Lookonchain APP

App Store

Tencent Unveils Flash Model for the First Time: 224 Billion Total Parameters, 12 Billion Activated Parameters, Cost of Some Office Tasks Cut by Over 20 Times

34 minutes ago

Beating AI News: Tencent Hunyuan has officially launched its first Flash-named model, Hy4 flash, forming two tiers in the Hy4 series: a flagship version and the low-cost Flash variant. Hy4 flash has a total parameter count of 224 billion, including 64 billion Engram memory parameters, with 12 billion activated per inference. Compared to Hy4’s 770 billion total parameters and 49 billion activated parameters, Hy4 flash is significantly smaller in scale, yet still supports multimodal functionality and a 1 million-token context window. Engram is a conditional memory technology previously introduced by DeepSeek. It stores common text patterns in a quickly queryable memory table, enabling the model to retrieve information via table lookups and reduce redundant computations. In terms of pricing, Hy4 flash costs just 0.5 yuan per million input tokens, 2 yuan per million output tokens, and 0.02 yuan per cache hit. Compared to Hy4’s standard API list prices, these rates are 12x, 9x, and 15x lower respectively. However, Hy4 currently offers a limited-time 50% discount, so the actual price gap between the two models will be smaller than the list price difference. Tencent’s official test results show that while Hy4 flash’s PPT collaborative editing score dropped from Hy4’s 78.9 to 74.0, its per-task cost plummeted from 1.714 yuan to 0.105 yuan—roughly one-sixteenth of the original. The automation task score also only fell slightly, from 56.2 to 52.2, with per-task cost dropping from 0.908 yuan to 0.038 yuan. The performance gap is more pronounced for complex tasks: In the Terminal-Bench 4.0 terminal operation test, Hy4 flash scored only 19.2 points, versus the flagship Hy4’s 44.4. In the ProgramBench programming test, its score dropped from 30.0 points to 16.0 points.

Relevant content

He Yi posted his trading results showing a $100 million loss over 30 days, consoling himself that "spot holdings are no cause for worry".

Binance co-CEO He Yi recently made a rare trading update post on Binance Square. The data screenshot she shared showed her account had a $101 million loss over 30 days, with trading losses amounting to $97.32 million. In the accompanying text, He Yi wrote "Drawdown is temporary, spot is not afraid"—but the post was clearly self-deprecating. The phrase is a classic crypto trading meme; it was originally a serious "lesson learned" for traders. However, around 2024, as altcoins plunged endlessly and Bitcoin also saw unexpected declines, the line gradually evolved into a self-mocking joke.

1 minutes ago

Codex "model full" errors have dropped significantly: the percentage of affected users fell from 6.82% to 0.085%

Beating AI Express reports that OpenAI Core Product Lead Tibo Sottiaux shared an internal chart, indicating the team may have finally resolved the long-plaguing "model at full capacity" error that has troubled users. Data shows that on October 19, approximately 6.82% of Codex daily active users encountered at least one request ending with a capacity error. By November 12, this proportion had fallen to 0.085%. Previously, numerous users complained that even with sufficient subscription quotas, Codex would suddenly prompt that the model was at full capacity, requiring a model switch. Ongoing programming tasks would be interrupted directly, and switching models was not always effective. The fix primarily centers on the automatic retry mechanism. Between October 15 and November 6, OpenAI implemented 6 rounds of adjustments. When servers are overloaded, Codex is now notified of the appropriate wait time before retrying. The client, upon receiving the instruction, can wait and retry automatically, minimizing task termination. OpenAI also increased inference computing power ahead of this adjustment. Some netizens joked that the drop in Codex errors might be due to users flocking to Claude Opus 5.5.

1 minutes ago

South Korea's KOSPI index closed down 3.6%, with SK Hynix falling 7.89%.

According to Bitget market data, South Korea’s KOSPI index closed 190.28 points lower on Monday, November 16, down 3.6% to 5100.25 points. SK Hynix dropped 7.89%, while Samsung Electronics fell 4.46%.

1 minutes ago

Morgan Stanley forecasts gold prices will reach $5,250 in the fourth quarter of next year.

Morgan Stanley forecasts that the gold price will reach $5,250 per ounce by the fourth quarter of 2027.

1 minutes ago

22V Research analyst: Bitcoin is approaching a head and shoulders top pattern, a break below the $60,000 neckline could signal a decline to $38,000.

Bitcoin hit $62,000 over the weekend and rebounded to around $65,000 during Asian trading hours on Monday. John Roque, technical strategist at 22V Research, said in a Saturday research note that Bitcoin is nearing completion of the right shoulder of a head-and-shoulders pattern, and has not yet broken below the $60,000 neckline; a break would signal a risk of a drop to $38,000. Bitcoin started the year at around $90,000, briefly approaching $98,000 in January, fell back to the $60,000 range in February, and has been very close to its year-to-date low over the past week. Data from CoinGecko shows that Bitcoin, Ethereum, BNB, XRP, and Solana have all fallen by 13% or more in the past seven days. Matt Maley, chief market strategist at Miller Tabak, also said on Saturday that Bitcoin breaking below its 50-week moving average is "not a good sign" and that it must hold the $60,000 level (the lows from February and June); a sustained break below this level would be quite bearish. Additionally, the BVIV index, which tracks Bitcoin's 30-day implied volatility from Volmex Labs, rose to 49.75 over the weekend, its highest level since June.

1 minutes ago

Amber Group issued a statement denying false allegations, stating its operations are normal and reserving the right to pursue legal accountability.

Amber Group released a statement via its official account, noting that it has become aware of recent false allegations targeting the company and its management members, which were spread through anonymous letters and online platforms. The firm firmly denies these claims, adding that its operations are currently normal and all business activities are proceeding in an orderly manner. The statement further pointed out that the relevant content involves the dissemination of personal identity documents and other private information, which seriously violates personal privacy and crosses the proper bottom line. Amber Group said it respects legitimate public discussions that are fact-based and respect personal privacy, is preserving relevant evidence, and reserves all rights to take legal action and pursue liability for acts of defamation, unauthorized disclosure of the company's confidential information, and infringement of personal privacy.

1 minutes ago

Popular tokens

BitcoinEthereumHyperliquidSolanaTRONBNBTetherAaveXRPPepeFartcoinOndoJupiterUniswapBonkPendleEthenaArbitrumAvalancheLidoChainlinkPolygonDogecoinCardano