DeepSeek shipped new open weights on Oct. 171. The model, V4.1-Flash-1017, updates the version it launched in September. It is DeepSeek's fifth flagship release since April. Moonshot, Zhipu AI and Alibaba's Qwen team have each gone two months or more since their last one.
A tip for this piece claimed Qwen's next flagship weights were still pending in August. The record disagrees. Alibaba released Qwen3.8's open weights on Aug. 12 and 14, after months of keeping its top models closed2.
That claim holds for the model after this one. Qwen 4, announced Sept. 22, stays in training. Liu Dayiheng, Qwen's project lead, called the release "very soon"3. Five weeks on, all four promised tiers remained in Alibaba's lab.
The score since February
Count every flagship open-weight release from DeepSeek, Moonshot, Zhipu AI, MiniMax and Qwen between January and October. The total comes to fourteen. Seven of the ten months carried at least one. Three of those months, April, June and August, carried three releases apiece.
DeepSeek carries 5 of China's 14 flagship releases since February
Flagship open-weight releases by month, five labs, February through October 2026
Each bar is one month with at least one release; a color is one lab, and the bar's height is how many labs shipped that month.
The chart counts only the seven months that carried a release. MiniMax's Sept. 27 preview stayed off Hugging Face, so it is left out.
Sources: The Decoder [1]; Dataconomy [2]; Yotta Labs [4]; Tech Times [5]; Miraflow [6]; MarkTechPost [7]; Hugging Face [9]; CoderSera [11]; DataNorth AI [12]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | DeepSeek | Moonshot | Zhipu AI | MiniMax | Qwen |
|---|---|---|---|---|---|
| Feb. | 1 | ||||
| Apr. | 1 | 1 | 1 | ||
| Jun. | 1 | 1 | 1 | ||
| Jul. | 1 | 1 | |||
| Aug. | 1 | 1 | 1 | ||
| Sep. | 1 | ||||
| Oct. | 1 |
DeepSeek shipped five times in that span: April, July, August, September and October14. Zhipu shipped four times, on a cadence close to two months: February, April, June and August. Moonshot shipped three times by July, and then it stopped.
Where the other four stand
Moonshot's last flagship was Kimi K3, whose weights landed Jul. 275. Eighty-three days had passed by Oct. 18, the same number still on its card. Online trackers describe a K3.1 in testing. Moonshot has kept its model card and its repository private so far8.
Zhipu AI ships on a tighter cadence. GLM-5 arrived in February. Three more followed roughly every two months: April brought 5.1, June brought 5.2, and Aug. 25 brought 5.3910. GLM-5.4 falls due next, between Oct. 8 and Nov. 9.
As of Oct. 18, Zhipu had kept GLM-5.4 back, though its own pace still leaves three weeks.
MiniMax last opened weights for M3 on Jun. 211. Its newest model, M3.1-Flash-Preview, arrived Sept. 27 with a different shape. The model runs only inside MiniMax's own coding tool, with its weights kept off Hugging Face entirely12.
MiniMax charges for that tool by the month: $22, $55 or $132, depending on the tier12. Qwen followed the same pattern earlier in the year. Alibaba kept Qwen3.5 and Qwen3.6's biggest models behind an API for seven months before Qwen3.8 opened one2.
Moonshot's flagship gap reaches 83 days, MiniMax's 138
Each lab's most recent flagship open-weight release, as of Oct. 18, 2026
Each point marks a lab's last flagship release; the note under each point counts the days back to Oct. 18, 2026.
Sources: The Decoder [1]; Dataconomy [2]; Tech Times [5]; Hugging Face [9]; CoderSera [11]. Chart by The AILately.com Desk.
The numbers behind this chart
| Date | Event |
|---|---|
| June 2, 2026 | MiniMax. 138 days ago: M3 |
| July 27, 2026 | Moonshot. 83 days ago: Kimi K3 |
| Aug. 14, 2026 | Qwen. 65 days ago: Qwen3.8-27B |
| Aug. 25, 2026 | Zhipu AI. 54 days ago: GLM-5.3 |
| Oct. 17, 2026 | DeepSeek. 1 day ago: V4.1-Flash-1017 |
Weighing four reasons
Four explanations compete for why four labs slowed while one kept its pace. The newest is China's own export plan. On Oct. 16, the commerce and science ministries proposed new registration rules for AI weights. Any model above 2 trillion parameters or 10^26 FLOPs would need one, though publishing it as open source wins a faster lane13.
Dean Ball, OpenAI's head of strategic futures, said the draft creates a choice Beijing can exercise later. "This move does not mean 'China is shutting down open-weight models,' but it does indicate the Chinese government establishing the optionality to shut down future open-weight models at will," he said13.
The timing argues against his read as the sole cause. Moonshot, Zhipu and MiniMax had each gone quiet in July or August, weeks before Beijing's proposal existed.
A second explanation reaches back further. Anthropic says it traced roughly 190 million exchanges to seven China-based labs harvesting Claude's reasoning traces14. It rolled out defenses for accounts opened after Aug. 31, cutting what a distiller can extract by 90 to 99 percent.
Zhipu's last flagship fell six days before that date. Moonshot's and MiniMax's gaps opened weeks earlier, which weakens a clean cause and effect.
Compute is a third reason. Washington's chip export rules still cap the Nvidia processors China's labs can buy. A January 2026 rule loosened the limit on H200s, but the cap remains15. Training a flagship model takes more compute than training a point release, and labs working under that cap may simply need longer between releases.
Nathan Lambert tracks open models at Interconnects. He argues speed itself, more than any explanation above, is what kept Chinese labs at the frontier all year. "The time to release for Z.ai is likely days, not months as with OpenAI or Anthropic," he wrote16. By his account, a slower Zhipu or Moonshot gives up the one edge that let them keep pace.
Alex Colville, an analyst with the Australian Strategic Policy Institute's cyber, technology and security program, expects the open releases to keep coming, selectively. "Stakeholders around the world can expect China to continue open-sourcing a wide variety of items that are used to build AI systems," he wrote17. A frontier model's full weights, he added, are the asset Beijing is likeliest to keep closed.
What to do
Plan a launch around DeepSeek's pace. It is the one lab that has shipped something new nearly every month since April. Treat Moonshot's and MiniMax's calendars as open questions instead.
Check the license on each model before you build a product on it. Qwen3.8's large model carries a custom Qwen3.8-Max license, while its 27-billion-parameter sibling carries Apache 2.0, a far more permissive one. The two halves of one release can carry two different sets of rules.
Mark Nov. 9 on your own calendar. If Zhipu's two-month cadence holds, GLM-5.4 lands by then. A miss would raise the same question this piece asked about Qwen in August.
By the numbers
- 14: flagship open-weight releases counted across five labs, February through October 2026124567911.
- 5: of those fourteen came from DeepSeek alone, one in each of five separate months14.
- 83 days: the gap between Moonshot's last flagship, Kimi K3, and Oct. 185.
- MiniMax's gap runs longer: 138 days since its last open-weight flagship, M3, shipped Jun. 211.
- 90 to 99 percent: the cut in distillation harvest Anthropic reports after its August defenses14.
- MiniMax charges $22 to $132 a month for the coding tool that holds its newest, closed-weight model12.
- 2 trillion parameters: the size that triggers registration under Beijing's Oct. 16 export draft13.
What to watch
Watch Nov. 9 for Zhipu's next flagship. Alibaba's own channels are the place to check for a ship date on Qwen 4, now five weeks past Liu Dayiheng's "very soon"3. Beijing's comment window stays open through Nov. 15; watch which labs argue loudest for the open-source fast lane13.
Sources
- Staff, "Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs," The Decoder, Oct. 17, 2026, https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/
- Staff, "Alibaba Unveils Open-source Qwen3.8-Max AI Model," Dataconomy, Aug. 3, 2026, https://dataconomy.com/2026/08/03/qwen3-8-max-ai-model/
- Staff, "Alibaba at Apsara 2026: Qwen 4 in Training, 10T Models Next," NYU Shanghai RITS, Sept. 22, 2026, https://rits.shanghai.nyu.edu/ai/alibaba-at-apsara-2026-qwen-4-in-training-10t-models-next/
- Staff, "DeepSeek V4: Models, Specs, Pricing, and What's Current (2026)," Yotta Labs, Apr. 27, 2026, https://www.yottalabs.ai/post/deepseek-v4-release-date-specs-how-to-access-2026
- Staff, "Kimi K3 Open Weights Arrive Sunday: Self-Hosting Cuts China Data Risk the API Never Can," Tech Times, Jul. 25, 2026, https://www.techtimes.com/articles/321551/20260725/kimi-k3-open-weights-arrive-sunday-self-hosting-cuts-china-data-risk-api-never-can.htm
- Staff, "Kimi K2.6 Explained: Moonshot AI's Open-Source Model That Ties GPT-5.5 on Coding (April 2026)," Miraflow, Apr. 21, 2026, https://miraflow.ai/blog/kimi-k2-6-explained-moonshot-ai-open-source-model-ties-gpt-5-5-coding
- Staff, "Moonshot AI Releases Kimi K2.7-Code: a Coding Model Reporting +21.8% on Kimi Code Bench v2 Over K2.6," MarkTechPost, Jun. 12, 2026, https://www.marktechpost.com/2026/06/12/moonshot-ai-releases-kimi-k2-7-code-a-coding-model-reporting-21-8-on-kimi-code-bench-v2-over-k2-6/
- Staff, "Kimi K3.1: The Name, and What Moonshot Actually Ships," OrcaRouter, Oct. 13, 2026, https://www.orcarouter.ai/blog/kimi-k3-1
- Maxime Labonne, "GLM-5: China's First Public AI Company Ships a Frontier Model," Hugging Face, Feb. 12, 2026, https://huggingface.co/blog/mlabonne/glm-5
- Staff, "GLM 5.4: Does the Model Exist, and What Z.ai Ships Today," OrcaRouter, Oct. 13, 2026, https://www.orcarouter.ai/blog/glm-5-4
- Staff, "MiniMax M3 Release Date: June 1, 2026 (Is M3.1 Out?)," CoderSera, Jun. 1, 2026, https://codersera.com/blog/minimax-m3-release-date-whats-new-2026/
- Staff, "MiniMax released M3.1-Flash-Preview," DataNorth AI, Sept. 27, 2026, https://datanorth.ai/news/minimax-releases-m3-1-flash-preview
- Staff, "China plans to tighten export controls on open-weight models and 23 key AI technologies," The Decoder, Oct. 16, 2026, https://the-decoder.com/china-plans-to-tighten-export-controls-on-open-weight-models-and-23-key-ai-technologies/
- Staff, "Chinese AI labs secretly used millions of Claude exchanges to train their models, Anthropic says," CNBC, Sept. 11, 2026, https://www.cnbc.com/2026/09/11/chinese-ai-labs-moonshot-deepseek-alibaba-anthropic.html
- Staff, "The U.S. banned Nvidia's best chips from going to China. Now it's trying to close a crucial loophole," CNBC, Aug. 19, 2026, https://www.cnbc.com/2026/08/19/china-ai-nvidia-chips-us-export-controls.html
- Nathan Lambert, "GLM-5.3: How Chinese labs keep stride with the frontier," Interconnects, Aug. 14, 2026, https://www.interconnects.ai/p/glm-53-how-chinese-labs-keep-stride
- Alex Colville, "Some Chinese AI will go closed-source, but not all," The Strategist (ASPI), Jul. 22, 2026, https://www.aspistrategist.org.au/some-chinese-ai-will-go-closed-source-but-not-all/