Six free models currently score at or near the top of our quality index, and all six cost nothing per million input or output tokens. Since price no longer separates them, the useful question is which model fits the work you actually do. This guide sorts them by architecture, input types, and intended role, using only what the vendors have published and what our scores show.
How to Read the Scores
Four of these models score 100 on our quality index. The other two score 98. A two-point gap at this level is small, so do not treat it as a ranking that settles every decision. Treat the score as a floor that tells you all six can handle serious work. Choose among them based on modality, deployment footprint, and the specific job.
Most of the list uses mixture-of-experts designs, where only a fraction of the total parameters activate per token. The active parameter count is the better guide to inference cost and speed if you plan to self-host the open weights. The total count tells you more about memory requirements.
NVIDIA Nemotron 3.5 Super VL
Nemotron 3.5 Super VL is an open multimodal mixture-of-experts model with 12B active parameters out of 120B total. It accepts text, images, documents, GUI screenshots, and video. It scores 100.
This is the model to pick when your inputs are messy and visual. The explicit support for documents and GUI screenshots makes it a strong fit for document extraction pipelines, screen-reading agents, and any workflow where a model needs to understand what is on a user interface. If you are building a computer-use agent or processing scanned forms, start here. The 12B active footprint keeps per-token compute moderate relative to its total size.
NVIDIA Nemotron 3 Nano Omni
Nemotron 3 Nano Omni is a 30B-A3B open multimodal model, meaning roughly 3B parameters activate out of 30B total. It accepts text, image, video, and additional inputs. NVIDIA designed it to work as a perception and context sub-agent in enterprise agent systems. It scores 100.
The intended role matters here. This model is built to sit underneath a larger orchestrating model, handling perception tasks and feeding context upward. If you run a multi-agent system and need a cheap, fast component to interpret images or video before a planner model acts, Nano Omni is the right choice. Its small active count makes it well suited to high-volume calls. For a single model that does everything in one pass, Super VL is the better pick.
Google Gemma 4 31B
Gemma 4 31B Instruct is a 30.7B dense multimodal model that takes text and image input and produces text output. It has a 256K token context window, a configurable thinking mode for reasoning, and native function calling. It scores 100.
Pick Gemma 4 31B for long-context work and tool use. The 256K window handles large codebases, long contracts, and extended research documents in a single prompt. Native function calling makes it a dependable base for agents that call APIs. The configurable thinking mode lets you turn on reasoning for hard problems and turn it off when you want faster, cheaper responses. Because it is dense rather than mixture-of-experts, every parameter runs on every token, so expect higher compute per token than the MoE options on this list.
Google Gemma 4 26B A4B
Gemma 4 26B A4B IT is an instruction-tuned mixture-of-experts model with 25.2B total parameters and 3.8B active per token. Google describes it as delivering near-31B quality. It scores 100.
This is the efficient sibling of the 31B model. If you like the Gemma 4 family but need lower latency or plan to self-host on limited hardware, the A4B variant gives you similar quality at a fraction of the active compute. Use it for high-throughput chat, classification, and summarization where the dense 31B model would be more than you need.
Dots Studio Dots3-Note Preview
Dots3-Note Preview is an open-weight mixture-of-experts model with 16B active parameters out of 280B total. It is the lightest model in the Dots 3 family. It scores 98.
It carries the largest total parameter count on this list and the highest active count, which suggests a large knowledge capacity behind each response. Note the "Preview" label. Preview releases can change, so it is a reasonable candidate for evaluation and experimentation, but you should test it carefully before building production workflows around its behavior. Run it against your own prompts alongside one of the 100-scoring models before committing.
Thinking Machines Inkling Small
Inkling Small is an open-weight multimodal mixture-of-experts model from Thinking Machines Lab, with 12B active parameters out of 276B total. It is positioned as the smaller, more efficient member of its family. It scores 98.
Inkling Small pairs a large total parameter pool with the same 12B active count as Nemotron Super VL. It is a sensible alternative to test if you want a multimodal model from a different vendor, which helps when you are comparing outputs or reducing dependence on a single provider.
Which One to Choose
For visual documents, screenshots, and GUI agents, use Nemotron 3.5 Super VL. For a lightweight perception component inside a larger agent system, use Nemotron 3 Nano Omni. For long documents and function calling, use Gemma 4 31B, and switch to Gemma 4 26B A4B when speed and hardware cost matter more. Keep Dots3-Note Preview and Inkling Small in your evaluation set as second opinions. Because all six are free, the cheapest way to decide is to run your twenty most common prompts through two or three of them and keep the one whose outputs need the least fixing.