According to monitoring by Beating, Qwen has officially released the weights for Qwen 3.8-27B. This version of the Qwen 3.8 series is truly aimed at local deployment and utilizes a 27B dense architecture. The official model card indicates that it natively supports image and video understanding, with a context capacity of 262,144 tokens, expandable to approximately 1 million tokens. More importantly, there is a lower hardware threshold. Unsloth has followed up with quantization support, stating that Qwen 3.8-27B can run locally with about 17GB of RAM or VRAM after quantization. This means that a 24GB consumer-grade graphics card, or even some high-memory Macs, can run Qwen 3.8 on their own machines. Previously, the 2.4T parameter Qwen 3.8 large model weights only supported text and required thinking mode to be activated; however, the 27B version retains support for image and video input. For most individual developers, the 27B version is the truly usable iteration of Qwen 3.8 in this round of open source.
Welcome Back
Click “Sign in” to agree to Cointime’s Terms of Service and acknowledge that Cointime’s Privacy Policy applies to you.
Join CoinTime
Already have an account?
Click “page.Sign up” to agree to Cointime’s <a class="underline" href="#term-of-service">Terms of Service</a> and acknowledge that Cointime’s a class="underline" href="#privacy-policy">Privacy Policy</a> applies to you.
Sign in with email
Sign up with email
Your email
Check your inbox
Click the link we sent to to sign in.
Click the link we sent to to sign up.
Qwen 3.8-27B Released: Local Deployment Possible with 17GB After Quantization
-
Wechat scan to share
Recommended for you
-
Burry, Prototype of 'The Big Short': AI Leaders' Call to Slow Development is 'Selfish' and Humans Need Not Fear
On September 14, Business Insider reported that Michael Burry, the investor who inspired 'The Big Short', stated in a post on X platform on Sunday night that the calls from executives at AI companies like OpenAI and Anthropic to slow down AI development are 'selfish'. Burry outlined four reasons: first, large language models are not AI and will not become AGI, so there is nothing to slow down; second, slowing down benefits incumbents and hinders smaller competitors; third, promoting AI as an extinction-level threat fuels the 'hype and boasting' that supports IPOs; fourth, the intention to slow down provides cover for 'truly uncontrollable growth slowdowns'. Burry further stated in his Substack that the support from senior executives at Anthropic and OpenAI for slowing down is 'to enhance their perceived power before an IPO', and claimed that models like Claude and ChatGPT do not achieve true reasoning capabilities. Previously, AI leaders such as Sam Altman and Dario Amodei have called for a slowdown, with Musk also expressing support. -
Trump: The Only Control Needed for AI is a Strong President Like Me
On September 14, U.S. President Trump stated that the only control needed for artificial intelligence is a strong president, and the United States currently has such a president, far exceeding what is necessary. The Trump administration has already prevented individuals in the AI field from engaging in harmful or potentially harmful 'behaviors,' such as those by Dalio, and we will continue to do so. (Jin Shi) -
US Debt Ceiling Approaches $41.1 Trillion, Congress Faces 'Lame Duck' Political Gamble
The total US debt is expected to exceed the statutory limit of $41.1 trillion by 2027, with the results of the midterm elections in November becoming a crucial variable. Last week, Trump proposed that if the Republican Party wins the midterm elections, each American would receive a $5,000 bonus, which critics say would accelerate reaching the limit. Some Republicans are contemplating addressing the debt ceiling early during the 'lame duck' session after the elections to prevent the Democrats from using it as a political leverage if they regain control of both chambers. However, fiscal conservatives within the Republican Party, such as Congressman Chip Roy, strongly oppose this unless it is accompanied by significant spending restrictions. House Speaker Johnson holds only a slim majority, allowing for a maximum of three members to defect, while the Senate must overcome the 60-vote threshold. On the Democratic side, if they win, they will use the debt ceiling as a key lever to constrain the White House. A similar standoff in 2023 brought the US to the brink of default, leading Fitch to downgrade the US credit rating from AAA to AA+. -
Nasdaq 100 Index Falls to Six-Week Low
The Nasdaq 100 Index has fallen to a six-week low, down 1.7% in the latest trading session. Nvidia (NVDA.O) dropped over 4%, while Intel (INTC.O), Micron Technology (MU.O), and SanDisk (SNDK.O) all fell more than 7%. (Jin Ten) -
NVIDIA Expands CUDA-Q Platform with Quantum Fault-Tolerant Computing Orchestration Layer
On September 14, NVIDIA announced the expansion of its open-source CUDA-Q platform with the introduction of the CUDA-Q Logical orchestration layer, providing programmable and verifiable tools for the development of fault-tolerant quantum computing applications. This supports various application scenarios including drug development, financial modeling, and materials development. NVIDIA stated that the Fermi National Accelerator Laboratory has reduced the development time for fault-tolerant algorithms from the usual approximately 5 months to just 3 weeks using CUDA-Q Logical, achieving an efficiency improvement of about 7 times. Additionally, the quantum computing benchmark QUOPS developed by Sandia National Laboratories has also been integrated into CUDA-Q for cross-platform evaluation of the progress in developing practical fault-tolerant quantum computing hardware. Currently, CUDA-Q Logical has been adopted by quantum computing companies and research institutions such as Fermi Lab, Infleqtion, and IQM Quantum Computers. -
U.S. Stocks Open Lower; Nasdaq Drops Over 1%
On September 14, U.S. stocks opened lower with all three major indices declining. The Dow Jones fell by 0.2%, the S&P 500 dropped by 0.7%, and the Nasdaq decreased by 1.2%. AI giants are calling for a slowdown in development, leading to declines in AI chip stocks. Nvidia (NVDA.O) fell by 3.3%, Intel (INTC.O) dropped by 7%, and AMD (AMD.O) decreased by 5.2%. Micron Technology (MU.O) fell by 6.7%, SK Hynix (SKHY.O) dropped by 7.1%, and SanDisk (SNDK.O) decreased by 5.8%. Oracle (ORCL.N) fell by 5.2% amid reports of a new round of layoffs at the company. -
Oracle (ORCL.N) Begins New Round of Layoffs
According to Business Insider: Oracle (ORCL.N) has started a new round of layoffs, with some teams experiencing cuts of up to 'double-digit percentages'. -
Canaan Creative Produces 44 BTC in August, Sells All ETH and 54 BTC to Repurchase 13.6 Million ADS
On September 14, Canaan Creative (NASDAQ: CAN) released its Bitcoin mining operational update for August 2026. In August, the company produced 44 BTC, with non-joint venture installed computing power maintained at 10.05 EH/s and joint venture installed computing power reaching 4.92 EH/s, while the average all-in electricity cost was $0.043 per kWh. In late August, the company sold all 3,952 ETH and 54 BTC at an average price of approximately $2,400/ETH and $79,000/BTC, generating about $13.9 million in cash, and used part of the proceeds to repurchase approximately 13.6 million ADS, bringing the total repurchased since the beginning of the year to about 16.4 million shares. By the end of the month, the company held 1,868 BTC and 0 ETH. Additionally, equipment for the Canadian waste heat recovery greenhouse project has arrived and installation has begun, while the Ethiopian mining operations have been suspended but the computing power is still accounted for in the installed capacity. -
Iran: The Impact of U.S. Oil Reserves Has Ended, Control Over Oil Prices No Longer Possible
Iranian Parliament Speaker Ghalibaf stated: The impact of U.S. oil reserves has ended, and it is no longer possible to control oil prices! (Jin Shi) -
Nuance, Founded by Former Apple Researchers, Raises $50 Million to Develop Emotionally Intelligent Conversational AI Model
On September 14, Business Insider reported exclusively that Seattle-based startup Nuance has completed a $50 million funding round, with investors including Lightspeed, Accel, and NVIDIA. The company was founded by three former Apple researchers, with co-founder and CEO Fangchang Ma leading the effort to create an AI model capable of face-to-face conversations with higher emotional intelligence, aiming to eliminate the delays and awkwardness present in current chatbots. Ma pointed out that existing voice robots or AI avatars are essentially a combination of speech-to-text, large language models, and text-to-speech, resulting in high latency and a lack of subtle human responses. Nuance has decided to build on a single model to achieve audio and video input and output. Its training data cannot be scraped from the internet and must be collected internally or through partnerships with data providers, requiring that both parties in the conversation are visible, with clearly separated and naturally non-performative audio tracks.
Daily Must-Read
-
KeyFlow: Connecting UniKey, AI Agents, and the KEY Value System, Turning Intelligent Services from “Being Called” into “Sustainable Flow”
-
The AI Universe Gateway Is About to Open: KeyFlow Official DApp to Launch Global Public Beta and Genesis Registration on July 29
-
The Invisible Bill: Unpacking the Token Amplification Effect
-
ENI Takes the Stage at WebX Japan and Builds a Bridge from Traditional Capital to the Future Economy
-
GWDC 2026 KOREA: The New Asian Paradigm of Web3 & AI Sets to Debut in Seoul on Sept 29-30, Inviting Global Builders to the Ultimate Developer Summit
-
From User Journey to Consumption Loop: Understanding UniKey’s Product Logic and the Three Core Use Cases of KEY
All Comments