New Deepseek, GLM 5.3, Grok 4.6, LTX 2.5, Qwen 3.8, Gemini 3.7: AI NEWS
AI Search showed 24 AI tools in this video on . 15 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-33.
Tools shown in this video, with timestamps
Open-source prompt-based video editor, a 16B multimodal autoregressive diffusion transformer generating 720p at 30+ fps with around one second latency. Quality matches Bernini R and Kling 3 Omni on an editing benchmark; Apache 2 license, 32.5 GB weights.
Live demoFramework giving video models precise camera movement control from an input image plus a camera path. Code released under Apache 2, built on Wan 2.2.
Live demo1.7T-parameter MoE open model with D-Spark speculative decoding. Competitive with leading open models and some closed models; very cheap at about 6 cents per task on Artificial Analysis. Raw size 893 GB.
DeepSeek's own agent harness for orchestrating its models, replacing third-party frameworks like Hermes or OpenClaw. Still a developer preview with breaking changes expected.
xAI's top closed model, matching GPT 5.6 on the Artificial Analysis index and beating it on GDPval, though behind on Deep SWE. Fast and cost-efficient, available in Cursor, Grok Build and via API.
Open-sourced 2.4T-parameter MoE with 95B active, 4.8 TB in size. Close to frontier closed models and ahead of DeepSeek V4 on the leaderboard, cheaper than GPT and Claude.
Generates complete audio scenes including speech, music, sound effects and ambient noise in one generation, with emotion and multilingual support. Under 12 GB, open-sourced on GitHub.
Live demoOpen-source video model with native audio and multi-shot, the fastest open video model. Tested against MiniMax H3 on several prompts and mostly lost on quality, winning only on camera orbit. Int8 version is 22 GB and works with LTX 2 LoRAs.
Live demoOpen-source video model compared against LTX 2.5; won most tests for coherence, consistency and text rendering, though slower, about twice the generation time.
Live demoTranslates sign language from the phone camera into text, trained on over 100,000 hours across 50+ sign languages. Launches with American Sign Language on Gboard and Live Transcribe on Pixel 11.
Ultra-fast mode for GPT-5.6 generating up to 750 tokens/sec, 14x standard, powered by Cerebras. Limited private preview for select customers.
27B dense multimodal open model that beats Opus 4.6 Max on many agentic and coding benchmarks, with 1M context. Full model is 56 GB; Unsloth GGUFs go down to 9 GB.
Post-trained version of GLM 5.2 with big gains in coding, long-horizon tasks and cybersecurity, leading on several benchmarks. Weights to be open-sourced in about two weeks after safety testing; available now via the Z Code plan.
Google's fast multimodal Flash model, leading small models on a frontier code benchmark with 340 tokens/sec. More expensive than GPT 5.6 Luna (40 cents vs 5 cents per task). Available in Antigravity, AI Studio, Android Studio and Gemini app for paid subscribers.
Live demoOpen-source music generator taking style prompts and structured lyrics. Model is 9.8 GB with an int8 version fitting low-end GPUs; the creator already made a full tutorial.
Open-source text-to-speech with few-second voice cloning, emotion and multilingual support, including dubbing Chinese to Spanish. Total size about 5.5 GB.
Live demoOpen-source 114B MoE video generator with native audio, 10-second generations and refiner up to 1080p. 228 GB model needing eight Hopper GPUs, so impractical for consumers.
Tiny 45M-parameter model in a 14 MB binary using about 28 MB RAM, for tool calling and device control on small hardware. Decodes 500 tok/s on Raspberry Pi 5; not suited to long-horizon reasoning.
Generates entire open 3D worlds from text using multiple agents that plan layout, terrain and materials, then generate assets coarse to fine and verify consistency. Open-sourcing not confirmed.
Live demoWorld action model trained on over 1 million hours of first-person human video, showing a scaling law where more human data improves performance on unseen robot data. Adapts to robots with little robot data.
Open 30B MoE model with 1M context, less intelligent than Qwen 3.6 but about twice as fast. About 22 GB in FP4.
Open-source model router that picks the best model per task. With Opus 4.8 and other models it completed more tasks at about three times lower cost than Opus 4.8 alone.
Project simulating human populations with 8.3 billion persona agents and 1,290 attributes to test apps, surveys and chatbots. Creator notes it is unclear how closely simulated data matches real people.
Open 30B dense model under Apache 2, around 60 GB. Meta's benchmarks look cherry-picked; Artificial Analysis shows it does not beat Qwen 3.6, so the creator calls it underwhelming.
