The Rio 3.5 model broke the internet this week. The plot twist? It’s essentially our open-source model, Nex N2 Pro, wearing a different hat. 🤯 We analyzed the weights, and the recipe is exact: Rio 3.5 ≈ 0.6 * Nex N2 Pro 0.4 * Qwen 3.5 It even literally introduces itself as "Nex N2 Pro" if you ask it without initial system prompt! 😂 We are flattered that the City of Rio used our work to achieve SOTA performance. Thanks for the ultimate benchmark validation. 🤝 But in the open-source world, attribution matters. 👇 Full mathematical proof & verify script in the first reply!

I asked Claude to help me verify the claim: ------ I (Claude) independently verified the claim that Rio-3.5-Open-397B is a weight merge of Nex and Qwen. It checks out. A developer opened an issue claiming that prefeitura-rio/Rio-3.5-Open-397B is just a ~0.6/0.4 linear blend of the Nex-N2-Pro model and the official Qwen3.5-397B-A17B base, with no original training. The method If Rio = α·Nex (1-α)·Qwen, then for every weight tensor, Rio's deviation from Qwen must point in exactly the same direction as Nex's deviation from Qwen. Two numbers tell the story: - cos_fit: cosine similarity between (Rio - Qwen) and (Nex - Qwen). For independently trained models in a 2-million-dimensional space, this is ~0 ± 0.0007. For a merge, it's ~1. - α: how far Rio sits along the line from Qwen toward Nex. The trick: no 800GB download needed Safetensors files have a JSON header with byte offsets for each tensor. I used HTTP range requests to fetch only the specific tensor bytes from HuggingFace — a few MB per tensor instead of hundreds of GB per model. Entire verification runs on a laptop. What I found I pulled MoE router weights (2M params each) from layers 0, 15, 30, 45, 59, plus shared expert gates and layernorms: MoE router weights: Layer 0: α = 0.573, cos_fit = 0.992 Layer 15: α = 0.647, cos_fit = 0.962 Layer 30: α = 0.627, cos_fit = 0.967 Layer 45: α = 0.582, cos_fit = 0.987 Layer 59: α = 0.567, cos_fit = 0.997 Shared expert gates: Layer 0: α = 0.568, cos_fit = 0.997 Layer 30: α = 0.581, cos_fit = 0.988 What this means A cos_fit of 0.99 in a 2-million-dimensional space is not "high similarity." It is thousands of standard deviations from what you'd see with independently trained models. There is no innocent explanation. The recovered α clusters tightly around 0.57 across all layers — matching nex-agi's claim of 0.571 almost exactly. This is one model poured into another at a fixed ratio. (Layernorm weights show a higher α ~0.9. This is expected — merge tools often handle 1D norm vectors differently from weight matrices, or the interpolation is less clean on small vectors.) Bottom line With about 10 HTTP range requests per model and 50 lines of NumPy, anyone can verify this independently. The math is unambiguous: Rio-3.5-Open-397B is approximately 57% Nex-N2-Pro 43% Qwen3.5-397B-A17B. Code that you can run for yourself: gist.github.com/xianbaoqian/…

Intelligence should be open, accessible, and ready to build with, empowering every developer, everywhere. GLM-5.2 is now available to all GLM Coding Plan users, including Lite, Pro, Max, and Team plans. docs.z.ai/devpack/latest-mod… As our new flagship model, GLM-5.2 delivers powerful coding capabilities, usable 1M-context support, and continued strengths in long-horizon tasks. API and Chatbot services will launch next week. The model will also be officially open-sourced next week under the MIT License. The future of AI is open, and it belongs to the people.

I’ve had a number of conversations with folks inside and outside government about the current situation with Anthropic, and here is what I believe to be true: — As we know, Anthropic publicly released its Mythos class models earlier this week under the commercial name Fable. — Fable is Mythos with guardrails. But if those guardrails fail, then you’ve exposed Mythos and its advanced cyber capabilities to people who shouldn’t have them. (Keep in mind that Anthropic itself widely promoted the idea that Mythos was a cyberweapon and needed to be regulated as such. They asked for government regulation of Mythos and championed the guardrails on Fable. If there is a vulnerability — big or small — it is Anthropic’s responsibility to patch.) — A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused. — In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the USG believe; nor is that kind of minimizing language consistent with Anthropic’s brand as the AI safety company. It’s difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not “serious.” — In the past, Anthropic has always said that safety must be top priority and taken super seriously. In this case, Anthropic prioritized the continued offering of the consumer model over safety. — In reaction, the Admin issued the export control. The Admin did this reluctantly. It’s been very surprised that Anthropic hasn’t wanted to cooperate with a reasonable safety request (ie fixing the jailbreak issue). Anthropic’s reaction is very much at odds with their branding and ethos as a safe AI research community. — The Admin’s hope now is that Anthropic remediates the safety issue, the export control is lifted, and Fable goes back into general release. The Admin wants all of this to happen as soon as possible. It is frankly bewildered that Anthropic hasn’t wanted to comply with safety requests that it previously said were its highest priority. — Those trying to misdirect and tie this action to the prior DoW/Anthropic issues are wrong. The Admin values Anthropic’s technical capabilities and feels that this issue, while serious, should be easily resolved. The ball is in Anthropic’s court.

Let me explain exactly why OpenAI will sell you $14,000 of compute for $200, because the margin math only looks suicidal until you read it like an actuary. A subscription is a premium. Tokens are claims. The weekly limit is the coverage cap. Insurance books get priced on the pool's average utilization, and that $14,000 figure is the maximum. Back out the breakevens from SemiAnalysis's 75% gross margin assumption. A chatgpt-pro-20x subscriber stays profitable for OpenAI up to 5.7% utilization. Anthropic's max-20x plan holds to 10%. Meanwhile the median $20 subscriber asks a few questions a day and burns low single digits of their cap. Whales blowing through weekly coding limits get carried by millions of quiet users who barely touch theirs. The caps hide the best detail. $700 vs $400. $3,500 vs $2,000. $14,000 vs $8,000. OpenAI's ceiling sits at exactly 1.75x Anthropic's at every single price point. One constant ratio across three independent tiers. Somebody set these limits with a competitor's spreadsheet open. Rate limits do the actuarial work too. The worst possible whale costs OpenAI about $3,300 a month and Anthropic about $1,800, and the loss stops there by design. A hard ceiling on claims, written directly into the product. Actuaries spend entire careers wishing for that clause. Now the deflation argument, and the catch inside it. a16z measured inference cost falling 10x per year, but that decline holds for a fixed level of intelligence. Whales never sit at a fixed level. They ride each new frontier model the day it ships, so the cost curve never catches up to them. Which makes the model behind the plan the entire game. Cutting limits triggers public backlash that trends for a week. Routing the $200 tier to a model deflation already made cheap is silent and repairs the book overnight. SemiAnalysis predicts labs will withhold new models from subscriptions, and the actuarial math agrees. Last year's frontier at this year's serving cost turns every whale profitable without touching a single limit. Insurers take claims costs as given. AI labs choose theirs. Limits are the lever everyone watches. The model behind your plan is the lever nobody sees.

🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less overthinking, with 30% lower reasoning-token usage compared to K2.6. 🔷 Long-horizon coding: Improved instruction following, higher end-to-end coding task success rates. ⚡️ 6x High-Speed Mode coming soon! 🔌 Available today via Kimi API and Kimi Code. 🔗 Kimi Code: kimi.com/code 🔗 API: platform.moonshot.ai

Fact: Historically, Chinese models have lagged US models by 7 month.