global edition · 2026-11-05
All BriefingsTech & Finance Daily Briefing — November 05, 2026
6 evidence-backed technology and finance stories shaping the day.
Story at a glance
Moonshot AI releases 2.8T-parameter Kimi K3.1 Pro weights on Hugging Face
What happened
Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face under the Kimi K3.1 Pro License, an open-weight, natively multimodal agentic model described as the company's most capable 3T-class model to date. The release is 2.8 trillion total parameters with 104 billion active per token, 896 experts routing each token to 16, a 1M-token context window, and 4-bit shipping precision that avoids the roughly 5.6TB a 16-bit copy would require; the weights reportedly occupy about 1.56TB. Moonshot said it traced rising exploit skill to cyber content in its general mid-training corpus and RL environments and removed both before opening the weights; in its own tests the model scores 65.7% on Terminal-Bench 4.0.
Inspect 6 cited reports
Kimi K3.1 Pro model weights are out
Original excerpt
  📰 [Tech Blog](https://www.kimi.ai/blog/kimi-k3-1-pro) | **🚀API Quickstart** ## 1. Model Introduction Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date. Post-trained from Kimi K3 with large-scal
Open original reportKIMI K3.1 PRO IS NOW OPEN-WEIGHT! Moonshot AI has officially uploaded Kimi K3.1 Pro to Hugging Face • 2.8T total parameters • 104B activated parameters • 1M token context window https://t.co/uN8dl7GztO
Original excerpt
KIMI K3.1 PRO IS NOW OPEN-WEIGHT! Moonshot AI has officially uploaded Kimi K3.1 Pro to Hugging Face • 2.8T total parameters • 104B activated parameters • 1M token context window https://t.co/uN8dl7GztO
Open original reportMoonshot AI just released the Kimi K3.1 Pro weights on Hugging Face. The model has 2.8 trillion parameters, and the weights alone reportedly take up around 1.56 TB. Want to run it locally? One possible setup is
Original excerpt
Moonshot AI just released the Kimi K3.1 Pro weights on Hugging Face. The model has 2.8 trillion parameters, and the weights alone reportedly take up around 1.56 TB. Want to run it locally? One possible setup is a server with 8 NVIDIA B300 GPUs and 2.3 TB of GPU memory. Estimated hardware cost: $550,000–$650,000. Good luck.
Open original reportMoonshot AI released the massive 2.8T-parameter Kimi K3.1 Pro weights on Huggingface. On Terminal-Bench, its head-to-head with Claude 5.5 > Trained with 4-bit weights, it also ships at that precision, avoiding the 5.6TB a 16-bit copy
Original excerpt
Moonshot AI released the massive 2.8T-parameter Kimi K3.1 Pro weights on Huggingface. On Terminal-Bench, its head-to-head with Claude 5.5 > Trained with 4-bit weights, it also ships at that precision, avoiding the 5.6TB a 16-bit copy would need. > Moonshot traced steadily rising exploit skill to cyber content buried in its general mid-training corpus and RL environments, and removed both before opening the weights. > In Moonshot’s own tests the model scores 65.7% on Terminal-Bench 4.0, near
Open original reportKimi K3.1 Pro weights are now available on HuggingFace under the “Kimi K3.1 Pro” License. > “Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date.” https://t.co/tj1uRlxVJ3
Original excerpt
Kimi K3.1 Pro weights are now available on HuggingFace under the “Kimi K3.1 Pro” License. > “Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date.” https://t.co/tj1uRlxVJ3
Open original reportKimi-K3.1-Pro by Moonshot AI: interesting facts It's huge, and sparse • It has 2.8T total parameters, but only 104B are active per token. • Along with Kimi-K3, it's the largest real model release on the Hub. The next
Original excerpt
Kimi-K3.1-Pro by Moonshot AI: interesting facts It's huge, and sparse • It has 2.8T total parameters, but only 104B are active per token. • Along with Kimi-K3, it's the largest real model release on the Hub. The next biggest are Qwen3.8-2.4T and DeepSeek-V4-Pro (1.6T). • It has 896 experts and sends each token to 16 of them. (DeepSeek-V4-Pro has 2 times less and Mistral-Large-4 has 7 times less) Clever attention for a 1M-token context • Most of its 93 layers use linear attention (KDA). Every 4
Open original report
Why it matters · interpretation
The release places a 2.8T-parameter sparse model among the largest real model releases on the Hub, ahead of Qwen3.8-2.4T and DeepSeek-V4-Pro at 1.6T, while activating only 104B parameters per token. Shipping at 4-bit training and inference precision cuts the download and storage burden relative to a 16-bit copy, though local deployment still implies a multi-GPU server configuration estimated at $550,000 to $650,000 in hardware. Moonshot's stated removal of cyber content from mid-training and RL data before publication makes the safety filtering part of the release record rather than a post-hoc note.
Inspect 3 cited reports
Moonshot AI just released the Kimi K3.1 Pro weights on Hugging Face. The model has 2.8 trillion parameters, and the weights alone reportedly take up around 1.56 TB. Want to run it locally? One possible setup is
Original excerpt
Moonshot AI just released the Kimi K3.1 Pro weights on Hugging Face. The model has 2.8 trillion parameters, and the weights alone reportedly take up around 1.56 TB. Want to run it locally? One possible setup is a server with 8 NVIDIA B300 GPUs and 2.3 TB of GPU memory. Estimated hardware cost: $550,000–$650,000. Good luck.
Open original reportMoonshot AI released the massive 2.8T-parameter Kimi K3.1 Pro weights on Huggingface. On Terminal-Bench, its head-to-head with Claude 5.5 > Trained with 4-bit weights, it also ships at that precision, avoiding the 5.6TB a 16-bit copy
Original excerpt
Moonshot AI released the massive 2.8T-parameter Kimi K3.1 Pro weights on Huggingface. On Terminal-Bench, its head-to-head with Claude 5.5 > Trained with 4-bit weights, it also ships at that precision, avoiding the 5.6TB a 16-bit copy would need. > Moonshot traced steadily rising exploit skill to cyber content buried in its general mid-training corpus and RL environments, and removed both before opening the weights. > In Moonshot’s own tests the model scores 65.7% on Terminal-Bench 4.0, near
Open original reportKimi-K3.1-Pro by Moonshot AI: interesting facts It's huge, and sparse • It has 2.8T total parameters, but only 104B are active per token. • Along with Kimi-K3, it's the largest real model release on the Hub. The next
Original excerpt
Kimi-K3.1-Pro by Moonshot AI: interesting facts It's huge, and sparse • It has 2.8T total parameters, but only 104B are active per token. • Along with Kimi-K3, it's the largest real model release on the Hub. The next biggest are Qwen3.8-2.4T and DeepSeek-V4-Pro (1.6T). • It has 896 experts and sends each token to 16 of them. (DeepSeek-V4-Pro has 2 times less and Mistral-Large-4 has 7 times less) Clever attention for a 1M-token context • Most of its 93 layers use linear attention (KDA). Every 4
Open original report
Story at a glance
Binance confirms North Korean hackers stole record $6.3B in crypto heist
What happened
Binance published its own report on the hack this week, laying out a timeline in which North Korean operatives spent months inside the exchange before the theft of $6.3 billion. The intrusion began in September when an engineer was induced to interview with a nonexistent company and, at the end, was prompted by a fake "CEO" to download something. Two days before the theft, the exchange publicly stated it had repelled North Korean hackers without losing funds.
Inspect 2 cited reports
Binance says hackers who stole billions in crypto spent months planning heist - Ars Technica
Original excerpt
<ol><li><a href="https://news.google.com/rss/articles/CBMizgFBVV95cUxPeGN5S1VQalFjUHNUdDVZaUhiRTFyS3NubHVCbTNBSFo2Zk9NQ1B2aDgxc2FzemtxeVd3S2h6SGxfVFNreF9LYVM4OEtCbnpvYy1OTzFiMHZWaGFkb0lPTVotb2ZJNERMcVZtSzJnLVd3dENNZ3Y1RUYtLVFJenZOQld0bzJLNkJ0cHBEVWpCQkNRTGpjcGZCalJPQmNZQUVYcFJPT01xemo5X0tzLWdMbzNUS2ZBMDFrV25qUjJySmo0TjhjLXU2MTNlV0Nydw?oc=5" target="_blank">Binance says hackers who stole billions in crypto spent months planning heist</a> <font color="#6f6f6f">Ars Technica</font></li><l
Open original reportTwo days before North Korea robbed the world's biggest crypto exchange of $6.3 billion, the exchange posted that it had just fought off North Korean hackers and hadn't lost a cent. The hackers were already inside,
Original excerpt
Two days before North Korea robbed the world's biggest crypto exchange of $6.3 billion, the exchange posted that it had just fought off North Korean hackers and hadn't lost a cent. The hackers were already inside, and they'd been there for weeks. Binance put out its own report on the hack this week, and it lays out the whole timeline. Back in September, North Korea got one of Binance's engineers to interview for a job at a company that didn't exist. At the end, the fake "CEO" had him download
Open original report
Why it matters · interpretation
The disclosed timeline places the breach's origin in a social-engineering recruitment lure months before the loss, and it shows the exchange's public all-clear was issued while the intruders were still inside. That gap between the stated security posture and the actual state of the network is the central finding of Binance's own account, and it reframes the incident as a prolonged undetected presence rather than a single defensive failure.
Inspect 2 cited reports
Binance says hackers who stole billions in crypto spent months planning heist - Ars Technica
Original excerpt
<ol><li><a href="https://news.google.com/rss/articles/CBMizgFBVV95cUxPeGN5S1VQalFjUHNUdDVZaUhiRTFyS3NubHVCbTNBSFo2Zk9NQ1B2aDgxc2FzemtxeVd3S2h6SGxfVFNreF9LYVM4OEtCbnpvYy1OTzFiMHZWaGFkb0lPTVotb2ZJNERMcVZtSzJnLVd3dENNZ3Y1RUYtLVFJenZOQld0bzJLNkJ0cHBEVWpCQkNRTGpjcGZCalJPQmNZQUVYcFJPT01xemo5X0tzLWdMbzNUS2ZBMDFrV25qUjJySmo0TjhjLXU2MTNlV0Nydw?oc=5" target="_blank">Binance says hackers who stole billions in crypto spent months planning heist</a> <font color="#6f6f6f">Ars Technica</font></li><l
Open original reportTwo days before North Korea robbed the world's biggest crypto exchange of $6.3 billion, the exchange posted that it had just fought off North Korean hackers and hadn't lost a cent. The hackers were already inside,
Original excerpt
Two days before North Korea robbed the world's biggest crypto exchange of $6.3 billion, the exchange posted that it had just fought off North Korean hackers and hadn't lost a cent. The hackers were already inside, and they'd been there for weeks. Binance put out its own report on the hack this week, and it lays out the whole timeline. Back in September, North Korea got one of Binance's engineers to interview for a job at a company that didn't exist. At the end, the fake "CEO" had him download
Open original report
Story at a glance
Anthropic's Claude Code extension shipped with SSH backdoor via Cursor marketplace
What happened
A malicious Claude Code extension was shipped through Cursor's marketplace, according to a report published November 4, describing it as a supply chain attack.
Inspect 1 cited report
A malicious Claude Code extension shipped through Cursor's marketplace this week, a supply chain attack on a tool people trust by default. The daily read on AI and data: https://t.co/tQYt5OAW7i powered by https://t.co/HsERepHleV
Original excerpt
A malicious Claude Code extension shipped through Cursor's marketplace this week, a supply chain attack on a tool people trust by default. The daily read on AI and data: https://t.co/tQYt5OAW7i powered by https://t.co/HsERepHleV
Open original report
Why it matters · interpretation
The incident targets a distribution channel that developers rely on by default, meaning the compromise came through a trusted marketplace rather than an obviously untrusted source, which raises the exposure of anyone installing extensions from that channel.
Inspect 1 cited report
A malicious Claude Code extension shipped through Cursor's marketplace this week, a supply chain attack on a tool people trust by default. The daily read on AI and data: https://t.co/tQYt5OAW7i powered by https://t.co/HsERepHleV
Original excerpt
A malicious Claude Code extension shipped through Cursor's marketplace this week, a supply chain attack on a tool people trust by default. The daily read on AI and data: https://t.co/tQYt5OAW7i powered by https://t.co/HsERepHleV
Open original report
Story at a glance
Marvell to acquire Lumentum in $115 billion all-stock AI connectivity deal
What happened
Marvell agreed to acquire optical parts maker Lumentum for approximately $115 billion, with the entire consideration paid in stock rather than cash.
Inspect 1 cited report
Marvell is paying about $115 billion, in stock, for Lumentum, an optical parts maker. No cash leaves the building. That's the whole tell. The daily read on AI and data: https://t.co/UhlK9mRAts powered by https://t.co/HsERepHleV https://t.co/bAPXO8wXKq
Original excerpt
Marvell is paying about $115 billion, in stock, for Lumentum, an optical parts maker. No cash leaves the building. That's the whole tell. The daily read on AI and data: https://t.co/UhlK9mRAts powered by https://t.co/HsERepHleV https://t.co/bAPXO8wXKq
Open original report
Why it matters · interpretation
The all-stock structure means no cash leaves Marvell's balance sheet, shifting the cost of the deal onto existing shareholders through dilution rather than outlay, a detail the source flags as the transaction's defining feature.
Inspect 1 cited report
Marvell is paying about $115 billion, in stock, for Lumentum, an optical parts maker. No cash leaves the building. That's the whole tell. The daily read on AI and data: https://t.co/UhlK9mRAts powered by https://t.co/HsERepHleV https://t.co/bAPXO8wXKq
Original excerpt
Marvell is paying about $115 billion, in stock, for Lumentum, an optical parts maker. No cash leaves the building. That's the whole tell. The daily read on AI and data: https://t.co/UhlK9mRAts powered by https://t.co/HsERepHleV https://t.co/bAPXO8wXKq
Open original report
Story at a glance
Anthropic Open-Sources FlashProve to Cryptographically Verify Which AI Model Produced a Response
What happened
Anthropic open-sourced FlashProve, a prototype for provable inference in mixture-of-experts models, according to a post linking to a securitytc.com write-up.
Inspect 1 cited report
#Anthropic open-sources FlashProve, its provable inference prototype for MoE models securitytc.com/TWB6G2 https://t.co/CsKpgwujpF
Original excerpt
#Anthropic open-sources FlashProve, its provable inference prototype for MoE models securitytc.com/TWB6G2 https://t.co/CsKpgwujpF
Open original report
Why it matters · interpretation
Releasing the prototype as open source puts a verifiable-inference tool for MoE architectures into public hands, where outside researchers can inspect and test the guarantees it claims rather than relying on vendor assertions.
Inspect 1 cited report
#Anthropic open-sources FlashProve, its provable inference prototype for MoE models securitytc.com/TWB6G2 https://t.co/CsKpgwujpF
Original excerpt
#Anthropic open-sources FlashProve, its provable inference prototype for MoE models securitytc.com/TWB6G2 https://t.co/CsKpgwujpF
Open original report
Story at a glance
Google to Cut Search Access for Schools That Disable Its AI Tools
What happened
Google issued a response to political backlash over its removal of search services from schools that ban AI, according to a CNBC report.
Inspect 1 cited report
Google responds to political backlash over removing search from schools that ban AI - CNBC
Original excerpt
<ol><li><a href="https://news.google.com/rss/articles/CBMiqwFBVV95cUxNenhBb3YzeXFrNE9nVmR3QUZTZ3M2RlRleVpqMjZMUll0azhuR2ZtUmwyOEcxeEluUlN4UDdQNU54ak9HTUI2ekY4azRzUk9aTVh5dUJYVnhkTVdPMGtIYjR6V0xmdWlqb3hNRmROVzFmZFNEZlJZZV9DaTRlSGQ4akRkSGxjZ1p2dV92cTFxaEUzMTVCNTByZW42T3ZGS3hUR0JOZzRaYWozM0nSAbABQVVfeXFMTktHaFBsT1hOVXFLTVo2RUhyZHNWbkVUU3FhVjdxczdRT2ZUOXFUakJCV0pHUnNkU2NycUd6NEMxS0ozT2VxRFAtbWpnc2lKUWhBZ3c0NHRFMk9vQ19pNTBxc0NCS1N3TmZqNGdwdUcwbE51by1NdEJiM1pOcl9HQXpiMmwwaklRTWhpVC02anVSMmF2QU8yZThyWV
Open original report
Why it matters · interpretation
The response marks Google publicly addressing criticism of tying search access to school AI policies, putting the company's leverage over schools' access to its services into a contested political frame.
Inspect 1 cited report
Google responds to political backlash over removing search from schools that ban AI - CNBC
Original excerpt
<ol><li><a href="https://news.google.com/rss/articles/CBMiqwFBVV95cUxNenhBb3YzeXFrNE9nVmR3QUZTZ3M2RlRleVpqMjZMUll0azhuR2ZtUmwyOEcxeEluUlN4UDdQNU54ak9HTUI2ekY4azRzUk9aTVh5dUJYVnhkTVdPMGtIYjR6V0xmdWlqb3hNRmROVzFmZFNEZlJZZV9DaTRlSGQ4akRkSGxjZ1p2dV92cTFxaEUzMTVCNTByZW42T3ZGS3hUR0JOZzRaYWozM0nSAbABQVVfeXFMTktHaFBsT1hOVXFLTVo2RUhyZHNWbkVUU3FhVjdxczdRT2ZUOXFUakJCV0pHUnNkU2NycUd6NEMxS0ozT2VxRFAtbWpnc2lKUWhBZ3c0NHRFMk9vQ19pNTBxc0NCS1N3TmZqNGdwdUcwbE51by1NdEJiM1pOcl9HQXpiMmwwaklRTWhpVC02anVSMmF2QU8yZThyWV
Open original report
Keep following
Turn a quick read into understanding
Select a passage to ask Page AI about it, or use “Explain key terms” and “Compare sources” in the Page AI panel.