Open-Weight AI Models Compared, October 2026: Size, Context, License and How Far Behind They Are

Open-weight models, whose trained weights anyone can download, now cover most of what a company needs from AI at a fraction of the closed-model price. They also come with very different licenses, sizes and hardware demands. This page compares the main families as of late October 2026.

Read this first. These specs come from vendor model cards as relayed by third-party roundups, not from our own testing, and the roundups disagree in places. Where they disagree, the table says "reported" or leaves the field open. Always confirm the license on the model's official page before shipping a product. For background on what "open" means, see our explainer on open-weight models, and for the closed side, our frontier model comparison.

The Models Side by Side

"Total / active" refers to mixture-of-experts models, which hold many parameters but use only a fraction for each token. Active parameters drive speed and cost; total parameters drive the memory you need to host the model.

Index scores are Artificial Analysis Intelligence Index values as reported by Fello AI and other roundups in October, which name MiMo-V2.6-Pro (46.3) and DeepSeek V4.1-Flash (45.1) as the leading open-weight models. Other scores are approximate and come from secondary roundups. Llama 4 Scout is an older release, shown because it remains widely deployed and offers the longest context.

Some entries are unsettled. Reports say DeepSeek's V4 Pro has been routed to V4.1-Flash since mid-September, so check which one you are getting before you benchmark. Mistral says Large 4 is coming, but at the time of writing its weights were not downloadable and its license was unpublished. Meta's Muse Spark Open is reported to grant broad rights subject to a usage policy, so it is not unrestricted.

Cost: Open Models Can Be Very Cheap to Rent

Hosted prices vary by provider and change quickly. Reported examples from October roundups include roughly $0.435 / $0.87 per million input and output tokens for MiMo-V2.6-Pro, $0.15 / $0.60 for DeepSeek V4.1-Flash off-peak (about twice that at peak), and $3 / $15 for Kimi K3. By comparison, the closed mid-tier models list at $2 / $10, and the top tiers at $4 / $20 to $10 / $50.

Hosting the weights yourself is a different calculation. Big mixture-of-experts checkpoints are large (Kimi K3's is reported at roughly 1.4 to 1.6 TB), so running them takes multi-GPU servers. The dense 27B and 31B models, Qwen 3.8 and Gemma 4, are the ones most teams can run on a single high-end GPU, especially when quantized.

How Far Behind Are They?

On the Artificial Analysis index reported by Fello AI, the best open-weight score is 46.3, against 57.6 for Claude Opus 5.5 and 51.8 to 53.4 for the other closed models we compared. That is roughly an 11-point gap to the top and a 5-to-7-point gap to the middle of the closed pack. Moonshot's Kimi K3.1 Pro scored 55.1 on the same list but is not among the open-weight leaders Fello names, so it is excluded here.

A composite index hides the shape of the gap, and we have not tested where it is widest. Our earlier open-weight explainer noted that frontier labs still tend to lead on the hardest reasoning and agentic tasks.

Choosing a License Matters as Much as Choosing a Model

• MIT and Apache 2.0 are the permissive ones. You can use, modify and ship the model in a commercial product without asking, subject to keeping notices. The MiMo, DeepSeek, Qwen 27B, Gemma and Mistral Large 3 entries above are reported under these.
• Custom licenses (reported for Kimi K3 and GLM-5.3, and Meta's Muse and Llama licenses) can add usage policies, attribution rules or user-count thresholds. Llama's community license, for example, requires a separate agreement above 700 million monthly active users.
• Larger siblings can differ. Reports say some of Qwen's bigger checkpoints use a custom license while the 27B is Apache 2.0, so a license on one size does not cover another.

How to Choose

This section is analysis, not reporting.

• Need to ship a product with no legal friction: start with the permissive-license list, then run your own tasks on two or three of them.
• Need to keep data in-house: the dense 27B to 31B models are the most practical to self-host. Move up to the mixture-of-experts models only if you have the GPUs.
• Want frontier-like quality cheaply and can use a hosted API: the 1M-context MiMo and DeepSeek models are the reported leaders, at a small fraction of closed-model prices. Check the data-handling terms of whichever provider hosts them.
• Need the hardest reasoning or agent reliability: closed models still lead the independent index in this snapshot. If the task justifies the cost, test them side by side with an open candidate and let the results decide.

What We Cannot Conclude

We did not test these models. Parameter counts, context windows and licenses are as relayed by third-party roundups and in some cases contradict each other, which is why several cells say "reported" or "unconfirmed." The benchmark numbers come from a secondary summary of Artificial Analysis. Prices vary by host. Check the sources before you decide.

Why It Matters

• Open weights are close enough for many jobs. An index gap of about 5 to 11 points still leaves open models competitive for coding and everyday tasks.
• The price difference is large. Several hosted open models are reported at under a dollar per million input tokens.
• License is a product decision. Two models with similar scores can differ in whether you may ship them commercially.
• Specs are still settling. Several of the details above changed or were disputed within the last month.