If you have not been watching Hugging Face, the last six weeks were crowded. Between August 12 and September 22, 2026, Chinese labs released six notable open-weight models, and the pattern across them is as interesting as any single score.
The timeline
| Date | Release | Size | License | Headline |
|---|---|---|---|---|
| Aug 12 | Qwen3.8-Max weights | 2.4T, 95B active | Qwen3.8-Max license | First Max-class Qwen to go open; 27B dense sibling followed Aug 14 (Apache 2.0) |
| Aug 14 | GLM-5.3 | 753B | MIT-style | Same base as GLM-5.2; Terminal-Bench 3.0 4.6 → 28.3 from post-training alone |
| Aug 26 | GLM-5.3-Flash | 320B, 18B active | MIT | Natively multimodal, $0.15 / $0.50, beats GLM-5.2 at about one-tenth the price |
| Aug 26 | Qwen3.8-Flash-Next | 125B + 51B N-gram, 6B active | Qwen community | Qwen4-architecture preview, trained at about 1/9 the cost of Qwen3.7-Plus |
| Sep 10 | DeepSeek V4.1-Flash | 552B, 8B/16B active | MIT | DeepSWE v1.1 74.2, KV cache one quarter the size |
| Sep 22 | MiMo-V2.6-Pro | 1.02T, 42B active | MIT | Omni-modal; Xiaomi reports 46 on the Artificial Analysis index |
Alibaba’s API-only Qwen3.8-Omni-Flash (Sep 18), built for audio-video agents, filled in the multimodal picture without releasing weights.
What the pattern says
- Efficiency beats size. Three of the six releases are “Flash” models — 6B active (Qwen), 18B (GLM) and 8B for input / 16B for output (DeepSeek). All of them lean on hybrid or sparse attention and smaller KV caches to make 1M-token contexts cheap.
- MIT is the default. Four of the six use MIT or MIT-style licenses (GLM-5.3, GLM-5.3-Flash, DeepSeek V4.1-Flash, MiMo-V2.6-Pro); Alibaba uses a community license for Flash-Next, a custom Max license for the 2.4T weights, and Apache 2.0 for the 27B.
- Flash-tier prices start at $0.15 input. GLM-5.3-Flash and Qwen3.8-Flash are $0.15 input, MiMo-V2.6-Pro is $0.435, and DeepSeek V4.1-Flash is $0.15 off-peak ($0.30 at peak).
- Multimodality is becoming the default. GLM-5.3-Flash, DeepSeek V4.1-Flash, Qwen3.8-Flash-Next and MiMo-V2.6-Pro all take images; MiMo and Qwen’s Omni model add audio and video.
- RL scale is the new lever. GLM-5.3 improved 24 points on Terminal-Bench 3.0 with post-training alone, and Xiaomi says its six-day RL run cost about $2.6M for Pro.
Qwen 4 is coming
At Alibaba Cloud’s Apsara Conference on September 22, Alibaba said its next-generation Qwen 4 is in training, with a Qwen 4.5 and Qwen 5 roadmap projected to reach 5 to 10 trillion parameters. It gave no release date, sizes or pricing. Qwen3.8-Flash-Next is the architecture preview, and Alibaba also reported its Qwen3.8-Max lifted its own Artificial Analysis score from 40 to 45 through automated self-improvement — a vendor claim.
Caveats
- Every score is vendor-reported, using different harnesses; cross-vendor numbers are not directly comparable.
- Closed models from Anthropic and OpenAI still lead the hardest long-horizon tasks.
- Self-hosting a 550B-plus model remains a multi-GPU job; the Flash tier is the realistic one.
For the wider model landscape, see the 2026 LLM guide.