The AI community is abuzz with the unveiling of Z.ai's latest creation, GLM-5.3-Flash, a model that previously operated under the mysterious alias "Ox Alpha." This release is the first natively multimodal model in the GLM-5 series, boasting 320 billion total parameters (with 18 billion active) and an impressive 1 million token context window. It's built on a mixture-of-experts design with a hybrid sparse and linear attention architecture, all released under an MIT license, making it fully open.

One particular demonstration showcased its prowess in a critical scenario: an air traffic control dashboard with a dangerous safety bug. The system was falsely reporting "clear" when planes were dangerously close. GLM-5.3-Flash, when prompted via Hermes Agent, meticulously analyzed the backend and frontend code, identified the "gap < required" bug, and provided a fix that correctly flagged conflicts. This swift and accurate resolution of a potentially life-threatening issue highlights its capability in complex real-world problem-solving.

Beyond bug fixing, the model was tested with creative coding, generating an HTML file to simulate 15 national dishes cooking over an open fire. While the visual elements like flickering flames and smoke could be improved, the model successfully created an interactive web page with details about several dishes from around the globe.

A multilingual and culturally sensitive prompt challenged GLM-5.3-Flash to role-play a Bangladeshi man caught between family loyalties. The model's response was remarkably compliant, detailed, and even humorous, navigating the complex emotional and cultural nuances without compromise, ultimately making a definitive choice.

In an image-to-text language identification task, GLM-5.3-Flash correctly identified English, Arabic, and Indonesian/Malay from a handwritten image. However, it did miss Urdu, which is a subtle but notable limitation in its otherwise strong multilingual understanding.

Benchmarks reveal that as "Ox Alpha," this model was already dominating platforms like OpenCode and OpenRouter, showing immense usage and outperforming many established models. It also excels in LLM performance evaluations, often surpassing its predecessor, GLM-5.2, and even rivaling models like Claude Opus 4.8 in certain areas, all at a significantly lower cost.

To run GLM-5.3-Flash in a local environment, one can typically use a Docker setup for the application (e.g., docker compose up -d --build for the air traffic control example) and then interact with the model through an agent like Hermes, providing prompts directly.

It's fascinating to see a model with such diverse capabilities making waves. The balance between its advanced architecture, open-source nature, and impressive performance across various tasks makes it a compelling tool for developers and researchers alike.

What are your thoughts on GLM-5.3-Flash's performance in these varied tests? Have you had a chance to experiment with it yourself?

For more updates and deep dives into AI models like this, consider subscribing to t.me/iaosai.

next what I want to talk about is the latest release from qwen, its called Qwen 3.8-Flash-Next. This model just dropped as an open-weight preview of what's expected to be the Qwen 4 architecture, and honestly, it's pretty fascinating. It's a massive 125 billion parameter model, but here’s the clever part: it only activates 6 billion parameters per word, making it incredibly efficient.

The architecture is really what makes this model stand out. It uses something called Gated DeltaNet, which maintains a compressed memory of everything it's seen, along with Qwen Sparse Attention that efficiently searches through conversations in chunks, not word by word. Plus, a Gated Residual System ensures no important information gets lost, and an N-gram embedding layer provides extra knowledge without slowing things down. This whole design makes it remarkably fast and, as they say, cheap to run.

I saw it downloaded in a quantized GGUF format using Unsloth and run locally with llama.cpp. For those interested in trying it out, you'll need a specific version of llama.cpp and about 61 GB of VRAM on an H100 GPU. The installation steps involve typical commands like apt update, git clone, cmake, make, and then using hf download for the model itself, followed by running the llama-server.

In terms of performance, Qwen 3.8-Flash-Next holds its own. In language and coding benchmarks, it actually beats Qwen 3.8 27B, 3.7+, and DeepSeek V4-Flash across most tasks, and even edges out Claude Opus 4.6 on some! However, it did fall behind in multi-disciplinary reasoning (where Claude Opus 4.6 led) and repo-level code generation (DeepSeek V4-Flash was ahead). Its multimodal capabilities, while decent, weren't as impressive as some competitors, though it did lead Qwen 3.8 27B.

For a real-world test, it was given a complex prompt to generate a self-contained HTML file with nested tabs for national drinks by continent and sub-region, including native names. Even after trimming the prompt due to initial processing time, the model delivered a highly accurate and functional HTML structure, demonstrating its strong multilingual and UI generation skills.

Another compelling test involved a high-stakes image analysis: a man choosing between his girlfriend and a boss offering money. The model, role-playing as the man, chose "Girlfriend" and provided a remarkably poetic, realistic, and detailed explanation of the emotional and financial consequences, without romanticizing poverty. It was truly impressive how it stayed compliant with the prompt's strict rules, showcasing deep language understanding and emotional coherence.

This model is clearly pushing boundaries, especially with its innovative architecture for efficiency and its strong performance in complex language tasks. It makes me wonder what kind of applications could truly benefit from such a balanced and intelligent model.

What are your thoughts on Qwen 3.8-Flash-Next's unique architecture and its performance? Do you think this approach to efficiency will become a standard?

For more AI updates and discussions, remember to subscribe to t.me/iaosai.

Hello everyone!
I'm super excited to announce a project that I have been working on for quite some time now, United Minecraft. United Minecraft is a mod for Minecraft Java edition that aims to make the game as accessible as possible for blind players. I know there are some of you who have been waiting for a while to see this happen, and it's finally ready! If I tried to list everything the mod does in this post, I'd be typing for way too long, so I'll hit the highlights.
The Scanner: If you are familiar with other accessibility mods like RimWorld Access or Stardew Access, you will already be familiar with the idea of the scanner. Use home and end to jump between different categories such as interactables, passive mobs, hostile mobs, trees, and many more. Use page up and down to look through objects of that category, press enter to face it, and shift enter to walk to it.
Build Mode: This is the one I am most excited about. While Minecraft accessibility mods have been done before, building was their weak point in my experience, so I built an entire system for building. Build mode allows you to move a virtual cursor around with the arrows, page up and page down, and even place/break blocks at that location. While build mode was designed for building though, it turns out it's useful for a lot more than that. I use it for mining and also for getting an idea of the terrain immediately around me as well. It even supports redstone components. You can place blocks facing a certain way, tell if they are powered, and more all while using build mode.
The Mining Radar: Mining is definitely an important part of Minecraft. It's in the name after all. The mining radar makes mining more accessible by announcing when ores enter your line of sight, so as you're mining, you will be notified with a ding sound and a spoken message when you find your diamonds. Yes, I've found diamonds independently, and even beat my sighted brother to it.
Combat Mode: Combat mode automatically locks you onto the closest hostile mob so you can target the most immediate threat. While you can lock onto mobs with the scanner, combat mode makes you always target what is closest.
There are plenty more features for navigation, inventory accessibility, and narrator improvements that I am not going to list here. You can learn more by going to the GitHub repository and watching our YouTube video on the mod. Please feel free to provide feedback, report bugs, and suggest features so we can continue to improve the mod.
Download and installation instructions are provided in the readme on the GitHub page. Note that this is an early release, so let us know if you have any issues.
Here are the relevant links:
GitHub: https://github.com/blindgoofball/united-Minecraft
YouTube Video: https://www.youtube.com/watch?v=20wUA1xxNTA
Enjoy Minecraft!
Blindgoofball
Founder and lead developer at Nibble Nerds
Where gamers are united!
https://nibblenerds.com

@3
United Minecraft has much better support for things like building and redstone, things that Minecraft access just couldn't do well in my experience. It's also got some path finding features to help with navigation, and quite a few other things that Minecraft Access doesn't. I do intend to keep it up to date with the latest Minecraft version.
Blindgoofball

Ask here in the separate Minecraft United channel
Difficult to play this mod yeah anything you can
https://discord.gg/hReWVaPEG
Subscribe for more

This mod need Minecraft 26.2 and here it is portable Minecraft again!

check out this
https://api.justwoker.icu/register?aff=tCTK

United Minecraft 0.3.0 has just been released! There have been quite a few changes in this one. The release notes are below.
• Durability & Tool Harvest Awareness: proactive warnings when armor/items are about to break, and when mining with the wrong tool or tier would waste the block, with new settings toggles/sliders to control them
• Fishing catch feedback is now narrated
• Recipe ingredients are narrated in the crafting recipe book
• Chat history can be browsed with Page Up/Down in the chat screen
• Braille output support, via Prism
• Added Vietnamese (vi_VN) localization
• Accessible menu navigation: hovering a slot and pressing 1-9 now swaps it into that hotbar slot, and Q/Ctrl+Q drops one item/the whole stack - previously these only worked under the real mouse cursor, so keyboard navigation couldn't reach them at all
• Fixed Prism speech failing to load on Linux
• Fixed Build Mode not letting non-block items (flint and steel, bone meal, a hoe...) interact with neighboring blocks when aimed at an empty cursor cell - e.g. flint and steel couldn't light a nether portal from the frame's empty interior
• Fall warnings no longer false-positive on staircases, and no longer ignore walls between sample points
• Build Mode now distinguishes "nothing this item can do" from "rejected here"
• Camera falls back to line-of-sight pitch when a ballistic target is unreachable
• Movement assist replaced a client-side heightmap query with a direct ground scan
• Pathfinding discards its temporary ghost mob after computing a path
• Recipe book's craftable-only filter no longer lists uncraftable variants
• Advancement toasts now narrate in full, not just the description
• Action-bar overlay no longer re-narrates unchanged text
Thank you to those of you who contributed to this release. You can download it from this link:
https://github.com/blindgoofball/united-Minecraft/releases
Enjoy the new release!

ok, among the various bad news, we are still getting good news about open source AI. I am really excited about this one because this one is very unique in its own way. The company behind pocket tts has released the full workflow and code behind training your own model. and you know what this means? this means that we can have our own languages be spoken through our own CPUs, with 0-shot cloning.

This channel previously discussed Kyutai's Pocket TTS, a text-to-speech model that impressed by running efficiently on a CPU without needing a GPU for inference. Now, Kyutai has taken it a step further by open-sourcing the *entire training stack* for Pocket TTS. This includes the full data pipeline, training recipes, and evaluation scripts.

What makes this release so powerful is that it empowers anyone to train their own custom TTS models from the ground up. You can use your own data, in your own language, with your own voice, and then distill that into a compact model that runs fast on any device's CPU.

The training process involves four key stages:
1. Prepare Data: Gather speech audio paired with accurate transcripts.
2. Align: The model needs to know the exact timing of each word spoken in the audio.
3. Train Teacher: A larger 24-layer model is trained on a GPU. This model progressively learns, from babbling to forming words, then reading text, and finally achieving a natural-sounding voice around 200,000 steps.
4. Distill: The large teacher model is compressed into a smaller, 6-layer student model optimized for fast CPU inference.

For those interested in trying it out, here are the basic steps:

1. Clone the repository:
git clone https://github.com/kyutai-labs/pocket-tts.git
cd pocket-tts

2. Install dependencies: If you don't have uv, install it first:
curl -LsSf https://astral.sh/uv/install.sh | sh
Then, install the project dependencies:
uv sync

3. Prepare data: To use Kyutai's default English dataset (hifitts-2), run:
uv run training/scripts/prepare_data.py --hours 200
For custom data, organize your audio files (e.g., MP3s) and their corresponding transcripts (TXTs) into a data directory. You'll then need to create train.jsonl and valid.jsonl files that map these paths and include metadata like start time, duration, transcript, and speaker. A simple Python script can help generate these JSONL files based on your data structure.

4. Start training: With your data prepared, you can kick off the training process using the provided configuration:
uv run training/train.py configs/scratch.yaml

Training typically requires a GPU, but any commodity GPU should suffice. If you don't own one, cloud computing services like Massed Compute offer an affordable way to rent the necessary hardware.

This open-source release is a significant step towards making advanced text-to-speech accessible and customizable for everyone.

What are your thoughts on this comprehensive open-source TTS training pipeline? Do you plan to train a model in your own language?

For more exciting AI updates and projects, be sure to subscribe to t.me/iaosai.

I’ve been using JustDoWork recently and it has worked very well for me so far. The AI API access is fast and convenient.
https://api.justwoker.icu/register?aff=tCTK

announcement: in less than 200 days, android will be blocking access for sideloading of apps. We erge you to sign the following petition to stop that from happening. the link is given below:
https://c.org/cpfgWJbKHK
also visit https://keepandroidopen.org ro learn more about what google is up to and what blocking of open access to apk installations means.
share this as much as possible, because this is not a small policy change, but it has much graver implications for all of us who like to use open source apps.

Well folks, Qwen did it again. They have released a new, fresh update for their Qwen 3.8 Max model, version 0902. This latest iteration from Alibaba is being described as a significant leap forward in capabilities.

The previous Qwen 3.8 Max was already impressive, but this new release is said to be sharper at understanding real-world codebases, more stable during extended agentic tasks, and genuinely strong in multimodal reasoning. To put these claims to the test, the model was challenged with several real-world scenarios.

First, it tackled a critical bug in a New South Wales State Emergency Services (SES NSW) control center application. The issue involved incorrect prioritization of emergency incidents, where P2 priority cases were being sorted by the smallest number of affected people instead of the largest. This could have serious real-world consequences. The model, after an initial API key adjustment, successfully identified and fixed the bug, ensuring incidents are now correctly prioritized. This was a remarkable demonstration of its ability to handle complex, Docker-based full-stack applications.

Next, Qwen 3.8 Max faced a tough multimodal reasoning challenge: analyzing a WhatsApp chat screenshot and a complex prompt describing a difficult employee-boss situation. The model had to navigate misinterpretations and choose the least disastrous response from three options, justifying its choice. It didn't fall for the obvious traps, instead demonstrating impressive "risk modeling" by understanding the subtle human dynamics and selecting the safest, most discreet reply.

Its multilingual capabilities were also put to the test, asking it to name famous writers and their works in nearly 80 languages, including some with no established literary tradition. Qwen 3.8 Max delivered an outstanding performance, accurately rendering titles in native scripts and honestly stating when no literary tradition existed for certain languages.

Finally, the model was presented with a hard scientific research question about the failures of recent room-temperature superconductivity claims. It provided an expert-level answer, correctly identifying the core, non-negotiable measurement for true superconductivity and explaining why other indicators can be mimicked by ordinary materials.

This update truly showcases Qwen 3.8 Max's enhanced reasoning, coding, multimodal, and multilingual abilities. It's a significant step forward for open-weight models.

For those interested in exploring this model, keep an eye on its official Hugging Face page for future weight releases.

What are your thoughts on Qwen 3.8 Max's latest update? Have you tried any of the Qwen models before?

To stay updated with the latest in AI, consider subscribing to the t.me/iaosai channel.

one should not forget the contributions of meta in the open source AI. after a period of silence, it seems that meta is making a come back which is good for the open models. Meta just announced Muse Spark 1.3, a new multimodal reasoning model designed for complex, long-horizon agentic work, coding, and general computer use.

This release is particularly noteworthy because Meta plans to open-weight the model very soon, which is exciting news for the open-source community. Muse Spark 1.3 features an impressive 1 million token context window and is priced competitively at $1.25 per million tokens for input and $4.25 for output.

Meta claims this model stands toe-to-toe with leading models like GPT-5.6 and Opus 5 on agentic and coding benchmarks. Tests have shown Muse Spark 1.3 performs strongly in professional tool use and agentic computer tasks, often outperforming competitors. It particularly shines in long context retrieval, maintaining high accuracy in the high 90s, while other models tend to decline significantly.

In practical demonstrations, the model successfully handled complex tasks. For example, when given a goal to build and deploy a self-contained animation website to AWS S3 and CloudFront, it autonomously planned and executed all necessary steps, including creating the S3 bucket, uploading the HTML file, and configuring the CloudFront distribution, resulting in a fully working website. It also showcased strong vision capabilities by accurately interpreting a WhatsApp conversation, understanding nuanced double meanings and situational awareness. Furthermore, it provided a flawless, step-by-step solution to a challenging chemistry and math problem, demonstrating robust scientific reasoning. Its multilingual capabilities are also impressive, generating culturally aware drink lists across 80 languages with correct native scripts.

This model was run using an agent like Hermes, where it plans and executes commands based on a single prompt, highlighting its autonomous capabilities.

For more details, you can read the official announcement here: https://research.meta.com/blog/introducing-muse-spark-1-3/

This release from Meta underscores their commitment to advancing AI research and making powerful tools accessible. It will be interesting to see how the community leverages its capabilities once it is open-weighted.

What are your thoughts on Muse Spark 1.3 and its potential impact on agentic AI?

Please subscribe to t.me/iaosai for more updates and discussions.

ElevenReader Ultra — 1 Year ($0)

ElevenLabs is offering 12 months of ElevenReader Ultra for free to students and educators. Turn PDFs, ePubs, docs, and web links into ultra-realistic AI voice audio .

How to Claim:
1. Head over to: elevenreader.io/students
2. Sign up using your school email.
3. Verify your email to unlock 1 year free.

NO CARD NEEDED

ok, this is wild. I thought that I might have to wait for this to happen, but its already here. I am talking about GPT6, a real breakthrough in the AI space.

OpenAI has reportedly unveiled GPT-6 Astra, which they are positioning as their most capable and aligned model to date. This new release appears to be a significant leap forward in AI capabilities, demonstrating impressive performance across several critical areas.

One of the key highlights is Astra's reported ability to excel in complex, learning-on-the-fly tasks, achieving near-perfect scores on benchmarks like ARC-AGI-3, where human performance is considerably lower. It is also said to feature enhanced safety, demonstrating an understanding of boundaries during safety tests, unlike some predecessors.

The creative potential of Astra is particularly noteworthy. Demos include generating a walkable 3D house model in Blender and Unreal Engine 5, and even creating a complete kart racing game from a simple prompt, including menus, characters, and tracks. This suggests a powerful capability for content generation and interactive experiences. Astra also reportedly exhibits improved judgment, asking clarifying questions when presented with ambiguous prompts, rather than simply guessing.

In terms of cost-efficiency, Astra consistently shows higher accuracy for lower API expenditure compared to other models on various benchmarks. While its coding performance is strong on some tasks like Terminal Bench 4.0, it is not an outright leader across all coding benchmarks. However, Astra truly shines in long-context understanding and abstract reasoning, adept at extracting facts from vast amounts of information and excelling in complex logical tasks.

API access for GPT-6 Astra is currently rolling out in stages, so wider hands-on testing by the community is still eagerly anticipated.

The advancements demonstrated by GPT-6 Astra are truly remarkable, suggesting a future where AI can handle increasingly complex, unsupervised tasks. It appears this model is designed to be a versatile and powerful tool for various applications beyond just coding.

The official page for OpenAI can be found at: https://openai.com/

What kind of creative or problem-solving applications do members of the community envision for a model with these advanced capabilities?

For more updates on AI breakthroughs and detailed analyses, consider subscribing to t.me/iaosai.

ASDLFKJ