This research reveals that quality trumps quantity in skill distillation, potentially reducing computational costs by over 90% while maintaining performance. For developers and researchers working with large language models, this method offers a practical way to optimize training efficiency and improve model performance on specific tasks like math contests.
微软基于 Qwen3.5-9B 出了个决策小模型，专管 Agent 流程里的判断打分，比 GPT-6 Sol 快 35 倍，输出还免费。
微软把决策类小模型做成了专项产品，比 GPT-6 Sol 快 35 倍还输出免费，做分类和判断任务的可以试试。
This specialized decision model addresses a critical inefficiency in AI applications - replacing expensive general-purpose models for simple choice-based tasks. For developers in Southeast Asia, it offers significant cost reductions (up to 200x) and speed improvements (up to 100x) for tasks like content moderation, intent recognition, and AI output evaluation. The model's availability through Microsoft Foundry and OpenRouter makes it immediately accessible for production use cases across the region.
微软用 Qwen3.5-9B 做了个专门快速打分决策的模型，比 GPT-6 Sol 快 35 倍，Foundry 上就能用。
想让你的 Agent 真刀真枪比一场？BenchFlow 办了场比赛，用免费 GPU 训 Qwen3.8-27B，看谁的环境设计能把模型调到最强。
The talk signals Alibaba Cloud's push to position its Qwen family as agent-native infrastructure rather than chat models, a framing that affects how enterprises in Singapore and Southeast Asia choose MaaS providers. Teams evaluating Alibaba Cloud against regional cloud rivals can watch the session to compare its agent stack and realtime audio capabilities. Builders planning multi-step agent workflows on Qwen models will find the execution-oriented framing directly relevant.
JetBrains 开源了新的编程模型 Mellum2.1，专门练了智能体写代码的能力，高负载下吞吐量几乎是 Qwen3.5-9B 两倍，还能本地部署保数据安全。
This research demonstrates that harness engineering can significantly outperform traditional approaches, offering developers a practical method to optimize AI-assisted coding workflows. Engineers working with AI coding assistants should consider implementing component-specific LLMs with dependency awareness to achieve substantial performance gains and cost reductions, particularly beneficial for resource-constrained development environments.
论文教你看 agent 失败原因来决定改运行时框架还是微调模型，Qwen3.5 上从 0.16 拉到 0.30，做 agent 的都该看看这套诊断方法。
一篇 MoE 训练论文：把固定 top-k 换成在 k 附近随机采样，Qwen3-30B-A3B 微调能多拿 2 分，推理成本不变。
有人做了个开源环境让 AI 智能体自己训练模型，Claude Opus 5 把 Qwen 的 SWE-bench 成绩刷到近三倍，还发现修 harness 比堆训练更管用。
Plain-chat safety evals can overstate how safe an agentic system actually is, since tool use measurably erodes refusal behaviour across every model tested. Teams building multimodal agents should benchmark safety with tools attached and consider re-inserting the original request before generation. The paper is on arXiv with a chat interface at academy.dair.ai for quicker exploration.
给 TTS 加了个情绪旋钮，不用重新训练模型，注入转向向量就能调情绪强度，Qwen3-TTS 上效果最高提升 7 倍。
这篇论文把 RAG 从'检索'升级成'主动算证据'，RouterLM 用 Qwen3.5-9B 就打赢了 Gemini 3.5 Flash，做 RAG 的可以看看思路。
Serving stacks have so far been split between LLM servers optimized for autoregressive decode and diffusion stacks built for denoising, forcing teams to glue disjoint runtimes together in production. vLLM-Omni replaces that with one control plane supporting session-shaped, long-lived workloads like robot loops and world models. Engineers building multimodal assistants on Qwen3-Omni or MiniCPM-o should evaluate it against their current inference setups.
Deploying text-to-speech models often fails on undocumented version pins and runtime quirks, and this repo documents exactly those gaps for Qwen3-TTS. Developers running inference infrastructure can copy the configs straight into their own vllm-omni deployments instead of debugging from scratch. It pairs Alibaba's Qwen audio model with open-source serving tooling, so teams evaluating self-hosted TTS alternatives to hosted APIs get a practical starting point.
The key idea is that observability and training share one dataset: the traces showing where an agent fails become the fine-tuning data and the eval set. Teams running agents on specialised documents can replace expensive frontier APIs with a small model tuned to their own workload. The comparison against GPT5.6 Luna on clause-citation accuracy gives a concrete baseline for anyone weighing cost against reliability in production.
有人发现给基础模型加一句"Okay"开头，Olmo-3-7B 在 MATH-500 就从 42% 涨到 78%，原理分析也很有意思。
12GB 显卡也能跑 125B 的 Qwen3.8-Flash-Next 了，Strata 开源，RTX 5070 上 94 词元/秒，配置门槛和 GitHub 链接都在文里。
不用 RL 训练，只在开头加个 "Okay" 就能让 Olmo-3-7B 数学成绩翻近一倍，这篇论文把原因挖得很透，做模型训练的值得看看。
给医学问答做了个按问题动态分配 LoRA 秩的微调方法，在 Qwen3-8B 上比 LoRA r16 高了 1 个多百分点，专科越多优势越大，做医疗方向微调的可以看看。
有人测了 TypeSafe 那个号称便宜快的 JEV，结果在 OpenAI 批量价格下根本不省钱，校准也没明显赢 Qwen3.8-27B，想做文本标注的可以看看。
GUI 智能体缺训练数据的老问题，这篇用图像编辑模型自己造数据，Qwen3-VL 微调后 AndroidWorld 成绩还涨了不少，代码开源可以直接试。
孟加拉语这种低资源语言上，Llama 4 Maverick 不微调就拿下 0.931 的 F1，比 BanglaBERT+XGBoost 还高，做低资源语言分类的可以看看。
Atomic Chat 把 NVIDIA TensorRT 接进自家引擎，RTX 5090 上跑 Qwen 系列模型比 llama.cpp 快一倍多，本地部署党可以看看实测数据。
Training on verified execution traces can bake retrieval-dependent prompt skills into model weights, reducing reliance on the agent finding and following instructions at runtime. Teams building coding agents can replicate the recipe by turning their own skill libraries into checkable sandbox tasks. The work comes from a Chinese lab and benchmarks on Claude Code with Qwen, making it a useful reference point against prompt-based skill systems like Claude Skills.
A 27B open-weights model running locally now solves multi-digit addition-in-words at near-perfect accuracy, a task GPT-4o fumbled two years ago. Practitioners evaluating small local models for arithmetic-adjacent workflows get a concrete benchmark on consumer-tier DGX Spark hardware. Comparisons across reasoning-on and reasoning-off settings make it easy to judge whether enabling reasoning is worth the latency.
Training in a scaffold nobody ships has been a quiet flaw in agentic RL; this proxy approach removes the need to reimplement harnesses as environments. Practitioners building agentic models can replicate the full pipeline on Hugging Face's guide and shape rewards for behaviors like tool efficiency. The results also show why harness-specific benchmark scores overstate real-world capability.
Figure 要发自己的 AI 助手 Hark 了，通用任务跑 Opus 5.5，还按 Pro 到 Pro⁴ 分四档收费，直接对标 Muse。
An 89% reduction in time to first audio means real-time voice responses from Alibaba's omni-modal model now fit on consumer hardware instead of data-centre GPUs. Developers building local voice agents can replicate the AWQ-4bit setup on an RTX 5090-class card. If the code lands upstream in vLLM-Omni, the gains become available to the whole vLLM community rather than one team.
一个开源项目把阿里 Qwen 的 125B 模型量化后塞进了 12GB 显存的消费级显卡，每秒还能跑几十个 token，本地部署玩家可以看看它怎么分配显存和内存。
不生成文本、只做分类决策的 Jev 三周就跑到日处理 1 万亿 token，OpenAI 和 Cloudflare 都出了同路线产品，挺值得了解的新方向。
