Twenty consequential open-weight launches by late August — April was the densest month, and million-token context went mainstream

Asked:assemble a release time line of major open source LLMs being released in 2026

A curated set of 20 major open-source and open-weight LLM releases from 2 February to 26 August 2026 — one event per model-family launch, with parameters, context window and license. 12 of the 20 ship under permissive open-source licenses (MIT or Apache 2.0); the other 8 are downloadable-weight models under custom or restricted terms, which does not satisfy the OSI open-source definition. 9 of the 20 advertise context windows of a million tokens or more.

permissive open source (MIT / Apache 2.0) open weights, custom or restricted terms open model, terms unverified ◇ month-level or inferred date · ● exact date

April is the densest month, with 5 flagship launches in 22 days (GLM-5.1, MiniMax M2.7, Qwen 3.6, Kimi K2.6, Mistral Medium 3.5) after the February ramp of Kimi K2.5, GLM-5 and Qwen 3.5 — z.ai

Million-token context became a frontier norm: 9 of the 20 launches claim ≥1M tokens, including MiniMax M3, the first open-weight model pairing frontier coding, 1M context and native multimodality — fazm.ai

Kimi K3 (2.8T parameters, 28 corroborating sources) is reported to match or surpass Anthropic's Fable 5 on key benchmarks at roughly half the cost — a vendor-favourable claim from press synthesis, not an independent audit — edgen.tech

DeepSeek V4 carries only a year-level date in the shaped record; the April 24 position is inferred from the publication date of its first-party Hugging Face model card — huggingface.co

Mixture-of-experts dominates: nearly every flagship here activates a small fraction of total parameters (e.g. GLM-5 at 744B total / 44B active; Inkling at 975B / 41B active) — best-ai.news

The 20 releases, in order

Date	Developer	Model	Parameters	Context	License	Evidence	Source

Curated timeline of major open-source and open-weight LLM releases, 20 rows, current through 2026-08-31. “Major” means a new flagship or model-family generation, or a consequential broadly released open-weight model — not every checkpoint, quantization, fine-tune or patch; the set is selective, not exhaustive. Built from broad 2026 web coverage plus first-party developer and model-card searches; rows marked multi-source synthesis carry their source count, and vendor benchmark figures are vendor-reported, not independently verified. DeepSeek V4's date is inferred from its first-party model-card publication (2026-04-24). Significance text is abridged in the table for space.