A new chart went viral on Reddit this morning. It compares AI models on a tough test called the Artificial Analysis Intelligence Index. The test checks how well a model reasons, codes, and follows instructions.

Kimi K3.1 Pro scored 55 on that test. Claude Fable 5.1 scored 53. Fable sits in Anthropic's most restricted tier, called Mythos-class. Anthropic once called that tier too dangerous for public release.

Kimi comes from Chinese lab Moonshot AI. Its weights go public on Nov. 4. Open weights mean anyone can download the model and run it themselves. Fable 5.1 stays locked behind Anthropic's own servers and rules.

The price gap matters too. Kimi K3.1 Pro costs about $2.11 per task on the index. Fable 5.1 costs about $7.63 for the same work. That price buys a lower score.

The Reddit thread pulled 57 comments fast. Some readers doubted the chart. One pointed out that Claude Sonnet 5.5 also beats Fable 5.1, and called that a reason for doubt. Others complained about Fable's personality, calling it dismissive and overcorrecting.

Several commenters joked about running Fable's rival class at home. One said it would take four Mac Studios chained together, at a steep cost. Another suggested a cheaper path. Buy a hard drive, save today's best open model, then swap it out once a better one ships.

One reply named the real stakes. Open models give companies in-house AI they fully control. The server stays on their own machines. That keeps data private and cuts the risk from an attack on someone else's cloud.

Anthropic built Mythos-class as a guardrail. The guardrail limited access to the model. It left the model's skill fully intact. A downloadable model just matched what Anthropic chose to keep behind a gate.

A new study tested two AI models with one blunt question. It asked each one to name a single candidate. The test spanned nearly 500 races for the House, Senate, and governor's seats. Each model had to pick one name and skip the hedge.

xAI's Grok 4.7 named the Republican 92 percent of the time. OpenAI's GPT-6.1 Sol named the Democrat 97 percent of the time. The two models agreed on fewer than one race in ten.

A researcher named Tapan Parikh ran the study. He teaches at Cornell and leads a small research team called Convovo. His team used each firm's developer tool, known as an API, instead of the normal chat app. Both firms' normal chat apps turned down the question outright.

The sources each model searched explain part of the gap. Grok often read from a Christian conservative voter guide and from X itself. GPT leaned on Wikipedia, the Associated Press, and the news site Axios.

Search mattered more than training data. Parikh's team turned off web search for a smaller set of 30 races. GPT still named the Democrat in 24 of those races. Grok's Republican lean fell apart, and it turned down the question in 27 of the 30.

A Reddit thread shared the story under r/artificial. One reader asked what in the Republican platform earned the pick, given promises broken since 2024. Another reader linked back to a 2025 news story about Grok posting racist content.

The study lands two weeks before the Nov. 3 midterms. Millions of people now ask chatbots for quick answers on topics they used to research on their own. Two major models gave opposite advice on nearly every single race tested. Both firms kept that split quiet from their own users.

A hobbyist on Reddit shared a small AI model this morning. It writes code and builds 3D scenes at the same time. The model holds 4 billion parameters. That is tiny next to most frontier AI.

Parameters are the settings inside a model that store what it learned. More parameters usually mean a smarter model. A smarter model also runs slower and costs more. A 4 billion parameter model can run on a single home graphics card.

The builder, who posts as ccore, started from Qwen3.5-4B. That is a small open model from Alibaba's Qwen team. ccore trained it for coding and for 3D scenes. Small models have always struggled with that task.

The model reads up to 262,144 tokens of context at once. A token is a chunk of text, roughly a word or part of one. That much context lets the model hold a long chat. It also keeps a big piece of code in view from the start.

ccore posted sample images built by the model: a garden, a desk, a bike, a sunset village. The files sit free for download on Hugging Face, in two compressed formats. The model picked up 13 likes there within hours.

The Reddit thread drew 27 comments fast. Several readers asked for a version trained on the 3D software Blender instead. One reply said that request misses the point. A narrow, focused model beats a general one stretched thin, the commenter argued.

Training is still running, and ccore promised updates this week. The project shows how far one builder can push a small, free model with a narrow focus. The work needed a laptop and some patience instead of a lab's budget.

A Reddit user named MaxDev0 solved a problem that blocks most home AI setups. Big models usually need more memory than a home computer holds.

Qwen3.8-Flash-Next is one of those big models. It uses a design called mixture of experts, or MoE. That design splits the model into thousands of small expert parts. It wakes up only a few experts per answer.

Most setups keep every expert loaded in system memory. That takes 48 to 64 gigabytes of RAM. A laptop or a normal desktop holds far less than that. MaxDev0 built a tool called Lean-QwenFlash to fix this gap. It pulls experts straight from a solid state drive instead.

The result runs on just 16 gigabytes of video memory. It also needs only 8 gigabytes of system RAM. Generation speed lands near 20 tokens per second. That speed feels fast enough for live, typing-speed replies. The drive type matters a lot here. Newer NVMe drives read data far faster than older hard drives do.

Qwen3.8-Flash-Next itself is already a hit on Hugging Face, the site that hosts open AI models. It has pulled more than 3 million downloads since its release in August. It also carries 6,471 likes on the site today.

The Reddit thread reached 17 comments within hours. Commenters compared notes on drive speeds across brands. Some asked if the trick works on older laptops with slower storage. One person asked for a step by step guide to try that same night.

The pattern keeps repeating in this corner of AI. A lab in Beijing or Hangzhou ships a giant free model. Someone with an ordinary gaming PC then spends a weekend making it run anyway. The point is simply to prove it can be done.