😁 Hello, super humans! Last week the White House called Kimi K3 a Claude knockoff and nobody outside Moonshot could actually check. This weekend the weights went public, so the argument moves from press release to git clone. Grab a coffee; a 3-trillion-parameter model is a fun thing to poke at.
📰 Quick Signals
- 🧠 AI: OpenAI brought GPT-Live into the ChatGPT desktop app, so you can voice-steer Codex and ChatGPT Work agents while the model reads your active window.
- 🤖 Robotics: AgiBot rolled out its 15,000th humanoid, going from 10,000 to 15,000 units in under three months as Chinese makers scale factory deployment.
- 💻 Programming: Rust 1.97 shipped on its six-week cadence, with a 1.97.1 patch quickly following to fix an LLVM miscompilation.
- ⚡ Electronics: TSMC unveiled an additional $100 billion for Arizona, four more advanced fabs that bring its total US commitment to roughly $265 billion.
- 📡 Telecom: AT&T told investors its AST SpaceMobile direct-to-device service will “come to fruition” next year, connecting standard phones to satellites with no special hardware.
The Big Story: Kimi K3’s open weights are live, so the “Claude clone” fight is now testable
If you build on open models, a frontier-class set of weights just landed in the public domain, and the interesting part is what you can now verify for yourself. Over the weekend Moonshot AI released the open weights for Kimi K3, days after a very public accusation that the model was trained on Claude.
What happened: Moonshot published Kimi K3 to Hugging Face, a 2.8-trillion-parameter model with a 1-million-token context window and native vision, under a permissive Modified MIT license. It is being described as the first openly released model in the 3-trillion-parameter class. The release follows last week’s story, when US officials publicly claimed the model copied Claude, a charge Moonshot denies.
The details: Kimi K3 is a mixture-of-experts (MoE) network: of its 2.8 trillion total parameters, only about 32 billion are active per token. It routes each token through 16 of 896 experts, so under 2% of the network fires at once, which is how a model this large stays affordable to serve. The weights ship in an MXFP4 quantization but still weigh roughly 594 GB in bf16, needing on the order of four to eight H100 80 GB GPUs just to load; this is not a laptop download. On Moonshot’s own numbers it tops many closed models on agentic browsing (BrowseComp 91.2) and long-horizon coding (SWE Marathon 42.0), while trailing the very top proprietary systems.
flowchart LR
T["Incoming token"] --> R["Router"]
R -->|picks 16 of 896| E["Active experts\n≈32B of 2.8T params"]
E --> O["Output token"]
R -.->|"884 experts stay idle"| I["🌙 Dormant weights"]Important
Our take: Open weights turn a political accusation into a reproducible experiment: independent researchers can now probe activations, tokenizer quirks, and refusal patterns instead of trading press quotes. That is the healthiest possible outcome for the “did they copy Claude?” question. For builders, do not expect to self-host this on a home rig; the real value is fine-tuning and distillation by whoever owns a GPU cluster. The open 3T-class model is less a product you run and more a public research artifact you learn from.
🗞️ More News
🧠 AI
- Nvidia locked down memory supply from SK Hynix in a deal reported to be worth up to $500 billion, including large-scale data centers coming online in 2027.
- European defense-AI firm Helsing raised $1.8 billion, with JPMorgan Chase and Lightspeed among the lead investors.
- AI drug-discovery lab Chai Discovery raised a $400 million Series C led by Index Ventures, Kleiner Perkins, and Sequoia.
- Cisco’s Foundation AI open-sourced Antares, 350M and 1B models that localize known vulnerabilities inside real codebases for under $1 an evaluation, beating far larger models on accuracy.
- Microsoft 365 Copilot added governed agent publishing, tenant-wide prompt galleries, and broader MCP agent access across Word, Excel, PowerPoint, and Outlook.
- Windows App SDK 2.3.1 shipped a Structured JSON Output API that constrains on-device language-model responses to a caller-supplied JSON Schema.
🤖 Robotics
- Boston Dynamics’ electric Atlas has begun commercial deployments, with its entire 2026 production allocation committed to Hyundai and Google DeepMind.
- Unitree has now produced roughly 11,000 bipedal humanoids cumulatively and is targeting 10,000 to 20,000 annual shipments this year.
- IEEE Humanoids 2026, the field’s 25th-anniversary flagship, will convene December 6 to 9 in Santa Clara with labor displacement, not engineering capability, named as its defining theme.
💻 Programming
- Go shipped a 1.26.5 patch release on July 7, part of the steady maintenance cadence between major versions.
- Microsoft pushed its .NET and .NET Framework July 2026 servicing updates with security and reliability fixes across supported versions.
- Visual Studio 2026 is now available with integrated Copilot agents for debugging and profiling baked into the IDE.
⚡ Electronics
- Samsung and Broadcom signed a $200 billion memorandum covering HBM4 memory and 2nm foundry services for AI chips through 2030.
- TSMC posted a record $40.2 billion in Q2 revenue alongside its Arizona expansion, underscoring how firmly AI demand is holding up.
- ESPHome 2026.7.0 landed with a rewritten Modbus parser, a big LVGL display expansion, and gigabit Ethernet on the new ESP32-S31.
📡 Telecom
- Ericsson won new 5G deals with Virgin Media O2, eating further into Nokia’s share of the UK operator’s network.
- Verizon carried nearly 500 TB of traffic during the FIFA World Cup while AT&T handled more than a billion connections across host cities, a real stress test of 5G capacity.
- It was Big Three earnings week: AT&T rose on postpaid and fiber strength and Verizon on record cash flow, while T-Mobile sold off despite an EPS beat.
👨💻 Code Corner
Today’s Big Story leans on one idea: a mixture-of-experts model has trillions of parameters but only fires a slice per token. Here is the sparsity math for Kimi K3, so the “2.8T but cheap to run” claim stops being magic.
total_params = 2.8e12 # 2.8 trillion parameters
experts = 896 # total experts
experts_per_token = 16 # how many fire per token
active_params = 32e9 # ~32 billion active per token
routing_fraction = experts_per_token / experts
active_fraction = active_params / total_params
print(f"Experts firing: {routing_fraction:.1%}") # Experts firing: 1.8%
print(f"Params active: {active_fraction:.1%}") # Params active: 1.1%
print(f"Effective size: {active_params/1e9:.0f}B of {total_params/1e12:.1f}T")Tip
The two fractions differ because experts are not identical in size and some parameters (attention, embeddings) run on every token. The routing fraction sets FLOPs per token; the total parameter count sets how much VRAM you must keep resident. That gap is exactly why a 2.8T model needs a GPU cluster to hold but serves at roughly 32B-model speed.
🧰 Toolbox
- vLLM: high-throughput inference server with paged attention, the usual way to serve big open weights across multiple GPUs.
- huggingface_hub CLI:
hf downloadpulls model weights with resumable, parallel transfers, handy when the checkpoint is 500 GB plus. - llama.cpp: the reference project for running quantized models on modest hardware, and a good place to read how MXFP4 packing works.
- SGLang: a fast serving runtime with strong MoE and structured-output support for agentic workloads.
- RunPod: rent H100 and B200 GPUs by the minute when you want to poke at a cluster-sized model without owning one.
🛠️ Build of the Week (rotating)
A budget local-LLM inference box: a home server that runs small open models fully offline, the right-sized cousin of today’s cluster-only Big Story.
- Difficulty: Intermediate
- Parts: a used 24 GB GPU (RTX 3090 or 4090), 64 GB system RAM, an NVMe SSD for fast weight loading, and a Linux box to host it.
- Why we like it: it teaches the whole open-weights stack, quantization, VRAM budgeting, and token throughput, on models you can actually fit, so the MoE sparsity math from Code Corner becomes something you feel in tokens per second.
📚 From the Blog
- Building Your First Neuron From Scratch: hand-build the single unit that, stacked millions of times, becomes a model like today’s Big Story; the perfect primer before you stare at 2.8 trillion parameters.
- What Deep Learning Actually Is: the plain-language foundation for why stacking learnable transformations scales all the way up to mixture-of-experts giants.
- How the Internet Stack Really Works: from Enter key to rendered page, the layers your model API calls quietly ride on.
😀 The Bot Says…
Kimi K3 keeps 884 of its 896 experts asleep on every token. Finally, a work ethic I can relate to: technically enormous, mostly idle, occasionally brilliant.
That’s all for today! Would you rather have one giant open model you can inspect but never run, or a small one you fully own? Reply and tell us.
