与「GLM53」相关的搜索结果

From Mecha Hitler to SOTA rare-disease diagnosis in children? @SpaceXAI's @grok 4.6 has taken the 👑 on RareBench, edging out @AnthropicAI Claude Opus 5 for about 1/3 the cost. This was not on my 2026 bingo card! @deepseek_ai's new v4-pro-0813 model underperformed my expectations, v4-flash, and seemingly the entire internet's. We accessed using DeepSeek's 1P API on the day of release and I almost wonder if they didn't switch over their model endpoint correctly. We will re-benchmark and report back. @Zai_org has attracted a following with GLM5.2, but they, too, underperformed. This doesn't surprise me because when I compared GLM and @Kimi_Moonshot K3 for coding use-cases, I found Kimi substantially stronger, but the internet seems to love this model.

貌似没人关心 Qwen3.8-2.4T-A95B 开源，其实 Qwen3.8-Max还行，就是太贵了。 DeepSeek-V4-Pro 感觉有点问题，不如用 Flash。让后训练再跑一会～ 目前国产模型我的选择： 超大杯：Kimi K3 大杯：GLM5.2 中杯：DeepSeek-V4-Flash 国外模型选择： 超大杯：Opus 5 大杯：GPT-5.6-Sol 中杯：Grok 4.5（4.6还在测试）

感觉 Codex 渐渐有垄断趋势了，至少在 X 和海外。 虽然模型与自己Harness 框架下执行会更好。 比如kimi cli搭配K3、zcode搭配 glm5.2、grok build 搭配 grok4.5 等，Deepseek 马上也要出 agent 框架。 但 Codex 更全面，有Browser Use、Computer Use，手机编程，加上插件还能随时切三方模型。 再不济，也能通过 skill 或 cli 调。 按这个趋势发展，其他框架生存空间会越来越小。 假如你负责推广 Deepseek 的 Agent ，你会打什么点？

先是 GLM5.2，然后是 Kimi K3，现在是 Qwen 3.8 Max，每次国产模型的新发布都更加接近 Coding 领域的 SOTA。从 6 月开始，我感觉国内大模型明显开始在 Coding 和 Agent 领域加速了，和海外顶级 SOTA 的差距正在肉眼可见地缩小。 目前看，国产模型的长程任务、自主运行、反馈回路、跨 Harness 泛化和视觉自我检查等，能力越来越强了。做一个谨慎的预测，预计到 2026 年的年底，对于重度开发者和 Vibe 用户来说，海外模型可能会成为辅助模型，国内的大模型将成为我们的主力工具。 模型用户没有忠诚度，大家会用脚投票的，拭目以待。

最近一直在硅谷交流，我发现中国Token出海，是真的在改变格局。 第一，硅谷创业公司已经大规模使用中国开源模型，因为便宜且好用。GLM5.2、Kimi编程模型都引发了不小震动。 第二，大厂也在动摇。Anthropic 6月收入增长放缓，Fable 5从限量变永久开放，看来也是有压力。 我觉得这对无论是中国、美国的创业者都是好事，因为只有竞争才能进步。

Before Argentina called Lionel Messi to play for their national team, Spain had the opportunity and tried to recruit him to play for them instead. Now, 20+ years later, the two teams face off for the World Cup final 🤯

2/ 中国顶尖代码模型（GLM5.2、Qwen3.7 Max）混合价约1美元/百万token，美股SOTA同档位要4-8美元，而且是在成本线以下卖。 高盛测算：价值型Agent模型目前EBIT利润率-30%，代码模型-39%，靠账上现金扛亏损，预计到2030年才转正。

Goldman Sachs: LLM primer A week or so ago there was a lot of questions about the model layer economics when LLMs are without a doubt viewed more and more as a commodity+ reaching a level where being on the frontier of intelligence is no longer the swaying factor. Economics 101, in a market with many substitutes like restaurants, price undercutting becomes a crucial factor and just like EV's is where China wins. Now these talks have been pushed to the background as the Lag 7 has revived on the back of Zucks considerations of an AI cloud business+producing an AI chip (positive read through to SUMCO/ I sold to early 😞). Back to the initial topic, Goldman provides a few insights into the landscape: → China's top coding models (GLM5.2, Qwen3.7 Max) sit at ~$1 per 1M blended tokens while US SOTA runs $4-8 for the same rung of output → And they are selling it below cost. GS pegs the value for money agentic model at a -30% EBIT margin today and the coding model at -39%, cash rich balance sheets eating the loss until it flips to +14% and +22% by 2030 on their numbers → The reason they can serve that cheap is architecture, sub-8% of params activated per token across the board, DeepSeek V4 Pro firing 49B of 1.6T and GLM5.2 40B of 744B, fewer FLOPs and a structural floor under the price → The adoption already shows up on OpenRouter where China models are 5-16% of spend by task but 85% of agent tokens and 89% of code tokens, winning wherever duration and volume make cost per task the number that matters → And the blended token price rolled over with it, SDLLMTK peaked around 2.07 in early June and sits at 1.67 now This seems somewhat similar to the EV playbook to me, we the consumers should win/benefit from a price war but the return to equity shareholders is more ify.

腾讯开始发力了，新发布的 腾讯混元Hy3 很有东西，而且两周内 API 免费用！！！ 价格也太很香：输入 ¥1/百万 token，输出 ¥4/百万 token，命中缓存 2毛5，低于 DeepSeek V4-Pro 295B 的 MoE 只激活 21B，等于用小模型的成本跑出了大模型的水平：推理能力够到 DeepSeek V4 pro、Qwen 3.7 max 这一档，Agent 能力卡在 GLM-5.1 和 5.2 之间。 更实际的是 token 效率，WorkBuddy里的办公任务对比 GLM5.2，做文档省 47%，做 PPT 省 49%，模型便宜一半等于预算翻倍。而且 Apache 2.0 随便商用。 目前 API 免费两周，大家快去薅起来～

Claude 封号封成这狗样 又是检测中转站，又是钓鱼邮件，又是中转站黑名单的…. 还在费尽心机坚持用官方号的朋友们 可以说是真爱了… 花钱用 token 还要偷鸡摸狗，这过的是啥日子啊 不过现在编程方面 codex 和 glm5.2 可以平替 claude 的模型了 写作和思考方面却没有一个能平替，deepseek 和 gemini 勉强能用，确实是个头大的问题