7*24

Hot topic

Market

Activities

On April 1, Bloomberg ETF analyst Eric Balchunas revealed that BlackRock's upcoming iShares Bitcoin Premium ETF will have the trading code BITA. This highly anticipated sequel to BlackRock's spot Bitcoin ETF IBIT has not yet disclosed its management fees, which are expected to be 38 basis points.

[BlackRock's Bitcoin Premium ETF Trading Code is BITA]On April 1, Bloomberg ETF analyst Eric Balchunas revealed that BlackRock's upcoming iShares Bitcoin Premium ETF will have the trading code BITA. This highly anticipated sequel to BlackRock's spot Bitcoin ETF IBIT has not yet disclosed its management fees, which are expected to be 38 basis points.

BlackRock's Bitcoin Premium ETF Trading Code is BITA

According to Bloomberg, sources indicate that Usman Ahmad, the CEO of Standard Chartered's cryptocurrency trading firm Zodia Markets, has left the company. The Zodia Markets website shows that Nick Philpott, who previously served as the company's head of partnerships, has been appointed as interim CEO. So far, spokespeople for both Zodia Markets and Standard Chartered have declined to comment on the matter.

[Bloomberg: CEO of Standard Chartered's Cryptocurrency Firm Zodia Markets Has Departed]According to Bloomberg, sources indicate that Usman Ahmad, the CEO of Standard Chartered's cryptocurrency trading firm Zodia Markets, has left the company. The Zodia Markets website shows that Nick Philpott, who previously served as the company's head of partnerships, has been appointed as interim CEO. So far, spokespeople for both Zodia Markets and Standard Chartered have declined to comment on the matter.

Bloomberg: CEO of Standard Chartered's Cryptocurrency Firm Zodia Markets Has Departed

On April 1, Cango Inc. (NYSE: CANG), a Bitcoin mining company, received a notice from the New York Stock Exchange (NYSE) informing it that it failed to meet the NYSE's continued listing price standards. As of March 9, 2026, the average closing price of Cango's Class A common stock had been below $1.00 per share for 30 consecutive trading days, violating the requirements of NYSE listing rule 802.01C. According to the regulations, Cango has a six-month period to rectify the situation from the date of receiving the notice. If, at the end of any calendar month during the remedy period, its closing price and the average closing price for the preceding 30 trading days are both at or above $1.00, compliance will be restored. Cango stated that it has notified the NYSE of its intention to rectify the situation and will continue to assess market conditions and feasible options. During the remedy period, Cango's Class A common stock will continue to trade normally on the NYSE. This notice does not affect the company's daily operations, obligations to the U.S. Securities and Exchange Commission (SEC), or other contractual obligations.

[Cango Receives Notice from NYSE for Non-Compliance with Listing Price Standards]On April 1, Cango Inc. (NYSE: CANG), a Bitcoin mining company, received a notice from the New York Stock Exchange (NYSE) informing it that it failed to meet the NYSE's continued listing price standards. As of March 9, 2026, the average closing price of Cango's Class A common stock had been below $1.00 per share for 30 consecutive trading days, violating the requirements of NYSE listing rule 802.01C. According to the regulations, Cango has a six-month period to rectify the situation from the date of receiving the notice. If, at the end of any calendar month during the remedy period, its closing price and the average closing price for the preceding 30 trading days are both at or above $1.00, compliance will be restored. Cango stated that it has notified the NYSE of its intention to rectify the situation and will continue to assess market conditions and feasible options. During the remedy period, Cango's Class A common stock will continue to trade normally on the NYSE. This notice does not affect the company's daily operations, obligations to the U.S. Securities and Exchange Commission (SEC), or other contractual obligations.

Cango Receives Notice from NYSE for Non-Compliance with Listing Price Standards

On April 1, according to The Wall Street Journal, asset management giant Franklin Templeton announced that it will acquire 250 Digital, a spin-off of venture capital firm CoinFund, to expand its cryptocurrency investment products. 250 Digital was spun off from CoinFund in January of this year.

[Franklin Templeton to Acquire CoinFund's Spin-off 250 Digital]On April 1, according to The Wall Street Journal, asset management giant Franklin Templeton announced that it will acquire 250 Digital, a spin-off of venture capital firm CoinFund, to expand its cryptocurrency investment products. 250 Digital was spun off from CoinFund in January of this year.

Franklin Templeton to Acquire CoinFund's Spin-off 250 Digital

On April 1, The Wall Street Journal reported that asset management giant Franklin Templeton announced it will acquire 250 Digital, a spin-off of venture capital firm CoinFund, to expand its cryptocurrency investment products. 250 Digital was spun off from CoinFund in January this year.

[Franklin to Acquire CoinFund's Spin-off 250 Digital]On April 1, The Wall Street Journal reported that asset management giant Franklin Templeton announced it will acquire 250 Digital, a spin-off of venture capital firm CoinFund, to expand its cryptocurrency investment products. 250 Digital was spun off from CoinFund in January this year.

Franklin to Acquire CoinFund's Spin-off 250 Digital

According to monitoring by 1M AI News, the AI cloud platform Together AI has open-sourced Aurora, an adaptive framework for speculative decoding based on reinforcement learning. Speculative decoding is the mainstream method for accelerating large model inference: it quickly predicts token sequences using a small 'draft model', which are then verified in parallel by a larger model, adopting hits directly and discarding misses for re-evaluation. The issue lies in the fact that the draft model is a static product trained offline; once production traffic drifts (due to model upgrades, changes in user demographics, or task type switches), prediction accuracy declines continuously, and retraining offline is both costly and delayed. Aurora redefines speculative decoding as a reinforcement learning problem: the draft model serves as the policy, the large model verifier acts as the environment, accepted tokens are positive rewards, and rejected tokens are negative feedback. The system consists of two decoupled components: the inference server, which processes requests normally and streams verification results into a buffer, and the training server, which asynchronously pulls data to update the draft model weights and hot-replaces them back to the inference side without interrupting service. In simulated traffic tests involving 40,000 requests across five domains (mathematical reasoning, Text-to-SQL, code generation, finance, and general dialogue), Aurora recovered acceptance lengths within about 10,000 requests after traffic domain shifts, achieving a speedup of 1.25 times over a well-trained static draft model. Notably, Aurora, which trains online from scratch, achieved an acceptance length of 3.08, surpassing the static baseline of 2.63 and the 'pre-trained then fine-tuned' baseline of 2.99, with a throughput stabilizing at 302.3 tokens per second, directly challenging the existing notion that 'speculative decoding must rely on large-scale offline pre-training'. In end-to-end tests, speculative decoding provided a 1.92 times speedup on Qwen3-Coder-Next (FP8) with a batch size of 1, and 1.63 times on MiniMax M2.5 (FP8). Aurora's predecessor, ATLAS, laid the foundation for the adaptive speculator, and this upgrade has resulted in a fully autonomous closed-loop system. The code has been open-sourced on GitHub.

[No More Offline Training for Draft Models: Together AI Open Sources Aurora for Speculative Decoding with Self-Learning]According to monitoring by 1M AI News, the AI cloud platform Together AI has open-sourced Aurora, an adaptive framework for speculative decoding based on reinforcement learning. Speculative decoding is the mainstream method for accelerating large model inference: it quickly predicts token sequences using a small 'draft model', which are then verified in parallel by a larger model, adopting hits directly and discarding misses for re-evaluation. The issue lies in the fact that the draft model is a static product trained offline; once production traffic drifts (due to model upgrades, changes in user demographics, or task type switches), prediction accuracy declines continuously, and retraining offline is both costly and delayed. Aurora redefines speculative decoding as a reinforcement learning problem: the draft model serves as the policy, the large model verifier acts as the environment, accepted tokens are positive rewards, and rejected tokens are negative feedback. The system consists of two decoupled components: the inference server, which processes requests normally and streams verification results into a buffer, and the training server, which asynchronously pulls data to update the draft model weights and hot-replaces them back to the inference side without interrupting service. In simulated traffic tests involving 40,000 requests across five domains (mathematical reasoning, Text-to-SQL, code generation, finance, and general dialogue), Aurora recovered acceptance lengths within about 10,000 requests after traffic domain shifts, achieving a speedup of 1.25 times over a well-trained static draft model. Notably, Aurora, which trains online from scratch, achieved an acceptance length of 3.08, surpassing the static baseline of 2.63 and the 'pre-trained then fine-tuned' baseline of 2.99, with a throughput stabilizing at 302.3 tokens per second, directly challenging the existing notion that 'speculative decoding must rely on large-scale offline pre-training'. In end-to-end tests, speculative decoding provided a 1.92 times speedup on Qwen3-Coder-Next (FP8) with a batch size of 1, and 1.63 times on MiniMax M2.5 (FP8). Aurora's predecessor, ATLAS, laid the foundation for the adaptive speculator, and this upgrade has resulted in a fully autonomous closed-loop system. The code has been open-sourced on GitHub.

No More Offline Training for Draft Models: Together AI Open Sources Aurora for Speculative Decoding with Self-Learning

According to monitoring by 1M AI News, the Technology Innovation Institute (TII) in the UAE has released two vision models, Falcon Perception and Falcon OCR, on Hugging Face. Both models utilize an 'early fusion' single Transformer backbone: image patches and text tokens share parameter space, with image tokens using bidirectional attention and text tokens using causal attention, eliminating the traditional 'visual encoder + text decoder' cascade design. This allows the models to truly understand spatial constraints and object relationships in natural language, rather than merely performing semantic retrieval of visual features. Falcon Perception has 0.6B parameters and is designed for open vocabulary image segmentation and localization. It achieved a Macro-F1 score of 68.0 on the SA-Co benchmark, surpassing Meta SAM 3's score of 62.3. TII also released the diagnostic benchmark PBench, which evaluates capabilities hierarchically. Falcon Perception shows the most significant lead in tasks requiring language understanding: 1. L2 (OCR-guided recognition, such as finding 'a bottle labeled 168'): 38.0 vs SAM 3's 24.6 (+13.4) 2. L3 (spatial relationships, such as 'the black car on the left' and 'the third window from the left'): 53.5 vs SAM 3's 31.6 (+21.9) 3. L4 (interactive relationships, such as 'the person holding an umbrella' and 'the person using a phone'): 49.1 vs SAM 3's 33.3 (+15.8) 4. Dense scenes (hundreds of instances coexisting): 72.6 vs SAM 3's 58.4 (+14.2). The gap for simple objects (L0) is only +0.8, confirming the trend of increasing disparity with language complexity. In terms of instance existence calibration (whether the target exists), SAM 3 still holds an advantage: MCC 0.82 vs 0.64. Falcon OCR has 0.3B parameters, reusing the same backbone but trained from scratch, specifically designed for document understanding. It scored 80.3 on the olmOCR benchmark (1.7 points from the top), leading all tested models in multi-column layouts (87.1%) and table extraction (90.3%); it scored 88.64 on OmniDocBench, surpassing larger parameter models or those relying on proprietary infrastructure like DeepSeek OCR v2, GPT 5.2, and Mistral OCR 3. According to TII, Falcon OCR is the highest throughput open-source OCR model, achieving a high concurrency test of 5,825 tokens per second on a single A100-80GB (approximately 2.9 images per second for the full process). Both models are now open-sourced on Hugging Face, with Falcon Perception offering an online Playground.

[TII Releases Two Open-Source Falcon Vision Models: 0.6B Segmentation Outperforms SAM 3 with Increasing Language Complexity]According to monitoring by 1M AI News, the Technology Innovation Institute (TII) in the UAE has released two vision models, Falcon Perception and Falcon OCR, on Hugging Face. Both models utilize an 'early fusion' single Transformer backbone: image patches and text tokens share parameter space, with image tokens using bidirectional attention and text tokens using causal attention, eliminating the traditional 'visual encoder + text decoder' cascade design. This allows the models to truly understand spatial constraints and object relationships in natural language, rather than merely performing semantic retrieval of visual features. Falcon Perception has 0.6B parameters and is designed for open vocabulary image segmentation and localization. It achieved a Macro-F1 score of 68.0 on the SA-Co benchmark, surpassing Meta SAM 3's score of 62.3. TII also released the diagnostic benchmark PBench, which evaluates capabilities hierarchically. Falcon Perception shows the most significant lead in tasks requiring language understanding: 1. L2 (OCR-guided recognition, such as finding 'a bottle labeled 168'): 38.0 vs SAM 3's 24.6 (+13.4) 2. L3 (spatial relationships, such as 'the black car on the left' and 'the third window from the left'): 53.5 vs SAM 3's 31.6 (+21.9) 3. L4 (interactive relationships, such as 'the person holding an umbrella' and 'the person using a phone'): 49.1 vs SAM 3's 33.3 (+15.8) 4. Dense scenes (hundreds of instances coexisting): 72.6 vs SAM 3's 58.4 (+14.2). The gap for simple objects (L0) is only +0.8, confirming the trend of increasing disparity with language complexity. In terms of instance existence calibration (whether the target exists), SAM 3 still holds an advantage: MCC 0.82 vs 0.64. Falcon OCR has 0.3B parameters, reusing the same backbone but trained from scratch, specifically designed for document understanding. It scored 80.3 on the olmOCR benchmark (1.7 points from the top), leading all tested models in multi-column layouts (87.1%) and table extraction (90.3%); it scored 88.64 on OmniDocBench, surpassing larger parameter models or those relying on proprietary infrastructure like DeepSeek OCR v2, GPT 5.2, and Mistral OCR 3. According to TII, Falcon OCR is the highest throughput open-source OCR model, achieving a high concurrency test of 5,825 tokens per second on a single A100-80GB (approximately 2.9 images per second for the full process). Both models are now open-sourced on Hugging Face, with Falcon Perception offering an online Playground.

TII Releases Two Open-Source Falcon Vision Models: 0.6B Segmentation Outperforms SAM 3 with Increasing Language Complexity

On April 1, according to CME's 'FedWatch' data, the probability of the Federal Reserve raising rates by 25 basis points in April has dropped to 0.5% (having briefly risen to 12.4% on March 23). The likelihood of a rate hike is now almost negligible, with a 99.5% probability of keeping rates unchanged.

[Probability of Fed Raising Rates by 25 Basis Points in April Drops to Just 0.5%]On April 1, according to CME's 'FedWatch' data, the probability of the Federal Reserve raising rates by 25 basis points in April has dropped to 0.5% (having briefly risen to 12.4% on March 23). The likelihood of a rate hike is now almost negligible, with a 99.5% probability of keeping rates unchanged.

Probability of Fed Raising Rates by 25 Basis Points in April Drops to Just 0.5%

On April 1, according to the Islamic Republic News Agency (IRNA), Iran launched multiple missile and drone attacks on Israel and U.S. military bases.

[Iran Launches Multiple Missile and Drone Attacks on Israel and U.S. Military Bases]On April 1, according to the Islamic Republic News Agency (IRNA), Iran launched multiple missile and drone attacks on Israel and U.S. military bases.

Iran Launches Multiple Missile and Drone Attacks on Israel and U.S. Military Bases

1. Twitter says Musk has become the sole director;
2. Hong Kong issues a policy statement on the development of virtual assets;
3. Bitcoin white paper has been published for 14 years;
4. RBI will initiate release of Digital Rupee on November 1;
5. Bitcoin miner Argo's $27M funding plan failed.

Detail

【Huobi Global Breaking News】1. Twitter says Musk has become the sole director;
2. Hong Kong issues a policy statement on the development of virtual assets;
3. Bitcoin white paper has been published for 14 years;
4. RBI will initiate release of Digital Rupee on November 1;
5. Bitcoin miner Argo's $27M funding plan failed.

Huobi Global Breaking News

Article

BTC broke through 67000  USDT

BTC broke through 68000  USDT

Whale Withdraws 6,905 ETH Worth $14.496 Million

Trump's Withdrawal Signal Boosts Nikkei Index by 4%

Thousands of U.S. Soldiers Deploying to the Middle East

US Spot Bitcoin ETF Sees $117.31 Million Net Inflow Yesterday

Huobi Global Breaking News

Isenção de responsabilidade：O conteúdo acima não representa as posições da HTX.，A HTX não faz quaisquer recomendações de negociação.。

Artigos relacionados

Todos os comentários0Mais recentePopular

Tópicos em destaque

Artigos relacionados