大模型网摘Open-Weight Flagships Compared (October 2026): Kimi K3.1 Pro, DeepSeek V4-Pro-0813, GLM-5.3 and Qwen3-Max
The open-weight camp has kept strengthening through October: Moonshot AI's Kimi K3.1 Pro pushes open-weight parameters to the 3T class, DeepSeek's V4-Pro-0813 continues to ship as open snapshots, while GLM-5.3 and Qwen3-Max represent two other tracks. This comparison puts the verified key specs in one place for quick shortlisting.
Kimi K3.1 Pro: 3T-class, 1M context
Released October 23, K3.1 Pro is a post-trained version of Kimi K3 with 3T-class parameters and a 1M-token context; the Highspeed variant trades roughly 2-3x speed, and full weights are promised by November 4. Among third-party-checkable benchmarks, it scored 65.7% on Terminal-Bench 4.0, between GPT-6 Astra (58.2%) and Claude Opus 5.5 (66.4%). See our launch coverage for details.
DeepSeek V4-Pro-0813 and V4.1-Flash-1017
V4-Pro-0813 ships as an open-weight snapshot with 1M context, 384K max output, and off-peak API pricing of $0.66 input (cache miss) and $1.98 output; our open-weights deep dive traces the architecture. V4.1-Flash-1017 is the speed tier of the family (552B MoE base), also with 1M context and vision input. Our dedicated comparison covers how the two tiers differ.
GLM-5.3 and Qwen3-Max: two other tracks
Zhipu's GLM-5.3 launched August 14, with the GLM-5.3-Flash tier following on August 26, both aimed at coding and agentic work. The Qwen3-Max line comes from Alibaba's Qwen team — confirmed at 1T parameters at its September 2025 launch, with Qwen3-Max-Thinking added in January 2026. Parameter details and license terms should be taken from the official pages; we have not audited them line by line, so this is a timeline for reference.
Selection notes
Raw parameter counts only say so much: context length, license terms, inference cost and the surrounding toolchain matter just as much. If self-hosting and toolchain completeness top your list, the DeepSeek line has become a fuller path since DeepInfer went open source; if terminal coding-agent scores matter more, K3.1 Pro and V4.1-Flash-1017 sit in a comparable range — see also our Flash vs GPT-6.1 Sol vs GLM-5.3-Flash comparison. The sensible move is a small in-house benchmark on your own workload before deciding.