阿里开源Qwen3.8，千问大模型全球下载超30亿次
Related
Alibaba Cloud: Meoo launches Night Plan, offering discounts of up to 20% on Qwen 3.7-Max at night.
According to Mars Finance, Alibaba Cloud announced the launch of the Night Plan for Meoo: From 22:00 to 08:00 (Beijing time) each night, users who initiate dialogues using a specified model on the web interface will be automatically charged at off-peak rates. Qwen 3.7-Max nighttime discounts are as low as 20% off, and Qwen 3.7-Plus nighttime discounts are as low as 40% off. Users do not need to change settings; simply log in during off-peak hours to build applications and automatically enjoy the discount. The discount only affects the price; model service quality and response speed are unaffected. Usage Notes: This is only available to Pro/Max paid plan users; Free plan users and users who have only purchased points are not currently supported. Coverage is limited to web interface dialogue building; Meoo CLI and AI service calls are not included in the discount. Requests outside of off-peak hours and other models will still be charged at the standard rate. (Cailian Press)
Alibaba's Qwen 3.7Max preview version of the 1000 Questions large model is about to be released.
According to BlockBeats, Alibaba's Qwen3.7Max and Qwen3.7Plus preview models will be released on May 19th. In Arena's evaluation, Qwen3.7-Max-Preview ranked 13th in the text domain, and is also the top-ranked domestic model. In the visual domain, Qwen3.7-Plus-Preview ranked 16th, again the top-ranked domestic model on the leaderboard.
Qwen3.7-Max becomes the world's second-largest AI programming model.
Mars Finance reported on May 26th that Code Arena, a globally authoritative third-party programming benchmark, released its results. Alibaba's latest flagship model, Qwen3.7-Max, scored 1541, surpassing models such as GPT-5.5, Gemini-3.5-Flash, GLM-5.1, and Kimi-K2.6, ranking second globally among major model manufacturers, only behind the Claude series. (Wide Angle Observation)
Alibaba Qwen 3.7-Max launches automatic implicit caching, reducing input costs by up to 80%.
According to Beating's monitoring, Alibaba's Qwen team announced that automatic implicit caching has been enabled by default for its flagship model Qwen3.7-Max on the Alibaba Cloud Bailian platform. Developers can directly benefit from caching and cost reduction without modifying code or specifying additional parameters. Under the new billing mechanism, the system automatically identifies and extracts repeated context prefixes in requests. Once a cache hit occurs, the cost of the input tokens for the hit portion is only 20% of the original unit price, directly eliminating 80% of the input cost. Implicit caching directly addresses the huge overhead in long text and Agent scenarios. With a long context window of 1 million tokens, Qwen3.7-Max needs to frequently and repeatedly read huge code bases or knowledge documents when running high-level tasks such as autonomous coding. One developer reported after testing Qwen3.7 that it consumed nearly 1 million tokens in less than an hour to build a tank battle webpage demo. If the agent is allowed to autonomously perform code review and iterative cycles in the background, the daily usage can easily reach hundreds of millions of tokens. The fierce competition among peers in caching pricing was another direct factor prompting Alibaba to lower its prices. Previously, DeepSeek V4-Pro attracted a large number of developers with its extremely low cache hit rate. After announcing a permanent price reduction at the end of May, DeepSeek V4-Pro's cache hit rate was reduced to only $0.003625 per million tokens (approximately RMB 0.025), equivalent to a 99.17% cost reduction based on the standard input price. Many developers, using tools like Reasonix, pushed the cache hit rate for a single session to a maximum of 99%, making the runtime cost of long-session AI agents almost zero. Faced with competitive pressure, Qwen3.7-Max not only launched an implicit caching mode requiring no configuration but also retained an explicit caching mode that required manually declaring the `cache_control` flag. Compared to automatic caching, explicit caching has higher hit certainty and a hit cost as low as 10% of the standard input unit price (a 10% discount). However, a 125% premium must be paid when creating the cache for the first time, and the cache block has a lifespan of only 5 minutes (which can be reset each time a hit occurs).
Qwen 3.7-Max officially launched its API, with Bailian providing Token Plan subscription and invocation capabilities.
According to Beating's monitoring, Alibaba's latest flagship large-scale model, Qwen3.7-Max, has been launched on the Alibaba Cloud Bailian platform, officially opening its API interface and Token Plan subscription service. The input is 12 yuan per million tokens, and the output is 36 yuan per million tokens. The Bailian Token Plan subscription service is also available on Qwen3.7-Max, and subscribers can access it directly without additional payment. According to Artificial Analysis's evaluation, the previously released Qwen3.7-Max scored 56.6 points, ranking fifth globally and first among domestic large-scale models. Qwen3.7-Max is designed specifically for intelligent agents and features architectural optimizations for long-term tasks such as autonomous programming and multi-tool invocation. In actual long-term tests, Qwen3.7-Max, in collaboration with frameworks such as Claude Code, OpenClaw, Hermes Agent, and Qwen Code, can independently complete enterprise-level complex tasks lasting up to 35 hours and involving over 1000 tool calls.
Qwen3.7-Max officially released: 1158 self-written codes in 35 hours, forging a 10x speedup operator on a domestically produced chip.
According to Beating's monitoring, Alibaba's Tongyi Qianwen officially released its new generation flagship intelligent agent foundation, Qwen3.7-Max. Official data shows that, without any chip architecture documentation or performance analysis data, the new model forcibly improved the Triton operator performance of the domestically produced Pingtouge Zhenwu M890 processor by 10.0 times in a fully autonomous kernel optimization task lasting 35 hours and spanning 1158 tool calls. During the optimization process, the model underwent five core evolution stages. It first uses Split-K partitioning to divide the prefix KV-cache along the token dimension to fill 36 SM cores; then it replaces the cudaMalloc synchronization between the host and device with pre-allocated PyTorch variables, and completely eliminates the synchronous cudaMemcpy action when querying the prefix length by using tensor metadata, thus completely removing the communication overhead between the host and device; in the final stage, the model refactoring operator processes all 4 query tokens simultaneously in a single thread block, and shared loading is used to distribute memory access overhead, completing the key architecture-level specialization refactoring. Operator optimization tests show that Qwen3.7-Max achieves a geometric mean speedup of 10.0x, significantly outperforming GLM 5.1 (7.3x) and Kimi K2.6 (5.0x). DeepSeek V4 Pro only achieves 3.3x and prematurely terminates the task in the latter half due to not issuing any tool calls for five consecutive rounds. To master general problem-solving strategies in diverse environments, Qwen3.7-Max decouples the task, runtime framework, and validator during training and avoids shortcut overfitting for specific benchmarks through cross-framework reinforcement learning training. On the general agent benchmarks MCP-Mark (60.8 points) and SpreadSheetBench (87.0 points), Qwen3.7-Max demonstrates extremely strong generalization ability, with its overall performance approaching that of Claude-4.6-Opus-Max.