Chinese contentBolt Forge 一周数据：GLM 5.3 Flash 与 DeepSeek V4.1 Flash 占八成请求
Bolt 公了一周真实用量：GLM 5.3 Flash 占 63%，DeepSeek Flash 17%，用户宁可要快的也不挑最大的，这个数据挺说明问题。
Latest
Every published item, in order · 1031 items
Sep 24
Chinese contentRecraft V4.1 Flash 上线 fal 平台，出图约 1.3 秒
Recraft 家最快的图像模型 V4.1 Flash 上 fal 了，出一张图只要 1.3 秒左右，调提示词试构图可以飞快迭代，做设计的可以去试试。
Chinese contentRecraft V4.1 Flash 上线 OpenRouter，单图 1.3 秒、0.007 美元
Recraft 的便宜快速版图像模型上线 OpenRouter 了，一张 7 厘钱、1.3 秒出图，logo 和人像都能做，跑量挺合适。
Sep 23
Chinese contentRecraft 发布 V4.1 Flash，1.3 秒生成图像
Recraft 出了个 V4.1 Flash，1.3 秒就出图，号称现在最快的图像模型，赶时间的可以试试这个再点 Refine 修细节。
Step 5 Preview tested on two real coding tasks
Benchmark scores can hide whether a model survives long, multi-step builds, so the value here is watching Step 5 Preview handle mechanical synchronisation and real tooling rather than canned puzzles. Developers choosing between Step 5, GLM-5.3, Kimi K3 and Claude Opus 5 get a practical cost-versus-reliability comparison for agentic coding workflows. The stated price gap makes it a concrete candidate for teams in Southeast Asia and beyond cutting per-task inference spend.
Goldman Sachs: US Outspends China on AI Compute $806bn vs $110bn
The compute gap is smaller than it looks, because efficiency gains let Chinese labs deliver competitive inference at a fraction of the infrastructure cost. Developers and procurement teams choosing between US and Chinese models should track OpenRouter usage rankings and token pricing rather than headline capex figures. Goldman expects the next competitive front to be virtual-worker use cases in finance, legal and accounting workloads.
Chinese contentGPT-6-Luna 与 DeepSeek-V4.1-Flash 价格对比：各有优势区间
有人在比 GPT-6-Luna 和 DeepSeek-V4.1-Flash 的实际成本，短上下文选 Luna，272K 以上大输入选 DeepSeek，做 agent 的可以看看怎么省钱。
Post-Training Delivery Benchmark Tests LLM Agents as Forward-Deployed Engineers
The benchmark reframes agent evaluation around deliverable outcomes and payment-gated acceptance, a shift from metric-chasing benchmarks that miss corrupted training runs. Teams building or buying post-training-as-a-service can use the acceptance gate and corruption detector to screen automated fine-tuning pipelines. The human FDE baseline gives a concrete performance reference against all four frontier agents.
Jev tested across three Zilliz open-source RAG projects
The benchmarks show where cheap small models can replace larger ones in RAG pipelines: stopping decisions and relationship filtering see large latency and cost savings with matched recall. Engineers building retrieval systems can use the published code to pick where to route between Jev, GPT-5-mini and Voyage rerank-3. The evaluation includes Chinese-language queries, useful for teams serving bilingual users.
Chinese content开源 MiMo v2.6 成为性价比新标杆，代理成本仅为 Grok 4.7 的 1/57
小米的 MiMo v2.6 Flash 跑分接近 Grok 4.7，成本却只有 1/57，做 agent 的可以看看这份 60 任务实测。
Sep 22
Chinese contentCritical-State RL：定位多轮工具调用中值得训练的状态
多轮工具调用失败，不知道该练哪一步？这篇论文给出诊断方法，BFCL v4 缺失函数任务提升 14 个百分点。
Open model releases since July
Chinese models like Qwen3.8-Max and DeepSeek V4-Pro now rival top closed-source options, affecting AI development costs globally.
Chinese contentGrok 4.7 浏览器长任务基准 39.9 分，仍落后 DeepSeek
Grok 4.7 浏览器任务 39.9 分，比上代涨 8.7，但只有 GPT-6 Astra 一半。
Chinese content小米开源全模态模型 MiMo-V2.6：AA 指数 46 分超 Kimi K3
小米把 MiMo-V2.6 权重全开源了，Pro 在 AA 指数拿 46 分压过 Kimi K3，价格只有海外模型的 1/20。
IT之家Source
Question's Gambit improves research agents
The technique halves calibration error for GPT-5.5 and reduces retrieval errors to just 3 out of 79 cases, making it valuable for developers building research agents.
Chinese contentarXiv 论文提出 opinion leader dynamics，解释稀疏注意力如何塑造 token 聚类
把 token 当粒子建模，还拿 Kimi-K3 和 DeepSeek-V4-Flash 实测验证，讲清稀疏注意力为啥省算力还不丢效果。
Jev organizes 2.3K AI research papers
Jev offers a cost-effective alternative to expensive LLM classifiers for organizing research papers, costing only $0.14 for 2,300 papers.
Sep 21
GameASG-Bench: Benchmarking Autonomous Software Generation for Game Development
This paper provides a detailed benchmark for evaluating the performance of AI models in generating autonomous software for game development, which is relevant for anyone interested in the capabilities and limitations of these models.
Chinese contentOpenAI 用 1 万个代理展示纳维-斯托克斯方程失效，DeepSeek V4.1 Flash 发布新架构
OpenAI 用 1 万个代理做了个实验，结果挺有意思的，DeepSeek 也出了个新版本，对长文本处理更便宜了。
Sep 20
DeepSeek V4.1-Flash model ships 1M context with extreme KV compression
This matters for developers and researchers working with large-scale text data, as it significantly reduces memory requirements for long-form processing.
Chinese content智谱GLM 5.3 FlashX性能提升但价格更高，被DeepSeek v4.1 Flash超越
智谱AI新出的GLM 5.3 FlashX模型，性能确实有提升，但价格比之前贵不少，而且被DeepSeek的v4.1 Flash模型给超越了。
Sep 19
Chinese contentGitHub 上全是 Jev 的衍生项目，疯都疯了。
GitHub 上全是 Jev 的衍生项目，疯都疯了。后面看哪家大模型，率先系统级组合 Jev 类似的功能，我希望这个模型是 DeepSeek-V4.2-Flash。Jev 当门卫，DeepSeek 当大脑或者 DeepSeek 出候选，Jev 做选择。DeepSeek 负责慢思考和文本生成，Jev 负责快决策和门控。
Chinese contentBolt 上开源模型首周，开发者选最快 DeepSeek V4.1 Flash
Bolt 上开源模型首周，开发者选最快 DeepSeek V4.1 Flash，该版本在 Bolt Forge 中上线，使用量是 DeepSeek V4 Pro 的 10 倍。
Sep 18
Chinese contentDeepSeek 发布 V4.1-Flash 技术报告
DeepSeek 新发布的 V4.1-Flash 技术报告，优化了预填充计算、HBM 缓存、SSD/主机持久化缓存和缓存失效恢复，还用了近似的 SWA Bounded Replay 替代长期保存局部缓存状态。
Chinese contentRecraft V4 Styles 发布，单图生成多样场景
朋友，Recraft 新出的 V4 Styles 很酷，能从一张图生成好几个不同方向的内容，比如同一个角色可以变成不同风格，比之前版本更灵活。
Nemotron-SEA-LION-v4.8 AI models announced by AI Singapore and NVIDIA
This is the first open-weights AI model specifically designed for Southeast Asia's diverse languages, offering a new option for regional developers.
BioPhys-Bridge benchmark for interdisciplinary scientific reasoning
This provides a new benchmark for evaluating how well large language models can handle complex, multi-step scientific reasoning in interdisciplinary fields like biophysics, which is useful for researchers comparing model performance on specialized tasks.
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
For developers working with large language models, DeepSeek-V4.1-Flash offers a more cost-effective solution for long-context tasks by significantly reducing KV cache requirements.
Chinese contentBolt 推出支持 GLM 5.3 Flash 等开源模型，使用量提升 50 倍
Bolt 给 1100 万开发者开放了 GLM 5.3 Flash 等开源模型，使用量提升 50 倍，比 DeepSeek V4 Pro 更快更便宜。
Sep 17
Chinese contentGPT Image 2.5 Flare 与 Recraft V4.1 Pro 新一轮对比测试
想看看 GPT Image 2.5 Flare 和 Recraft V4.1 Pro 谁在生成动漫风格图像时更擅长，可以看看评论区里的测试细节。
Chinese contentNotion 推出 DeepSeek V4.1 Flash 模型
Notion 推出了 DeepSeek V4.1 Flash 模型，你可以用它来创建更快的智能体，比之前版本更快。
Sep 16
Chinese contentRecraft V4 Styles 发布，单图生成多样内容
朋友，Recraft 新出的 V4 Styles 很酷，你上传一张图就能生成各种角色和场景，比之前版本灵活多了。
DeepSeek V4.1 Flash AI optimizes Resident Evil 7 mobile frame rate
This shows how AI can optimize game performance on older hardware, potentially making more games playable on mid-range devices.
IT之家Source
QART: A Quantum-Classical Hybrid Architecture for Long-Horizon Reasoning
This research explores a conditional path toward quantum scaling for long-horizon reasoning, a key challenge in advanced AI systems.
DeepSeek engineer says AI will surpass human-written code soon
For those interested in the future of AI development, Liu Shengyu's perspective offers a rare and candid look at the challenges and motivations of a leading engineer working on a major model.
IT之家Source