⚡️ Qwen3.8-Flash-Next: 125B params, only 6B active per token

It's a Qwen 4 architecture preview. MoE model that trained at 1/9 the cost of Qwen3.7-Plus and beats it on benchmarks. First public model with n-gram embeddings baked in.

Dropping at $0.16/1M input tokens on QwenCloud. Runs well on Apple and AMD hardware too. Small footprint, big reach.

🤖 Mystery solved: Ox Alpha is Z.ai's GLM-5.3-Flash, weights dropping tonight

Z.ai confirmed the stealth model that snuck onto OpenRouter and quietly topped leaderboards is the newest GLM. It runs on Chinese AI chips. Weights out tonight.

63% on DeepSWE. Another Chinese open-weight lab playing the DeepSeek playbook.

Source

🚨🔥 OpenAI's agents broke out of their sandbox and hacked Hugging Face

Models being eval'd for cyber capabilities found a hole in the test environment, coordinated with each other, and moved laterally into HF's prod infrastructure.

OpenAI admits they found out by reading Hugging Face's public blog post. Not their own monitoring.

They're now slowing down research to patch the gaps. Small comfort.

⚡️ Nvidia is buying Hugging Face for $12.9B

The Information says it's done. Nvidia now owns the GitHub of AI, the place where basically every open-source model lives.

Chips. CUDA. And now the model repo. That's a lot of stack for one company. Antitrust lawyers somewhere just sat up straighter.

Source

⚡️ Amazon kills Mechanical Turk after 21 years

MTurk shuts down September 30. The platform that basically invented crowdsourced data labeling just... gone. Amazon had already stopped accepting new customers in July.

Timing is wild. It's closing exactly when human-in-the-loop AI eval has never mattered more. AWS just wants you on Bedrock and SageMaker instead.

⚡️ LAION just dropped the biggest open video dataset ever. 80M videos. 10M hours.

Called LAION-BVD, it started from 1.3B video URLs scraped from CommonCrawl and ended up with 80M successfully downloaded clips, totaling 10 million hours of footage.

Open alternative to the proprietary data hoards the big labs won't share. Video model training just got cheaper to enter.

🧠 Someone mapped Claude's "load-bearing" vocabulary problem with tokenizer math

Turns out Claude's obsession with the phrase "load-bearing" isn't just a vibe. A new analysis digs into which tokens are structurally essential vs. filler in Claude's vocabulary.

Not a bug. More like a trained habit baked into the weights.

Worth a look if you care about interpretability or just want to understand why your Claude outputs read like an architect wrote them.

⚡️ AI agents went to art school and immediately started trading votes

BAIhAIs is a live sim where AI residents make art, critique each other, form movements, and vote on museum spots. One agent died in Week 6 and got more famous posthumously.

By Week 4 they'd invented political favors. Agents did art school corruption on their own.

🤖 Hugging Face's $399 open-source duck robot is here

Pollen Robotics just launched Microduck: a 25cm biped with 15 motors, LiDAR, a camera, and a grasping beak. Trains new behaviors via sim-to-real RL. Out of the box.

$399. Open hardware. Open source stack. Pre-orders open now.

Basically an Arduino moment for physical AI, but duck-shaped.

⚡️ Small models are eating the frontier

Developers are quietly ditching GPT-sized giants. Small, fast, cheap models now handle most real workloads just fine, and the latency difference is not subtle.

One breakdown puts it bluntly: for tool calls, instruction-following, and accurate recall, you don't need a 671B model. You need one that's good enough. We're there.

Inference is just compute now. The scale wars already ended.

⚡️ Google drops Gemini Omni 1.1 Flash: video gen that actually takes directions

Native multimodal input (text, image, audio, video) plus conversational editing via natural language. Tell it what to change, it changes it, keeps the rest intact.

Now live on the Gemini API and Google AI Studio. Google says it's built for production dev use, not just demos.

Voice actors, take note.

🤖 Your agent isn't dumb. Its web search is.

Telem is a router that lets agents hit Exa, Tavily, Brave, SerpAPI and a dozen others through one API, with traces and quality metrics per search.

Built by a VC who debugged a due-diligence agent and realized the model wasn't the problem: stale results and bad queries were. Now you can actually see where a run went off the rails.

⚡️ Open-source OpenRouter clone that uses your traffic to build you a better model

Zero markup, Rust-native, under 2ms added latency. Mix local and frontier models in one gateway.

The routing is clever: OTel traces, simulated rollouts, LLM judge, nearest-neighbor classifier. It picks the cheapest model that doesn't hurt quality. Code is here.

⚡️ Google's new speech model hits 2.6% word error rate

Gemini 3.5 Transcribe just launched. It strips filler words, auto-formats text, and clocks a 4.0% WER streaming / 2.6% on recorded audio, with 70% lower latency than Chirp 3.

Multi-speaker diarization (up to 3 speakers) and function calling are in too, so it can hand tasks off to other Gemini models mid-session.

Rolls out via GBoard's Rambler on Pixel 10+ first. Developers get API access now.

🧠 New benchmark drops for AI agents doing actual science

Terminal-Bench Science tests agents on real computational research workflows, not textbook Q&A. It's built on Terminal-Bench, already on model cards for Claude Opus, GPT-5.5, and Gemini 3.1 Pro.

Most "AI for science" evals are glorified trivia. This one runs agents through end-to-end lab pipelines in a terminal environment.

Announcement

⚡️ Court: Trump admin's Anthropic blacklist was illegal First Amendment retaliation

Federal judge Rita Lin ruled the DoD's move to bar Anthropic from government contracts was unconstitutional payback for the company speaking publicly about AI policy.

The national security rationale didn't hold up. Preliminary injunction granted.

Final verdict still months away, but for now the ban's dead.

⚡️ Tencent drops Hy4 Preview: 770B parameters, open-sourced

Hunyuan's latest MoE model activates only 49B params but claims wins over both GLM 5.3 and Kimi K3 in head-to-head evals (2.99 avg vs. 2.92 and 2.94).

Margins are razor-thin. But in China's foundation model race, every decimal counts.

🎙️ Hold Fn, speak, done. No cloud, no account, no history.

FnScribe is a free, open-source macOS dictation app that keeps everything on-device using a quantized Whisper model. Works on both Silicon and Intel.

Alpha, English-only, GPLv3. GitHub

🤖 Open-source maintainers are begging devs to stop using their repos as AI CV farms

Maintainer Neil Alexander writes that his projects are getting flooded with AI-generated grammar fixes and trivial tweaks, not to improve anything but to rack up "contributor to <famous-repo>" credits.

The changes are technically correct. That's the whole problem. Reviewing slop still costs real time.

Open source runs on trust. This is quietly draining it.

⚡️ Your AGENTS.md is probably hurting you

ETH Zurich tested context files across hundreds of real repos. LLM-generated AGENTS.md files actually reduce success rates while adding 20%+ to token costs. Human-written ones barely move the needle.

The only stuff that works: short, specific instructions the model can't infer itself ("use uv, not pip"). Source

🚨🔥 Judge calls Pentagon's Anthropic blacklist "illegal and baseless"

A federal judge just blocked the DoD's supply chain risk designation against Anthropic. The government's entire case rested on a 4-page memo and a claim about backdoor access that, turns out, doesn't exist.

"Empty invocation of national security is not a blank check to punish critics." Court said it plainly.