Lookonchain APP

App Store

Qwen2 7B Modified to Diffusion Agent: Speed Surges to Over 500 Tokens Per Second, Vendor Self-Test Shows Single Metric Outperforms Sol

42 minutes ago

Beating AI Express reports: Celeris has launched Celeris-1 Magnus, a hybrid diffusion Agent model modified from Qwen3.8-27B. The company states the model can achieve over 500 tokens per second in many scenarios. On 97 bank customer service tasks from τ³-Banking, Magnus recorded a 41.2% completion rate when low reasoning was enabled, with a median latency of 55 seconds. For comparison, GPT-5.6 Sol delivered a 38.1% completion rate and a 79-second median latency in the same test. Traditional models generally generate tokens one at a time, while Magnus can produce multiple tokens in a single pass before iteratively correcting them. For content requiring strict sequential generation, it processes in parallel, enabling more tokens to be advanced per computation step.

Relevant content

Tailscale open-sources Tailcat: No login or IP assignment required, enabling direct tunneling between two machines

From Beating AI Express: Tailscale has open-sourced Tailcat, a tool that allows two machines to connect directly without signing up for an account or joining the same Tailnet—all you need is a single address string. It retains Tailscale’s WireGuard encryption, NAT traversal, and DERP relay functionality, but eliminates the need for a Tailscale account or control plane. Running Tailcat on one machine generates a connection address string; sharing this with another machine enables file transfers, port forwarding, or SSH access. Connections first use DERP to locate the peer, then attempt a direct peer-to-peer UDP connection; if hole punching fails, traffic falls back to the DERP relay. The entire process runs in user space, with no modifications to system routes or DNS settings. Tailscale lead engineer Brad Fitzpatrick first built a Tailcat prototype back in 2023, but the project was largely sidelined. He revived it recently as AI agents began connecting en masse to sandboxes, remote VMs, and test machines. Fitzpatrick uses Tailcat to let agents connect temporarily to Raspberry Pi devices, EC2 instances, and Windows hosts—no pre-configured accounts or network access for the agents required. Tailcat is not a full replacement for Tailscale: it lacks user identities, permission policies, and device management, and its connection address must be treated like a password. Tailscale-hosted Tailcat DERP relays log metadata, have rate limits, and do not offer uptime SLAs; for greater stability or privacy, users can deploy their own DERP relays.

1 seconds ago

UK National Crime Agency freezes $13.6 million in English Premier League-related funds tied to crypto game Sorare.

UK National Crime Agency (NCA) has frozen over $13.6 million in funds belonging to the Premier League. The funds were held in a Barclays account under the name of "English Premier League Limited". The freezing order was issued by Westminster Magistrates' Court in January 2025 under the Proceeds of Crime Act, aimed at preventing the funds from being transferred during the investigation. The sum is the first installment of a four-year sponsorship agreement between crypto fantasy football card game Sorare and the Premier League. The two parties announced their partnership in January 2023, with the deal set to expire after the 2025-26 season; the total contract value was estimated at around $163 million at the time. Meanwhile, Sorare is facing criminal proceedings from the UK Gambling Commission. The regulator accused Sorare in 2024 of offering gambling services without an operating license, with the trial scheduled to begin in June 2027. Sorare denies its products qualify as gambling under UK law, stating that game outcomes depend primarily on player skill rather than luck. No evidence has yet emerged linking the Premier League to any illegal activity. The top-flight league previously declined to comment on the fund freeze and plans to apply to adjust the existing freezing order.

1 seconds ago

Wall Street turns to AI to slash legal fees: Beyond saving man-hours, firms are also lowering their fee quotes.

Beating AI News (Insight) | Goldman Sachs, Morgan Stanley, and Citigroup are now requiring major U.S. law firms to directly incorporate efficiency gains from AI into their service quotes. When billing by the hour, AI typically cuts 10 hours of work to just 4, which would naturally lower bills. But the banks are pushing for deeper price cuts: they want firms to calculate AI-related cost savings upfront and submit lower total prices accordingly. Citigroup has already asked law firms to bid on projects, explicitly stating how much AI would reduce their costs. Morgan Stanley plans to source most external legal work via bidding by the end of this year, and adopt more fixed-price arrangements. Goldman Sachs is also pressing law firms to quantify exactly how much AI has improved operational efficiency. This move directly impacts the long-standing revenue model of big law firms, which relies on "hours worked × hourly rate". While AI can accelerate tasks like research, contract review, and litigation document retrieval, hourly rates at top U.S. law firms continue to rise. Wall Street now aims to deduct the cost savings from AI directly from legal fees.

1 seconds ago

Owlto Finance is offering 10,000 free gas subsidy slots for Robinhood.

According to official announcements, Owlto Finance is sponsoring gas fees for Robinhood users and has opened 10,000 free gas subsidy slots. Eligible users include Owlto users, OWL token holders, ARB token holders, and Robinhood trading users. OWL users qualify if they meet either of the following criteria: having completed any activity on Owlto such as bridging, swapping, contract deployment, daily check-in, or staking, or holding more than 100,000 OWL tokens. Each wallet can claim the subsidy at most once every 24 hours. For Robinhood users, they must hold more than 10,000 ARB tokens or have conducted transactions on Robinhood Chain, with each wallet limited to one claim only.

1 seconds ago

Cango's Q2 revenue reaches $50.8 million, its AI infrastructure business enters the commercialization phase.

According to official announcements, NYSE-listed company Cango reported Q2 2026 revenue of $50.8 million, with BTC mining contributing $47.4 million and other business income accounting for $3.4 million. As of June 30, the company held 1,056 BTC, with cash and cash equivalents of $10.1 million and long-term debt of $31.2 million. In terms of mining operations, Cango is continuing to phase out less efficient S19 series mining rigs, partially shifting to a hash rate rental model. As of June 30, the company’s self-operated hash rate stood at 19.84 EH/s, with rental hash rate at 7.74 EH/s, bringing total operational hash rate to 27.58 EH/s. In Q2, the company mined a total of 656 BTC, with an average cash mining cost of $73,313 per coin, down approximately 5% from Q1. The company has also begun implementing a BTC hedging strategy to mitigate the impact of price volatility on operating cash flow. On the AI infrastructure front, Cango’s LN mining site in Georgia, U.S., completed renovations in early July, now supporting a maximum capacity of 3 megawatts with reserved space for future expansion. Containerized equipment has been fully installed, while GPU hardware is arriving in batches. The company has signed its first AI customer contract, marking the official shift of its business from infrastructure construction and technical verification to the commercialization stage, with related revenue expected to be recognized starting in Q3. Additionally, Cango plans to develop two models simultaneously: bare-metal GPU hosting and colocation services, though no formal colocation contracts have been signed to date. The company has also deployed test nodes in Texas and the U.S. West Coast to meet customers’ demand for localized deployment, and is evaluating more potential sites and the possibility of building its own facilities.

1 seconds ago

Bitcoin’s Kimchi Premium reappears, marking its longest duration since early May.

Recently, Bitcoin has re-emerged with a price premium in the South Korean market. On Tuesday, Bitcoin quoted in South Korean won on Upbit was once about 1% higher than the corresponding price on Binance, a gap that has persisted for roughly a week, marking the longest such stretch since early May. For most of the past months, Bitcoin prices in South Korea have traded below overseas markets; in early June, the discount reached 3.1%, and the overall average discount in August stood at around 0.25%. The return of this premium signals a recovery in risk appetite among South Korean investors.

1 seconds ago

Popular tokens

BitcoinEthereumHyperliquidSolanaTRONBNBTetherAaveXRPPepeFartcoinOndoJupiterUniswapBonkPendleEthenaArbitrumAvalancheLidoChainlinkPolygonDogecoinCardano