35 open-weight text-capable models landed July–mid-October 2026, and the big ones run on 2–6% of their parameters

Asked (summary):

Which open-weights text-inference LLMs were released in July, August, September and October 2026? For each: total parameters, active parameters, the releasing lab or entity, that entity's home country, and whether it accepts image or audio inputs alongside text.

A cross-confirmed inventory of 35 official open-weight releases or separately released size variants from 14 labs across 5 countries — 10 in July, 12 in August, 9 in September, 4 through October 16, 2026. It is a researched inventory, not an exhaustive universe: no global registry of open-weight releases exists, and every row here is supported by at least two independent source hosts. A broader single-source candidate list appears in the appendix and is not counted in the 35.

Every release on a log scale: pale bar is total parameters, solid bar is active per token

total parameters (log scale) active — China active — United States active — United Arab Emirates active — France active — Germany

Hover or tap a row for full details. Dense models show active = total; one active figure is phase-specific (DeepSeek-V4.1-Flash activates 8B during prefill and 16B during decode — plotted at 16B). DeepSeek-V4-Flash-Vision-Exp has no verified active-parameter figure, so only its total bar is drawn.

Moonshot AI's Kimi K3 is the largest release of the window at 2.8T total parameters, yet activates only 104B (3.7%) per token — github.com

China accounts for 17 of the 35 cross-confirmed releases; the United States 9, the UAE 6 (all from one coordinated six-size K2 Horizon drop on September 3) — en.wikipedia.org

Only 5 releases accept audio input, and every one of them also accepts images: Inkling, Inkling-Small, GLM-5.3-Flash, both MiMo-V2.6 sizes and Meta Muse Spark Open — huggingface.co

DeepSeek-V4.1-Flash is the clearest case of phase-specific activation: 552B total, 8B active during prefill and 16B during decode — huggingface.co

All 35 cross-confirmed releases

Date	Model	Total params	Active params	Lab / entity	Country	Image	Audio	Source

Active parameters are the weights engaged per token in mixture-of-experts models; for dense models the whole network is active, so active equals total. Country links go to the entity-country source. "Not verified" means the sources did not state the modality either way.

Appendix: supplementary candidates, not in the headline count

A broader list of 74 plausible releases found by targeted searching. Many have only one independent source host or are naming variants of models above, so they are kept out of the 35-model cross-confirmed total.

Method: cross-confirmed inventory of official open-weight, text-capable foundation-model releases (including separately released size/checkpoint variants) from July 1 to October 16, 2026 — 35 rows, each backed by at least two independent source hosts; compiled as of 2026-10-16. Total parameters are the full weight count; active parameters are weights engaged per token (phase-specific where noted). Third-party fine-tunes, quantizations, intermediate checkpoints, embedding/reranker, image-only and speech-only models are excluded. The 74-row appendix lists single-source or variant candidates and is excluded from headline totals. Entity-country mapping uses normalized entity aliases.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.