The Nightly Build.
Tech News

Moonshot's Kimi K3 cuts the lag to frontier models to three to five months

Independent testing puts Kimi K3 in the same benchmark tier as Claude Fable and GPT-5.6 Sol Max, and finds a hallucination rate near 51 percent that Moonshot's own charts leave out.

6 min read2026-07-27
Open-Weight AI

Moonshot's Kimi K3 closes to within months of the frontier1

Kimi K3 activates roughly 50 billion of its 2.8 trillion parameters per token, routing each one through 16 of 896 experts under what Moonshot calls Kimi Delta Attention, which the company says improves scaling efficiency by roughly 2.5x.1 The model carries a Modified MIT license, a 1-million-token context window, and native tool calling, web browsing, and multi-step planning, and Moonshot released the weights on Hugging Face at 00:00 UTC on July 27.1 Independent researcher Nathan Lambert ranks K3 third on Artificial Analysis's Intelligence Index, behind only Claude Fable and GPT-5.6 Sol Max, and first in the Frontend Code Arena.2 He reads those rankings as closing the historical lag between open and closed frontier models, from six to nine months down to three to five.2 Artificial Analysis's own testing, reported independently by TechTimes, found a hallucination rate near 51 percent, up from 39 percent on the prior K2.6 generation even as factual accuracy rose from 33 percent to 46 percent, a tradeoff missing from Moonshot's own published charts.3 Anyone weighing K3 for an agentic deployment inherits that tradeoff: a model that answers more questions correctly and invents more of the ones it gets wrong, in close to equal measure.

Open-Weight AI

DeepSeek retires its old API names four days after V4 goes generally available4

DeepSeek's V4 family ships as two models: V4-Pro, with 1.6 trillion total parameters and about 49 billion active per token, and V4-Flash, with 284 billion total and about 13 billion active.4 Both carry a 1-million-token context window, MIT-licensed weights, and a Mixture-of-Experts design DeepSeek calls Manifold-Constrained Hyper-Connections.4 The company previewed V4 on April 24 and moved it to general availability on July 20, adding agentic capabilities, stronger math reasoning, and dedicated code-generation improvements at that point, according to independent reporting from Tech Insider.5 Four days later, at 15:59 UTC on July 24, DeepSeek shut off the two legacy endpoint names, deepseek-chat and deepseek-reasoner, forcing every integration still calling them onto V4 on a compressed timeline.5 V4-Pro's list price is $0.435 per million input tokens on a cache miss, pricing Tech Insider describes as 57 times cheaper than Claude Fable 5 on the same basis.45 That comparison holds only for published list prices. It says nothing about the quality gap DeepSeek would need to close for the price gap to matter to a buyer choosing between them.

Robotics

NVIDIA cuts simulated surgical-robot training from hours to minutes6

NVIDIA's new Medical Physics Simulation framework, released July 22 as an open-source module inside Isaac for Healthcare, runs 8,192 robot-training environments in parallel on GPU-native simulation.6 That cuts the time to train a surgical policy from more than five hours to under two minutes, the company says.6 It combines classical physics modeling via NVIDIA's Warp and Newton engines, for friction, contact, and rigid or flexible mechanics, with real-time generative physics from a system NVIDIA calls Cosmos-H Dreams.6 CMR Surgical demonstrated the framework alongside its Versius Plus system at the Society for Robotic Surgery conference on July 23.7 It is the clearest sign yet that a commercial surgical-robot maker plans to train on simulated anatomy rather than solely on cadavers, which are expensive and offer limited anatomical variety.7 The trade coverage of that demonstration describes the collaboration and the conference showing but stops short of any clinical data. There is no published timeline yet for when, or whether, a policy trained this way reaches a Versius Plus arm in an actual operating room.7 The speedup is real and measured in simulation. What it buys in surgery is still a demo, not a deployed system.

Quantum Computing

Hitachi and Intel target a 1,000-qubit silicon chip by 20308

Hitachi, Intel K.K., and Japan's AIST institute announced a NEDO-backed program on July 22 to move silicon quantum computing from academic demonstration to industrial fabrication, built on silicon spin qubits manufactured on Intel's 18A process.8 The roadmap sets three dated milestones: a cloud platform for remote access to the hardware in fiscal 2027, a 100-qubit error-correcting prototype in fiscal 2028, and a 1,000-qubit system built through 3D chip integration in fiscal 2030.8 Right now the program covers chip designs and manufacturing toolkits. No working qubits exist yet, and the earliest of those three milestones is still more than a year out.8 The announcement landed one day before IBM said it was acquiring HRL Laboratories, a competing bet on silicon-based quantum hardware, putting two of the industry's largest players behind silicon over superconducting or trapped-ion approaches in the same week.9 Whether silicon spin qubits scale the way both companies are now betting on is a question fiscal 2028's prototype, not this announcement, will start to answer.

Chips

AMD's Venice server chips ship first on TSMC's 2-nanometer process10

AMD's EPYC 9006 series, code-named Venice, is the industry's first high-performance computing processor built on TSMC's 2-nanometer node, and the company says its Zen 6 cores deliver a 70 percent CPU performance gain over the prior Zen 5 generation.10 The top configuration reaches 256 cores and 512 threads, with a 16-channel memory system and PCIe Gen 6 connectivity, and AMD is positioning the line for agentic AI, enterprise, and HPC workloads.10 Independent hardware coverage from TweakTown confirms the core count and process node but stops at AMD's own performance figures: no third-party benchmark of the 70 percent claim has been published yet.11 AMD says consumer desktop parts on the same Zen 6 architecture are coming, just not yet. For now, Venice ships only to the cloud providers and enterprise buyers AMD names as its first customers.11

Sources

  1. Hugging Face · Moonshot AI, "Kimi-K3" model repository and model card
  2. Interconnects.ai · Nathan Lambert, "Kimi K3: The Open Weights Escalation"
  3. TechTimes · "Kimi K3 Open Weights Drop July 27: Near-Frontier Coding, Undisclosed Hallucination Risk"
  4. DeepSeek · "DeepSeek V4" official product and specifications page
  5. Tech Insider · "DeepSeek V4 Reaches General Availability"
  6. NVIDIA Blog · "Medical Physics Simulation Framework Speeds Surgical Robot Training"
  7. MassDevice · "NVIDIA Unveils Simulation Framework for Surgical Robotics"
  8. Hitachi · "Hitachi Launches R&D on Silicon Quantum Computers in Collaboration with Intel K.K. and AIST"
  9. TechTimes · "Japan Funds Silicon Spin-Qubit Development Using Intel's 18A Chip Process"
  10. AMD Newsroom · "6th Gen AMD EPYC 9006 Series Processors Deliver Purpose-Built Solutions for Agentic AI, Enterprise and HPC"
  11. TweakTown · "AMD Announces 6th Gen AMD EPYC 'Venice' CPUs: Up To 256 Cores And 512 Threads, And Built For AI"