouter catalog | |
08 Mistral Large 4 ↗Mistral · highOpen weightsPreview | 72.0 | 71.3 | — | 71.8 | — | $11.2 | 1.05M | $0.7 / $2.1OpenRouter catalog | |
09 DeepSeek V4 Flash 0731 ↗DeepSeek · source defaultOpen weights | 71.1 | 78.8 | — | 74.2 | — | $1.15 | 1.05M | $0.0553 / $0.3OpenRouter catalog | |
10 Inkling ↗Thinking Machines · xhighOpen weights | 68.9 | 64.5 | — | 71.9 | — | $18.1 | 524K | $1 / $4.05OpenRouter catalog | |
11 GLM-5.2 ↗Z.ai · source defaultOpen weights | 68.5 | 67.3 | — | 73.2 | — | $13.66 | 1.05M | $0.006 / $6.8OpenRouter catalog | |
12 GLM-5.3 Flash ↗Z.ai · source defaultOpen weights | 67.2 | 77.0 | — | 71.6 | — | $0.4682 | 1.05M | $0.0269 / $0.0996OpenRouter catalog | |
13 DeepSeek V4 Pro ↗DeepSeek · source defaultOpen weights | 67.1 | 61.2 | — | 71.6 | — | $22.4 | 1.05M | $1.6 / $3.2OpenRouter catalog | |
14 Kimi K2.6 ↗Moonshot AI · thinkingOpen weights | 66.9 | 66.7 | — | 70.5 | — | $13.32 | 262K | $0.65 / $3.41OpenRouter catalog | |
15 Kimi K2.7 Code ↗Moonshot AI · source defaultOpen weights | 64.8 | 65.7 | — | 68.4 | — | $13 | 262K | $0.6 / $3.5OpenRouter catalog | |
16 Nemotron 3 Ultra ↗NVIDIA · 550B A55B · source defaultOpen weights | 63.8 | 66.0 | — | 67.4 | — | $10.8 | 262K | $0.6 / $2.4OpenRouter catalog | |
17 DeepSeek V4 Flash ↗DeepSeek · source defaultOpen weights | 61.6 | 71.6 | — | 65.5 | — | $1.27 | 1.05M | $0.091 / $0.182OpenRouter catalog | |
18 Qwen3.6 27B ↗Alibaba · source defaultOpen weights | 60.0 | 64.4 | — | 64.0 | — | $9.9 | 262K | $0.45 / $2.7OpenRouter catalog | |
19 Nemotron 3.5 Lightning ↗NVIDIA · source default · thinking budget unspecifiedOpen weights | 41.6 | 58.6 | — | 44.5 | — | $1.1 | 262K | $0.07 / $0.2OpenRouter catalog |
— means a comparable result is unavailable. Consensus combines results for only 10 matched configurations. Model IDs, reasoning settings, source links and pricing differences are available on every model page. Downloadable model files can still have restrictions on commercial use.
LOOK BEYOND ONE NUMBER
| Measure | Claude Opus 5.5max | GPT-6 Solmax | DeepSeek V4.1 Flashmax |
|---|---|---|---|
| Agent Fit | 79.4 | 74.7 | 78.9 |
| Practical Value | 60.7 | 62.2 | 83.2 |
| Evidence Consensus | Insufficient evidence | Insufficient evidence | Insufficient evidence |
| Workload cost | $80 | $40 | $1.45 |
| Context window | 1.00M | 1.05M | 1.05M |
| Tool calling | Supported | Supported | Supported |
| Reasoning | 92.2 | 88.7 | 86.7 |
| Coding | 89.3 | 81.8 | 80.0 |
| Agentic coding | 71.7 | 52.9 | 77.3 |
| Mathematics | 97.1 | 96.4 | 93.3 |
| Data analysis | 80.3 | 81.2 | 79.3 |
| Language | 86.3 | 85.3 | 81.2 |
| Instruction following | 65.7 | 68.6 | 70.0 |
| Input / output per 1M | $4 / $20 | $2 / $10 | $0.045 / $0.5 |