AI Model Release Tracker

• Official release introduces DeepSeek-V4-Flash-Vision-Exp as the first experimental multimodal V4 model with native visual modules, delivering gains...

J-Zero jointly trains a Challenger, Solver, and Judge in a self-play loop, delivering 8.0-point average gains on unverifiable tasks and 4.2 points on...

Two new arXiv works highlight the shift from static vision benchmarks to closed-loop, embodied control.

• EASEL benchmark tests dexterous visual tool...

GPT-5.5 posts the higher aggregate public score at 72.92 versus 51.55, with non-overlapping 90% intervals.

• Most category averages remain...

Sora 2 Pro offers up to 1024p resolution and enhanced controls for maximum creative freedom and visual quality on demanding projects. It's built...

Opus 5 posts 43.3% on Frontier-Bench v0.1, more than doubling Opus 4.8’s 18.7% and topping Fable 5’s 33.7% at identical $5/$25 pricing.

• Early users...

Newly released models dominating the Terminal-Bench 2.1 leaderboard may reflect benchmark-specific optimization rather than genuine frontier terminal reasoning. Top 20 models were retested on TB-fn, which reformulates the same 89 tasks.

• Licensing edge: Stability's Community License permits commercial use for organizations under $1M revenue, unlike non-commercial FLUX.1 [dev]
• Real...

• Self-improvement loop: Sonnet 5 read research, generated ideas, created ~2,400 examples, and iterated to refine an early Opus 4.8 version.
  -...

• MoE Scale: 770B total parameters with 49B active per request
• Target Domains: Software engineering, research, financial analysis workloads
  -...

Video generators create visually convincing clips yet fail to reproduce the true distribution of physical outcomes when rolled out repeatedly.

-...

GPT-5.6 Sol leads Big Finance Bench at 0.530 on complex financial reasoning tasks, ahead of two other OpenAI entries. All share 1.1M context windows with API costs ranging $0.20–$5.00.

公开资料中两款模型无共同基准结果，无法得出质量优劣结论。只能参考GPT-4.1已披露的单项指标（如Coding 54.6、Knowledge 66.3）和成本数据（1K输入+500输出约$0.006），Lyra Mini多类数据未公布或缺失API费率。

MiniMax H3 generates AI videos from text, images, and reference clips at up to 2K with synchronized sound. The browser demo at https://minimax-h3.im/ enables fast experiments with short-form concepts.

Meta's SIGReg method for LeVJEPA delivers a 20x FLOP efficiency improvement over VJEPA1/2 pretraining by using only a simple sigreg + prediction loss,...

Tencent released the open-source Hy3 model with 770B parameters on Aug 28, 2026, targeting coding, research, and financial analysis. No benchmarks or deployment details were provided in the announcement.

• Ox Alpha identification: Secretly released via OpenCode/OpenRouter before Z.ai confirmed it as GLM-5.3 Flash
• Leaderboard result: 4th on AI Arena...

Could a specialized LLM meaningfully cut documentation workload for knee osteoarthritis and osteoporosis in Indonesian primary care?

• Workload...

Gemini 3.5 Flash posts a higher public score estimate (64.73 vs 51.07) yet the 90% intervals overlap, so the difference is only a lead, not a settled...

Interfaze Beta leads the VoxPopuli WER benchmark at 2.4%, the sole confirmed result among evaluated models. This standardized word-error metric isolates speech-recognition gains without relying on general rankings.