Your central hub for AI news and updates on Models. We're tracking the latest articles, discussions, tools, and videos from the last 7 days.
Fyra's Brief
Google’s Gemini 3 Deep Think AI reasoning mode has been updated to address science, research, and engineering challenges. The updated mode is now available to Google AI Ultra subscribers and can be tested via the Gemini API.
The updated Gemini 3 Deep Think AI reasoning mode is a significant advancement in the field of AI research and has the potential to solve complex science, research, and engineering challenges.
Fyra's Brief
OpenAI's GPT-5.3-Codex-Spark model codes 15x faster than GPT-5.3-Codex, targeted for real-time coding and collaboration
OpenAI's GPT-5.3-Codex-Spark model offers a trade-off between speed and accuracy, suitable for real-time coding and collaboration.
Fyra's Brief
Hollywood organizations and Disney condemn Bytedance's Seedance 2.0 for enabling 'blatant' copyright infringement and unauthorized use of U.S. copyrighted works.
The swift backlash against Bytedance's Seedance 2.0 highlights the need for AI companies to prioritize copyright protection and intellectual property rights.
Fyra's Brief
Anthropic's AI chatbot Claude experiences significant growth in US downloads following Super Bowl ads, surpassing its predecessor ChatGPT's gains. Claude's daily average downloads increased 32% from 37,400 to 49,200.
This surge in US downloads highlights the effectiveness of targeted marketing and key differentiation in driving growth for AI-powered consumer apps.
Fyra's Brief
DeepSeek V3.2 achieves 7360 TGS throughput on a single GB300 GPU, but still has significant room for improvement.
DeepSeek V3.2 still has significant room for improvement, and its performance relies heavily on the choice of parallelization strategy and model architecture.
Fyra's Brief
ByteDance's next-gen AI model, Seedance 2.0, can generate clips based on various inputs, offering a substantial leap in generation quality.
ByteDance's Seedance 2.0 AI video generator represents a significant improvement in AI video generation capabilities, which may disrupt the industry and have implications for AI professionals.
Fyra's Brief
Databricks has released generative AI features in AWS GovCloud, enabling government agencies and organizations to access AI technology and data intelligence capabilities.
This release expands the capabilities of government AI initiatives and demonstrates Databricks' commitment to innovation in highly regulated environments.
Fyra's Brief
ByteDance is working to enhance safeguards on its Seedance 2.0 AI video generator after receiving copyright infringement complaints from major studios and Hollywood trade groups.
ByteDance's move to strengthen safeguards on Seedance 2.0 highlights the growing need for responsible AI development and adherence to intellectual property laws.
Fyra's Brief
Anthropic's Claude 4.6 Opus surpasses ChatGPT in multiple areas, making it a strong alternative for those seeking better productivity and writing quality.
For AI professionals, Claude 4.6 Opus offers significant improvements over ChatGPT, making it a compelling alternative for those seeking enhanced productivity and writing quality.
Fyra's Brief
A global outcry has erupted as OpenAI decides to sunset ChatGPT's GPT-40 model, prompting Chinese fans and users worldwide to express their discontent.
The controversy surrounding OpenAI's decision to discontinue GPT-40 highlights the complex relationship between humans and AI models, and the responsibility that AI developers must take in handling their users' emotional connections with technology.
Fyra's Brief
Anthropic has upgraded its Claude AI free plan with four new features: file creation, connectors, custom skills, and longer conversations, potentially making paid subscriptions less necessary.
These enhancements make Claude AI more competitive, but subscribers may still want to consider the benefits of their paid accounts.
Fyra's Brief
InclusionAI launches Ling-2.5-1T, an open-source instant model with 1T parameters, delivering comprehensive upgrades in model architecture, token efficiency, and preference alignment.
The release of Ling-2.5-1T marks a significant step towards making AI more universally accessible and highlights the ongoing advancements in instant models.
NVIDIA just added `z-ai/glm5` to their NIM inventory, and I've updated `free-claude-code` to support it fully. You can now run Anthropic's Claude Code...
An ultra-fast model for real-time coding in Codex Discussion | Link...
NVIDIA Nemotron Nano v2 12B VL model enables multi-image reasoning and video understanding, along with strong document intelligence, visual Q&A and su...
Community talk
Qwen 3.5 Open Source: Native Multimodal, Ultimate Efficiency!
Qwen3.5-397B-A17B Unsloth GGUFs
New DeepSeek update: "DeepSeek Web / APP is currently testing a new long-context model architecture, supporting a 1M context window."
The new Gemini Deep Think incredible numbers on ARC-AGI-2.
How is this not the biggest news right now?
GLM-5 scores 50 on the Intelligence Index and is the new open weights leader!
Lead product + design at Google AI Studio promises "something even better" than Gemini 3 Pro GA this week
GLM-5 Officially Released
Zhipu is rolling out GLM 5 now!
DeepSeek has launched grayscale testing for its new model on both its official website and app. 1M content length!
The ‘updated Chat model’ planned for this week, that released today.
I owe the "it's gotten worse" crowd an apology regarding ChatGPT 5.2
How are Chinese models so strong with so little investment?
DeepSeek-v4 Benchmarks Leaked
The Hollywood Nerf has begun: Seedance 2.0 is already being lobotomized. Local Weights are our only defense.
Gemini 3 Deep Think multi-modal understanding: math images to zero-shot visualization (this is a standalone HTML page)
ChatGPT-4o remains available to the other plans until April 3rd but the ChatGPT-4o-latest API will be deprecated on February 17th (repost)
Nemotron3 Super/Ultra: FP4 pre-training, H1 2026 release, "NVIDIA is a company of volunteers" (all from recent NVIDIA interview)
KaniTTS2, our text-to-speech model with frame-level position encodings, optimized for real-time conversational AI.
LLaDA2.1 (100B/16B) released — now with token editing for massive speed gains
WTF WTF WTF
Gemini 3 Deep Think SVG Pelican Riding a Bicycle
Introducing GPT‑5.3‑Codex‑Spark
Introducing GPT‑5.3‑Codex‑Spark
''The MiniMax M2.5 model weights will be open-sourced on HuggingFace'' - from the official MiniMax account on X
Minimax M2.5 Officially Out
Anthropic and OpenAI dropped their coding models 20 minutes apart. This rivalry is getting wild
Z.ai implemented new guardrails in the GLM-5.
Google Gemini 3.1 Pro Preview Soon?
GLM 5.0 & MiniMax 2.5 Just Dropped, Are We Entering China's Agent War Era?
MiniMax M2.5 Released
GLM 5 Released
DeepSeek just updated to a 1M context window!
Fine-tuned FunctionGemma 270M for multi-turn tool calling - went from 10-39% to 90-97% accuracy
Opus 4.6 is really a goated all-around model, the best since GPT-4 in my opinion
GPT-5.2 Just Solved a 15-Year Physics Mystery — Then Scored 0% on the Physics Exam
Opus 4.6 v Codex 5.3 w. Extra High
Rumored/maybe confirmed? SOTA model - Seed 2.0 Pro - by ByteDance
ChatGPT 4o is being retired today, and some users are very unhappy about it. A petition with around 20,000 signatures is urging OpenAI not to remove it. The same group is also calling for a mass cancellation of subscriptions in protest.
AI might f white collar this year, not in a few years
MiniMax-M2.5 Checkpoints on huggingface will be in 8 hours
minimax 2.5 is only 230B / 10B active. Insane performance of its size.
Step 3.5 Flash is a beast?
Seeking alternatives to ChatGPT after GPT-4o retirement — which AI platforms are best for long-term use?
GLM-5 is here
Claude Sonnet 4.5 playing Pokemon TCG against me
Z.ai releases GLM 5
OpenAI admits the Pro model lacks memory. It's devastating for me. Does it matter to you?
The tides have turned. Codex-5.3 is super good. Congrats OpenAI
Emotions with Seedance 2.0
GPT-5.2 solved a previously unsolved problem in quantum field theory. A top physicist said: "It is the first time I’ve seen AI solve a problem in my kind of theoretical physics that might not have been solvable by humans."
GPT-5.2-xHigh & Gemini 3 Pro Based Custom Multi-agentic Deepthink: Pure Scaffolding & Context Manipulation Beats Latest Gemini 3 Deep Think
Why can't Anthropic increase the context a little for Claude Code users?
Hollywood is cooked
Difference Between Opus 4.6 and GPT-5.2 Pro on a Spatial Reasoning Benchmark (MineBench)
Difference Between Opus 4.6 and GPT-5.2 P on a Spatial Reasoning Benchmark (MineBench)
Minimax claims M2 is Opus 4.6 competitor on SWE-Bench Verified
Oh my God, the update that they did at 5.2 is absolutely insane.
I tested this update and...
Update on the First Proof Questions: Gemini 3 Deepthink and GPT-5.2 pro were able to get questions 9 and 10 right according to the organizers
ChatGPT 5.3 is coming VERY soon
Intent Model
Claude.ai is using very short prompt caching time limits for Opus 4.6, causing it to eat through limits very quickly if you spend even a few minutes between consecutive prompts.
I want to throw some love toward Haiku 4.5
Is there a known reason why long ChatGPT conversations seem to degrade instead of failing explicitly?
Did anthropic just replace sonnet 4.5 with opus 4.5
ChatGPT can’t identify obvious celebrities now? wtf am I paying $20 for