Breaking

Qwen and Z.ai Open-Weight Releases Lead Hugging Face After Late-August Drops

September 1, 2026 at 09:17 EDT

Foundation Models
Open Source
Software Dev & Coding

Alibaba’s Qwen team and Z.ai have packed Hugging Face with newly released open-weight checkpoints, led by Qwen3.8-Flash-Next, a multimodal mixture-of-experts model that previews the Qwen4 architecture and activates only 6 billion of 125 billion parameters per token. Beside it sit Z.ai’s GLM-5.3-Flash and flagship GLM-5.3, Qwen’s dense Qwen3.8-27B, and Unsloth’s Qwen3.8-Flash-Next GGUF conversions that make the Qwen4 preview runnable on local machines.

September 1, 2026 · Qwen · Z.ai

Open weights, two labs, five checkpoints

Qwen3.8-Flash-Next previews Qwen4 at 6B active parameters. Z.ai’s GLM-5.3-Flash and GLM-5.3 join Qwen3.8-27B and Unsloth GGUFs at the center of Hugging Face traffic.

Training cost

Flash-Next vs Qwen3.7-Plus

1/9

Qwen3.8-Flash-Next

1×

Qwen3.7-Plus

Qwen says Flash-Next trained at about one-ninth the cost of Qwen3.7-Plus, with stronger coding and office-agent scores.

Flash-Next active / token

8.6×

Prefill vs Plus at 1M

GLM native context

~5M

Qwen3.8-27B downloads

Active parameters

Compute used per token

Flash-Next

18B

GLM Flash

27B

Qwen3.8-27B

40B

GLM-5.3

What is working

Qwen3.8-27B fits a single consumer GPU under Apache 2.0. 1-bit Flash-Next GGUF loads in about 75GB RAM. GLM-5.3-Flash is MIT-licensed, multimodal, and cheap enough that an eight-hour coding run produced 15 pull requests for about $4.

The catch

Full Flash-Next weights are ~360GB. Cold 262K prefill on a 27B can take 6 minutes 14 seconds to the first token. GLM-5.3-Flash streams at 48.7 tok/s, slower than GLM-5.3. The flagship GLM-5.3 license adds a $10B MaaS security review.

Qwen4 preview

Qwen3.8-Flash-Next is the first open-weight model on the architecture Qwen says will underpin Qwen4.

Gated DeltaNet + Qwen Sparse Attention · 51B n-gram embeddings offloadable to RAM · 262K native context, 1M with YaRN · production API as Qwen3.8-Flash.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.