This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Per CCTV, the 2026 World Internet Conference Wuzhen Summit opened Nov 3 under the theme "Open Source, Open Sharing." Unlike past years, "open source" isn't a side topic - it's the core: the country's first open-source-themed ecosystem pavilion opened Nov 1, and the main forum has a dedicated "AI Wuzhen Dialogue." Official figures:
The biggest draw is the twin-flagship face-off: media report "Tencent is open-sourcing the Hy4 GA at the summit, colliding with Alibaba's Qwen4." To be precise - as of the evening of Nov 3, neither has released GA weights or scores (DataLearner still marks Hy4 "not yet released"). The two hands:
Take: a month ago the industry debated whether open source was "ebbing"; today the official side puts 10 billion downloads on the table and two flagships in the same hall - the November domestic-model showdown starts at Wuzhen.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| SenseNova 6.8 Flash GA (SenseTime) | 11-2 | Multimodal agent model | Three pillars: long-term memory (remembers why a customer declined across tasks and sessions, proactively re-follows-up), proactive interaction (flags affected items when relevant info changes), design delivery (dynamic PPTs/posters/Agentic 3D - agent + Blender multi-turn output as editable files, Figma-ready). Vendor claims "beyond some international frontier models" on agent collaboration, data analysis, and presentation generation, but no comparison targets or eval details - unreplicated. |
| Hyra 2.0 (Tencent Hunyuan) | 11-2 | AI research system | Shifts from "solve problems with AI" to "general agents developing science models and tools": an Overall/Mission/Worker three-tier architecture, WeCom and repo integration, reusable skills. Vendor first results: 1.96M-param HyInv protein design 57x faster than ProteinMPNN; HyQEC ~48% more efficient quantum error-correction; Xiangshan processor +5.79% per-GHz. Unreplicated. |
| DeepSeek "Creator mode" | 11-3 | Agent product capability | One sentence turns an idea into a runnable AI plugin: the agent writes, installs, and runs the plugin live, on the Cordis "everything-is-a-plugin" architecture. Official demo: ~5.5 minutes to build, install, and verify a floating pomodoro plugin. Vendor demo. |
| Mol-1 (Molence Bio, open code + weights) | 11-3 | Biomolecular foundation model | One model covers antibody-antigen, protein-peptide, protein-small-molecule, and molecular-glue complex prediction. Vendor: FoldBench-AB 75.0% (near Isomorphic's IsoDDE 75.6%), 88.3% with epitope constraints; large-complex modeling beats Anthropic's published results in most tests. GitHub open, online trial; third-party replication pending. |
| XuanYuan 4.5 (Du Xiaoman x HKUST) | 11-2 | Finance LLM | Released during Hong Kong FinTech Week, aimed at financial knowledge, complex business understanding, and finance-scenario deployment - another "big-company + university" combo in vertical LLMs. |
| Strata 0.2.10 (open local tool) | 11-3 | Local LLM runner | Runs 128K context in 8GB VRAM, aimed at low-spec hardware long-text; a practical engineering piece for local-deployment players. |
Today's keyword is "putting it on the table." Wuzhen put 10 billion downloads and 671 companies on the table, the four giants put training-data rules on the table, Anthropic put its pre-IPO books on the table - the race is moving from hidden benchmark battles to open ledgers. Next, watch when the Hy4, Qwen4, and Kimi weight cards actually land.
(Data from official announcements and public reports, 2026-11-03; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa