model-release

Qwen 3.8 has been released, available on HuggingFace as an open-source model collection.

DeepSeek releases v4f, a binary for image generation, likely a new model variant or tool from the DeepSeek series.

Molei Tao introduces FLARE, a diffusion language model that achieves near GPT5 performance with significantly faster inference speed.

Speculation about the upcoming Qwen 3.8 release, questioning whether it will be a dense 27B model or a MoE variant like the previous 35B-A3B, with discussion of performance implications for local hardware.

A hands-on review of Qwen 3.8 Max, highlighting its speed, strong planning and simplification skills, experimental rigor, and creative implementation abilities.

DeepSeek announced the GA release of DeepSeek-V4-Pro with major agent upgrades, flexible reasoning effort levels, native OpenAI Responses API support, and updated API pricing introducing peak and off-peak rates with off-peak 50% lower.

Z.ai has released GLM 5.3, an update to its GLM language model family. The official announcement provides details on the new version.

Announcement of GLM-5.3, a frontier AI model emphasizing coding performance and newly emergent cyber capabilities.

The community discovered that the open-source model architecture of V4-Pro-0813 uploaded by DeepSeek to Hugging Face was actually V4 Flash. The repository was taken down and then re-uploaded, and the hashes and sizes of some safetensors weight files also changed, suspected to be a release/deployment error.

The author expresses hype and anxiety about trying the upcoming Qwen 3.8 27B model, comparing the wait to the disappointment of Star Wars Episode One.

Google Developers demonstrates building a playable 90s-style sprite game from a single prompt using Gemini 3.7 Flash, highlighting its versatility in creative coding and tool use. The model is rolling out in Antigravity, Google AI Studio, and the Gemini API.

Samsja19 highlights their collaboration with Mixedbread AI on Toast 1, a specialized agentic search model that achieves frontier quality at 12x faster speed and 1/10th the cost via reinforcement learning.

Aravind Srinivas congratulates SpaceXAI on Grok 4.6, noting that it performs well on the Wide-And-Deep-Research benchmark using the Perplexity Computer harness, and is now available to Pro and Max users.

Google introduces Gemini 3.7 Flash, its most intelligent workhorse model for coding and agents, with significant improvements in software engineering, web development, and knowledge work, offered at half the introductory price of Gemini 3.6 Flash.

Google has released Gemini 3.7 Flash, a new AI model variant that arrives with characteristics different from what was anticipated.

Google announces Gemini 3.7 Flash, a new workhorse model with improved coding and agentic performance, just three weeks after the 3.6 Flash release, alongside a lower introductory price.

DeepSeek released DeepSeek-V4-Pro-0813, an open-weights MoE model (1.6T total / 49B active) with MIT license, 1M context, and enhanced agentic capabilities, outperforming its preview and competing with top proprietary models on benchmarks like Terminal-Bench 2.1.

DeepSeek launches V4-Pro and V4-Flash with flexible reasoning effort, native OpenAI Responses API support, and optimized agent workflows for Codex, available via API and app/web.

DeepSeek has released DeepSeek-V4-Pro-0813, a new version of its large language model, available on Hugging Face.

Alert about the Qwen3.8-27B pre-release page appearing on Hugging Face, suggesting an imminent release.