The harness you wrap around an open model can swing both solve rates and cost by large margins — opencode ran to US$3.35 per solved task, 11.7 times Pi's cost, on identical models. Teams self-hosting GLM, DeepSeek or Kimi models should pick a harness per model rather than a single default. The companion Harness Lens documents prompts, tools and compaction across 34 harnesses with source links.
Provider choice silently changes model quality: quantisation, custom kernels and output caps mean the same model can differ by 20 points or more depending on the endpoint. Developers picking serverless providers should check endpoint accuracy alongside price and latency rather than assuming all hosts are equivalent. Given Zhipu's GLM models are widely served across Southeast Asian inference providers, the GLM-5.3 Flash findings are directly relevant to regional API buyers.
Revolut 嫌 Claude Code 一人一月烧 1 万美元，直接买了 56 张 B200 自己跑 GLM-5.3，成本降到十分之一，还讲了量化调优的具体参数。
Long-sequence post-training on biological data is a rare public data point, showing that million-token fine-tuning is practical on 64 GPUs rather than a cluster of thousands. The open-source stack means teams doing AI for science or other long-document domains can reproduce the setup directly. GLM-5.3-Flash originates from Zhipu (Z.ai), so it also demonstrates a Chinese open-weight model being repurposed by US labs.
Artificial Analysis 数了各模型写破折号的频率，GLM-5.3 一条回复平均 7.4 个，Claude Opus 5.5 几乎为零，看你的模型像谁就靠这个。
如果你在华为云上跑 openPangu-2.0，11 月底就得换了，官方指的替代是 DeepSeek-V4.1-Flash 和 GLM-5.3，记得提前迁移。
一个把'指挥'和'干活'拆开跑的智能体框架，人能直接改文件插手，断线还能接着干，成绩追平 Codex 和 DeepSeek Harness。
华为云的盘古两款开源模型要下线了，官方还给了替代方案，用着的朋友记得提前迁移评测。
A specialised open model beating a frontier closed model on a science benchmark strengthens the case for domain-specific post-training over general-purpose scale. Research labs and computational chemistry teams can download and adapt the weights for their own discovery workflows. Its base model, Zhipu's GLM-5.3, also shows Chinese open models being repurposed at the frontier of applied science.
The method cuts indexer cost, which accounts for over half of prefill time at one million tokens, without the recall degradation that two-stage block approaches like HiSA suffer as context grows. Serving teams running long-context inference on models with learned sparse attention, such as GLM-5 or DeepSeek-V4, can use it to raise prefill throughput 1.3-1.8 times while preserving retrieval accuracy. Both target models come from Chinese labs, making it directly relevant to practitioners deploying these open-weight models.
Deep Principle 把 Z.ai 的 GLM-5.3 改造成了 753B 的科研模型，专攻化学和材料，还开源了 MIT 权重，搞科研计算的可以直接拿来跑。
超长上下文跑不动？这篇论文用分层聚类做稀疏注意力，10M token 下比全注意力快 72 倍还不掉点，做长上下文的值得看看。
同价差 499 元，两台 256GB 用雷雳 5 连起来跑 GLM-5.3 快了近一倍，买 Mac Studio 跑大模型的看过来。
Mistral 出了个 1 万亿参数的开源模型，编程跑分超过了 DeepSeek V4.1Flash，价格还挺便宜，写代码的可以看看。
字节把 Kimi、GLM、DeepSeek 直接塞进自家办公套件，一个 credits 体系随便切，做企业选型的话很省事。
Mistral 把 1T 参数的 Large 4 权重放出来了，编程智能体成绩追平 GLM-5.3 还省近一半 Token，做微调或自部署的可以看看。
Mistral 放出了全欧训练的万亿参数模型 Le Chonk，编码能力强、速度快还便宜，不过分数上还是被 DeepSeek V4.1 Flash 和 GLM-5.3 压一头。
Z.AI 的 GLM-5.3 居然能自己跑完整套网络攻击流程，研究者已经出来敲警钟了，搞安全的朋友得看看。
The v5 rebuild shifts the open-vs-closed frontier gap from 3 points under v4.3 to 12 points, meaning prior open-weight scores may have been inflated by public-question contamination. Practitioners choosing models for real coding work should look at FrontierCode 2.0's merge-based scoring, which measures restraint, style fit and regression risk rather than passable code. Kimi K3.1 Pro nominally leads open-weight models at 46 overall, but its weights only ship by 4 Nov 2026.
模型评测榜 v5 版来了，60% 改用私有题防刷榜，Claude Opus 5.5 拿了第一，开源和闭源差距拉大到 12 分，选模型前可以看看。
Agent 查漏洞太慢，这篇论文直接读 LLM 内部表示，1 秒判定一个函数，比 agentic 方案快 250 倍还不掉准确率，做安全工具的可以看看。
用 GLM 5.3 或自建 SGLang 推理服务的值得看看，Cloudflare 公开了 KV cache 跨请求泄露的排查和修复细节。
Cost and privacy trade-offs in agent execution do not move together — once task completion converges, strategy choice is mostly a privacy decision, and the best-scoring setup is also the most exposed. Teams deploying on-device agents can pick a collaboration mode based on edge model size: route execution to the cloud for small models, use cloud advisory for 9B-class ones. The comparison against Tsinghua-affiliated GLM-5.3 makes the findings directly applicable to Chinese open-source model stacks.
腾讯做了个“懂了鸭”，老师家长打句话就能生成能拖能玩的 3D 课件，背后接了 DeepSeek 和 Kimi 一堆模型，跟豆包爱学走的是完全不同的路子。
一篇研究智能体自我改进的论文，Qwen3.5-9B 用它这套框架跑赢了 DeepSeek-V4-Pro 和 GLM-5.3，做 agent 的可以看看思路。
The same serving stack now optimises both NVIDIA and AMD paths, giving operators a concrete reason to benchmark MI355X deployments against GB300 rather than defaulting to CUDA. Teams running agentic inference traces can enable the mode through SGLang and expect P90 latency gains without changing models. The plugin system also means third-party kernels — starting with GLM-5.3, K3, and MiniMax M3 — will arrive without waiting for a core release.
The non-obvious insight is that correctness alone no longer separates models; style metrics like Slop and Conviction now offer a measurable way to pick a model for tone-sensitive use-cases such as legal drafting or creative writing. Teams building user-facing products can use the eight indexes to shortlist models matching their brand voice. The data covers Chinese models like GLM-5.3, DeepSeek V4.1 Flash and Qwen3.8 27B alongside Western ones, making direct stylistic comparison across the two ecosystems possible.
DeepSeek 更新了 V4.1 Flash，智能指数涨到 45 分，追平 GLM-5.3 成开源第二，幻觉还少了好多，价格没变。
Efficiency is the headline: the model used ~189M output tokens versus 324M for GLM-5.3-Flash, so low cost comes from token frugality as much as pricing. Its 93% hallucination rate on AA-Omniscience means teams should add verification layers for factual workloads. An open-weights release on 23 October 2026 would make it directly comparable with DeepSeek's open models for ASEAN startups building on local hosting.
有人发现 DeepSeek 这个代号奇怪的模型在 Terminal-Bench 4.0 上拿了 42.3，还比 GLM-5.3 强，等 Artificial Analysis 复测。
The interesting shift is that code execution as the agent substrate lets weaker models close the gap with frontier ones by writing programs instead of clicking precisely, which changes how agent-vs-model capability should be evaluated. Developers building computer-use agents can reuse the harness, SDK and macOS app directly and compare against GUI-only and hybrid designs that the paper treats as special cases. The OSWorld-v2 results give a concrete head-to-head of GPT-6 Astra, Claude Opus 5 and Chinese models like GLM and Kimi under the same harness.
Epoch AI 的网络安全基准显示差距从 6 分扩到 11 分，GLM-5.3 得 160，美国闭源前沿 171，数据挺扎眼。
DeepSeek 新出的 V4.1-Flash-1017 是开源的，DeepSWE 1.1 跑分 77%，把 GPT-5.6 Sol 和 Opus 5 都比下去了，价格还便宜。
Meta 放了个 816B 的开源权重模型 Muse Spark Open，美国开源第一，1M 上下文还能塞进两台 Mac Studio 自托管，不过跑智能体任务不如 GLM-5.3。
The 6–7 month gap is a quantified benchmark of how quickly open-weight models absorb frontier-level cyber offensive capability, a metric security teams and policymakers can use to set review timelines. Chinese open-weight models closing in on US closed models is the notable trend here, relevant to anyone comparing model sourcing for security-sensitive workloads across ASEAN and beyond.
Anthropic 把反蒸馏说得很细：三层防御怎么设计、为什么思维链被偷走最伤，还直接点名智谱 GLM-5.3 和 Kimi，信息量不小。
A 35B model matching frontier-scale systems on MLE-Bench Lite suggests agentic MLE performance depends more on training method than raw parameter count, lowering the hardware bar for teams building autonomous experimentation agents. The WIRL reward design over multi-turn execution windows is a reusable recipe for anyone training long-horizon agents beyond ML engineering. The paper also ships an executable task factory, so practitioners can adapt the harness to their own domains.
Anthropic拟10月在新加坡设亚太第五个办事处 开放本地招聘 数据显示，按人口比例计算，新加坡的Claude.ai使用水平在121个国家中排名第二。 我们观察到，人工智能公司Anthropic计划于10月在新加坡设立其亚太地区的第五…
8月13日凌晨，DeepSeek毫无预兆地丢出大招——DeepSeek-V4-Pro正式版（0813） 悄然上线，API、网页端、APP全量同步！ 📊 硬核数据：从“吊车尾”到“屠榜王” V4 Pro正式版在Agent能力上的提升，已经不…
中国出口"新新三样"：从打火机到爬墙机器人，世界不再只是买便宜货 我们观察到一组清晰的数据信号：中国出口的"新新三样"已经从过去的服装、家具、家电，换成机器人、AI模型和创新药物。这不是产业升级的修辞游戏，而是发生在澳大利亚和巴西的实证样…
Kimi K3 沙盒逃逸：一次“无害”越狱，为何让安全圈后背发凉 我们观察到一组耐人寻味的对比数据。Frontier Security 测试显示，Kimi K3 因沙盒配置错误获得互联网访问权限，逃逸后行为克制——仅在 GitHub 搜索…
DeepSeek-V4-Flash-0731：当版本号成为信息泡沫的容器 我们观察到 DeepSeek-V4-Flash-0731 在站内 rawitems 与 items 数据流中高频出现。需要先声明一个基本口径：截至本文撰写时，Dee…
