Between August 12 and 14, 2026, the open-weight frontier moved twice. Alibaba released the first downloadable Max-class Qwen model, and Z.ai shipped GLM 5.3 with agentic coding numbers strong enough that it deliberately delayed the weights for a safety review (they landed on August 28). Every “best open model” page on the internet now needs a rewrite, and as of this writing, nobody has published a direct comparison of the two.
This is that comparison, grounded entirely in what both vendors have published, updated August 28, 2026 when GLM 5.3’s weights landed on Hugging Face. The short version: GLM 5.3 is the capability story and, since August 28, an availability story too; Qwen3.8 got there first. Which one wins for you now depends on the benchmarks you care about and the license terms you can live with.
On this page · 6 sectionsOpen
- Update, September 2: Qwen upgraded Qwen3.8-Max in place as Qwen3.8-Max-0902 (September 1), at the same $2/$6 price and 1M context, with agentic coding scores roughly doubling (TerminalBench 3.0 11.3 to 29.0) and JobBench up ten points. The comparison below reflects the August snapshots; see the September 2 section for what moved.
- GLM 5.3 (Z.ai, August 14, 2026) is a post-training upgrade on the ~744B-parameter GLM 5.2 MoE base, claiming a 50 percent coding improvement and state-of-the-art open-weight results on Terminal-Bench 3.0.
- Qwen3.8-Max (Alibaba, August 3) is a 2.4T-parameter MoE with ~95B active parameters; its open-weight sibling Qwen3.8-2.4T-A95B landed August 12 as the first downloadable Max-class Qwen.
- Update, August 28: GLM 5.3’s weights are now on Hugging Face under a permissive MIT-style license, ending the safety hold; Qwen’s open checkpoint remains text-only with a 262K context versus the hosted 1M.
- Qwen3.8-Max has published API pricing ($2 per million input tokens, $6 per million output); GLM 5.3’s API is listed as coming soon with pricing not yet confirmed.
- GLM 5.3’s distinctive edge is agentic and cyber capability: 84.5 percent on CyberGym per Z.ai, strong enough that the weights release was deliberately slowed.
- For agentic coding today, GLM 5.3 via its subscription plan is the capability pick; for self-hosting and multimodal API work, Qwen3.8 is the availability pick.
§ 01Update, September 2: Qwen3.8-Max became Qwen3.8-Max-0902
On September 1, 2026, Qwen shipped Qwen3.8-Max-0902, an upgraded snapshot of the same 2.4T-parameter model, further post-trained on coding and collaborative agent work. Nothing about the availability picture changed: same $2 per million input and $6 per million output, same 1M context, same API. What changed is the capability side of this comparison. On Qwen’s own table, agentic terminal coding went from 11.3 to 29.0 on TerminalBench 3.0, black-box software replication from 10.5 to 28.0 on ProgramBench, and professional job tasks from 53.4 to 64.0 on JobBench. The weakest coding rows in the August snapshot, the ones that made GLM 5.3 the clear capability pick for agentic coding, are the rows that moved most.
| Benchmark | Qwen3.8-Max (August) | Qwen3.8-Max-0902 |
|---|---|---|
| TerminalBench 3.0 | 11.3 | 29.0 |
| ProgramBench | 10.5 | 28.0 |
| JobBench | 53.4 | 64.0 |
The verdict below still holds in shape: GLM 5.3 remains the open-weight capability pick with its weights now on Hugging Face, and Qwen remains the availability and multimodal pick. The gap on agentic coding is narrower than it was in August. The full table, its footnotes, and where Claude Opus 5 still leads are in our Qwen3.8-Max-0902 post. The sections that follow are the August 16 comparison with the August 28 licensing update, kept as written.
§ 02The two releases, side by side
| GLM 5.3 | Qwen3.8-Max | |
|---|---|---|
| Vendor, date | Z.ai, Aug 14 | Alibaba, Aug 3 (open weights Aug 12) |
| Architecture | ~744B MoE (GLM 5.2 base) | 2.4T MoE, ~95B active |
| Context | 1M tokens | 1M hosted; 262K open checkpoint |
| Modality | Text | Multimodal hosted; text-only open |
| Open weights | On Hugging Face since Aug 28 (GLM-5.3 License) | Qwen3.8-2.4T-A95B downloadable (reduced) |
| API pricing | Not yet published | $2 / $6 per million tokens |
Two different bets are visible in that table. Z.ai spent its effort on post-training: same base as GLM 5.2, dramatically better behavior on long-horizon agentic tasks. Alibaba spent its effort on scale and openness: a model three times larger by total parameters, and a historic decision to let people download a Max-class checkpoint at all.
§ 03What the benchmarks actually say
Per Z.ai’s launch materials, GLM 5.3 improves on GLM 5.2 by roughly 50 percent on its internal coding benchmark and takes state-of-the-art open-weight positions on Terminal-Bench 3.0 and Agents’ Last Exam (CLI). The most striking number is CyberGym, a vulnerability-discovery benchmark, where Z.ai reports 84.5 percent versus 77.2 for GLM 5.2, ahead of the closed frontier models it compared against. In testing, the model reportedly found 2,436 real vulnerabilities across 269 open-source projects. That is the direct cause of the delayed weights: capability strong enough to warrant a safety hold.
Qwen3.8-Max’s published results position it near the closed frontier on coding and agentic suites, and its open checkpoint arrived with day-zero support in the major serving stacks. The honest caveat on both sides: these are vendor-reported numbers days old, and independent verification takes weeks. Treat every figure here as a claim with a date on it, not settled fact.
§ 04Openness, licensing, and what “open” means this month
Neither release is simply open, and the fine print matters.
GLM 5.2 shipped under MIT. GLM 5.3’s weights landed on August 28, 2026, right at the stated two-week mark, under a custom GLM-5.3 License: MIT-style permissions (use, modify, distribute, sell) with one condition, that Model-as-a-Service operators above 10 billion dollars in trailing-12-month revenue pass Z.AI’s security review before commercial use. For nearly every team that reads as permissive; the carve-out targets hyperscale API resellers.
Qwen’s open checkpoint is downloadable today but reduced: text-only where the hosted Max is multimodal, 262K context where the hosted product offers 1M, and released under Alibaba’s own license terms rather than a permissive MIT-style grant, with revenue-share conditions reported at release. For most teams this changes nothing; for products built directly on the weights, read the license before you build.
§ 05Which one, for what
- Best agentic coding capability right now: GLM 5.3, through the subscription. The Terminal-Bench and agentic numbers are the strongest any open-weight vendor has published, and the 1M context is real today in the hosted product.
- Self-hosting on your own hardware: a real two-way choice since August 28. GLM 5.3’s full-capability weights are on Hugging Face, and Qwen3.8-2.4T-A95B has been downloadable since August 12. Either way, plan for serious multi-GPU infrastructure.
- Multimodal API work: Qwen3.8-Max, the only one of the two with vision and published API pricing.
- Security research and code auditing: GLM 5.3’s CyberGym results make it the one to watch, and the one to handle carefully. A model this good at finding vulnerabilities deserves the least-privilege treatment when you wire it to real systems.
- The wait resolved: GLM 5.3’s weights landed under a permissive license on August 28, at the stated two-week mark, and the self-hosting calculus did change again. That is the pace this category now moves at.
§ 06What this means above the model layer
Here is the pattern worth internalizing: this was one week, and it rearranged the open-weight frontier twice. If your workflow requires re-evaluating your model choice every launch cycle, launch weeks are homework. If you work above the model layer, they are pure upside.
That is the practical argument for super-agent platforms and AI employees: the platform routes to the best available frontier models as they land, and the things you actually depend on (memory that carries between working sessions, permissions, isolation, a task board) don’t churn weekly. When we migrated CellCog to Fable 5 earlier this summer, no user re-evaluated anything; the work just got better. On CellCog that routing is visible as modes rather than model pickers: Agent, Agent Creative, and Agent Team, each at a Flash, Core, or Max tier (Creative starts at Core), so you choose depth per task and the platform chooses engines. The harness underneath has receipts of its own: #1 on Deep Research Bench (July 2026) (context on our benchmarks page), and user-reported coding wins over Claude Code on identical work, detailed with caveats in our August harness rankings. GLM 5.3 and Qwen3.8 will be absorbed the same way across the industry: quietly, by the platforms, while the model-picker debates rage on.
For the rest of the stack this fits into, see our August 2026 harness rankings. For the GLM 5.3 release in depth, continue to what GLM 5.3 means for AI agents.
Q1What is GLM 5.3 and when was it released?
GLM 5.3 is Z.ai’s flagship language model, released August 14, 2026. It keeps the ~744B-parameter mixture-of-experts base of GLM 5.2 and gains capability through scaled post-training, with a claimed 50 percent coding improvement, a 1M-token context window, and state-of-the-art open-weight results on terminal and agentic benchmarks per the launch materials.
Q2What is Qwen3.8-Max and what makes the August release special?
Qwen3.8-Max is Alibaba’s frontier model: a 2.4T-parameter mixture-of-experts with roughly 95B active parameters per token, a 1M-token context on the hosted API, and multimodal input. On August 12, 2026 Alibaba released Qwen3.8-2.4T-A95B, the first Max-class Qwen with downloadable weights, in a reduced form: text-only, 262K native context.
Q3Why are GLM 5.3's weights delayed?
Z.ai held the weights for roughly two weeks of additional safety evaluation because the model showed unusually strong cybersecurity capability, including finding thousands of real vulnerabilities across open-source projects in testing, per its launch materials. The hold resolved on schedule: the weights were published on August 28, 2026.
Q4Which model is cheaper to use?
Qwen3.8-Max is the only one with published API pricing: $2 per million input tokens and $6 per million output. GLM 5.3 is currently available through Z.ai’s coding subscription tiers, with standalone API pricing listed as coming soon; its predecessor GLM 5.2 was priced around $1.40 and $4.40 per million, which is the community’s reference point until real numbers land.
Q5Does the model choice matter if I use an agent platform?
Less than you’d think, and that’s the point. Platforms that route across models pick up frontier gains as they land, without you re-evaluating every launch week. What matters at the platform layer is the architecture around the model: memory, permissions, isolation, and whether the worker persists between sessions.