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.

Two days after Google's world model Genie 4, Tencent Hunyuan dropped a side-by-side comparison video: its own open model Hy-Nomas recreating all 8 of Genie 4's official demos one by one. The same day Tencent delivered the technical report (arXiv 2610.36729) and open-sourced the benchmark Hy-Nomas-Bench.
The cards on the table: Hy-Nomas was actually open-sourced back on Oct 21 - this round adds the "manual + exam" and rides the buzz. Its architecture is a two-piece kit:
Media's per-demo comparison: interactivity and content completeness overall near Genie 4, with bright spots (the Japanese station scene reflects the shooter in the train body - deep understanding of light and object relations). The gaps are clear: Genie 4 has synced audio (the announcer actually speaks), Hy-Nomas only outputs visuals; Genie 4's camera moves and expressions are more natural; Tencent's style consistency is weaker; and some object-relation errors (a robotic arm not gripping the pipette, the tube floating as if stuck to the arm).
Take: Genie 4 is still research preview, closed, invite-only; Hy-Nomas is download-anything, play-anywhere. The real meaning of this comparison is proving that "mature open LLM as director + small video diffusion as camera" can reach ~80% of a closed flagship with open components. In the world-model race, a Chinese team is now keeping pace the open-source way.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| FastH3 v1 (MiniMax x NVIDIA x FastGen, open source) | 10-31 | Video generation acceleration | Pitched as "near-realtime": one Blackwell card generates 15s of 768p video in 13 seconds, up to 14x speedup, text-to-video with synced audio; weights, training code, and deployment guides all open. An earlier community build (8x B300) measured ~8.7s per 10s - faster than playback. Team- and community-reported, not independently replicated. |
| Open-H3-Regenerate-2K (NVIDIA x MIT, open weights) | 10-30 | Video enhancement / regeneration | A MiniMax H3-based regeneration pipeline up to 4K: latent upscale + three-step refine + 2K decode, from 87.26s to 22.45s (~3.9x); blind tests preferred it over the official API 87% of the time. MiniMax H3 hit 24M+ downloads in a month with 300+ derivatives - big players are now building its ecosystem; the open-source flywheel spins. |
| MOSS-Image (OpenMOSS x Fudan NLP x Shanghai Chuangzhi, open source) | 10-31 | Text-to-image | A 5B diffusion model focused on long-text rendering (signs, UI text rarely mangled) and JSON layout control, Apache 2.0, image-to-image, local text edits, 2x upscale, ~18s for 1.5MP on a 5090. Note: Civitai records show a first release Oct 18 - this is reporting spread, not a new launch. |
| Kitesurf (Cloudflare, open source) | 10-31 | Agent browser engine | A lightweight browser for AI agents: runs in Workers' V8 isolates, Rust/WASM, strips tabs/extensions ("human features"), keeps only machine-readable output; screenshot and HTML extraction cost 3-7x less CPU/memory than full Chromium, compatible with Playwright, Puppeteer, and MCP. A low-cost base for scraping and PDF tasks. |
Today's keyword is "open-source catching up." On the most expensive track - world models - Tencent used an 8B model plus an open toolkit to reach 70-80% of Google's flagship. Next week Qwen4, Hy4, and Kimi weights land in a bunch, and the closed-source lead keeps getting eaten, piece by piece, by "modular + open recreation."
(Data from official announcements and public reports, 2026-11-01; 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