Open-Weight Models in 2026: How Many Launch Open?
Few launches since July said their weights were public, and none came from the three biggest US labs. Open weights now arrive mostly as specialist parts.
The signal
- LaunchesClaude Sonnet 5.5 (28 Sept 2026): Faster and cheaper than Sonnet 5 on most work, with better agentic coding scores. Closed weights, $2/$10 per million tokens.
- LaunchesFLUX 3 Action (23 Sept 2026): Open 7B world-action model that predicts robot motion and video together. Released under the FLUX Kommunity License.
- LaunchesTencent's Hy4 Preview (28 Aug 2026, per launch trackers) is the only open-weight launch in the feed before 23 Sep.
Of 51 named model launches since 6 Jul 2026, 22 stated their weights: 18 closed, 4 open. All 16 from OpenAI, Google and Anthropic were closed.
Named model launches by stated weight status, 6 Jul 2026 – 28 Sept 2026. Not stated = neither the lab post nor the tracker said.
Launch-feed rows as listed; a post naming several variants counts once. 29 of 51 rows do not state weights; all of them are tracker rows (6 Jul 2026 – 22 Sep) that rarely record it, so the open count is a floor. In the lab-post sweep from 23 Sep (7 rows) every row states it: 3 open, 4 closed. NV-Reason-CT rests on a single aggregator listing. DeepSeek V4.1 Flash has conflicting reports, so it is left as not stated.
Show the data (5 rows)
| Weight status | Launches | Labs |
|---|---|---|
| Open weights | 4 | Tencent, Black Forest Labs, NVIDIA |
| Closed: OpenAI, Google, Anthropic | 16 | OpenAI, Google, Anthropic |
| Closed: other labs | 2 | Alibaba Qwen, Meta |
| Not stated: labs in China | 17 | Tencent, Moonshot, Alibaba Qwen, ByteDance, Z.ai, DeepSeek |
| Not stated: other labs | 12 | Meta, xAI, NVIDIA, Runway |
The Mode read
Anthropic shipped Claude Sonnet 5.5 today with closed weights, like every OpenAI, Google and Anthropic launch since July. On 23 Sept 2026, 5 days earlier, Black Forest Labs and NVIDIA released 3 open models. So how often do labs put weights in the open?
Rarely, where we can tell. Of 51 named launches since 6 Jul 2026, 22 stated their weight status: 18 closed and 4 open. All 16 launches from OpenAI, Google and Anthropic were closed. The other 29 did not say, and 17 of those came from labs in China. The open ones are telling: 3 of the 4 are specialist models, for robot actions, speaker diarization and CT scans, not general assistants.
The general frontier is a closed product. Open weights now arrive as parts: small, focused models that a team can run on its own hardware and tune on its own data. A part is worth most to whoever controls where it runs and what data it sees. That is usually a company already inside the customer's building, such as a hospital, a factory or a robot fleet, not the lab that trained it.
The frontier stays closed. Open weights arrive as parts, and parts go to whoever owns the assembly line.
What we would back: Vertical products that run open specialist models inside the customer’s own walls: a hospital, a factory, a robot fleet.