3D modeling comparison test between DeepSwarm V4.1 Flash and Claude Opus 5.5

WTF!!! DeepSeek is absolutely godlike—the new version beats Fable, Opus, Astra? 🤯.. Next Monday's US stock AI sector is probably going to get crushed… Even more absurd is that DeepSeek's just-open-sourced V4.1-Flash 1017 costs only $0.15 per million input tokens and $0.30 for output, and on the DeepSWE benchmark it achieved a 77% success rate—higher than GPT-6.1 Sol—at one-third the per-run cost 🤯! If you just look at the benchmark scores, you might think it's just another routine update, but when you put the bill and the weight size together, it directly smashes the cost of calling frontier coding models through the floor. Fucking awesome, On the DeepSWE benchmark, which measures real software engineering capability, it achieved a 77% success rate with a per-task cost of only about $0.20, while the comparison GPT-6.1 Sol's highest-configuration success rate is about 75%, yet each task costs $0.65. At one-third the price, it beats the top closed-source foundation models on core development tasks—that kind of contrast was hard to imagine before, In the past, if you wanted an Agent to repeatedly trial-and-error in a codebase and batch-fix bugs, calling top-tier models was like burning a taximeter. Now input costs only $0.15 per million tokens and output only $0.30. Converted to the most common development scenario—100K input plus 10K output in a single run—the bill is about 86% to 93% cheaper than Sol, and about 93% to 97% cheaper than Opus 5.5, And if you pair it with DeepSeek Harness to get context caching running, the actual cost under high-frequency hits will drop by another order of magnitude, basically making it possible to call frontier models as freely as tap water. Besides cutting the bill to the bone, its size also throws a bucket of cold water on Silicon Valley's brute-force approach of piling on compute, Compared with the rumored 10T-parameter American closed-source giants, or the 2.8T Kimi K3, this Flash version keeps total parameters at 763B, with net weights of about 510.3GB, and it directly releases the weights for free. Based on this size, you don't need to rent a large data center specifically—4 DGX Sparks or a top-spec 512GB Mac Studio with one layer of quantization has a chance of running it locally on a desktop. What's most brutal is how it improved its scores this time: it didn't burn tens of thousands of cards to train a new foundation model, V4.1-Flash was originally just an old model that had gained some fame for being cheap and easy to use. This time, the official team only introduced reinforcement learning with self-feedback for post-training on the original foundation, and forcibly turned a budget model into SOTA level. There's an extremely counterintuitive phenomenon hidden here. Previously, to improve accuracy, models often had to crazily lengthen their chain of thought and spit out more nonsense in exchange for scores, But this time, the 1017 version tuned through post-training not only raised the success rate on development tasks, but also pushed down processing time and output token volume together—meaning the model got smarter, said less, and conveniently saved users money. To put it plainly, this is a wake-up call for all of us who use AI to write code and do business every day, When post-training can simultaneously max out accuracy and conciseness, what comes next is no longer about who can afford to call the most expensive closed-source API, but who dares to roll out all-day automated testing and whole-repo refactoring tasks—previously avoided because they were too expensive—into daily workflows at a hundredfold scale. But from a more realistic perspective, the original poster is honest in saying it has not yet comprehensively surpassed Opus 5.5 in all dimensions. A comeback on a single benchmark does not mean it wins across all scenarios, and the claim that the US stock AI sector will crash next Monday is more of an emotional vent, In addition, trying to stuff 510.3GB of raw weights into a machine with 512GB of memory—counting system overhead and context caching—will still blow out memory without quantization compression. The actual losses of local deployment still await the first batch of real-world tests. Feeding models bigger by piling on compute is just the brute-force work of the first half; using post-training to make models say less nonsense and do more of the right work is the real craft of the second half that creates the cost gap.

View original

WTF！！！DeepSeek这是超神了啊，新版本超越了Fable、Opus、Astra？🤯。。下周一的美股 AI 板块估计要被干崩了吧… 更离谱的是DeepSeek 刚开源的 V4.1-Flash 1017，输入每百万 tokens 只要 0.15 美元、输出 0.30 美元，在 DeepSWE 基准上用三分之一的单次成本跑出了比 GPT-6.1 Sol 还高的 77% 成功率🤯！ 如果单看跑分可能觉得只是又一次常规更新，但把账单和权重体积放在一起算，它直接把前沿编程模型的调用成本砸穿了地板，牛逼牛逼， 在衡量真实软件工程能力的 DeepSWE 基准测试上，它跑出了 77% 的成功率，单次任务成本只要约 0.20 美元，而作为对比的 GPT-6.1 Sol 最高分配置成功率约 75%，单次任务却要花掉 0.65 美元。  用三分之一的价格，在核心开发任务上跑赢了顶尖闭源底座，这种反差放在以前很难想象， 以前想让 Agent 在代码库里反复试错、批量改 bug，调用頂級模型就像在烧计价器，现在输入每百万 tokens 只要 0.15 美元，输出只要 0.30 美元。  换算成一个平时最常见的开发场景，单次喂进去 10 万输入加上 1 万输出，账单比 Sol 便宜了约 86% 到 93%，比 Opus 5.5 便宜了约 93% 到 97%， 要是再配合 DeepSeek Harness 把上下文缓存跑起来，高频命中下的实际开销还会再降一个量级，基本实现了像自来水一样随便调前沿模型。 除了账单砍到脚脖子，它的体积也给硅谷那套靠堆算力硬砸的路线泼了一盆冷水， 相比于传闻动辄 10T 的美国闭源巨无霸，或者 2.8T 的 Kimi K3，这版 Flash 总参数控制在 763B，净权重约 510.3GB，而且直接免费公开了权重，按这个体积推算，不需要专门去租大型数据中心，4 台 DGX Spark 或者顶配 512GB 的 Mac Studio 做一层量化，就有机会在本地桌面跑起来。  最狠的是它这次提分的方式，并不是重新烧几万张卡训了一个新底座， V4.1-Flash 原本只是一个因为便宜好用而小有名气的旧模型，官方这次只是在原底座上引入了带自我反馈的强化学习做后训练，硬生生把一个平价模型拧成了 SOTA 级别。  这里藏着一个极其反直觉的现象，以前模型想要提高准确率，往往得靠疯狂拉长思考链、吐出更多废话来换取分数， 而这次通过后训练调出来的 1017 版本，在拉高开发任务成功率的同时，反而把处理时间和输出 token 量一起压了下来，等于模型变聪明了，话变少了，还顺手帮用户省了钱。  说白了，这给我们所有每天用 AI 写代码、做业务的人提了个醒， 当后训练能够同时把准确率和精简度拉满，接下来拼的就不再是谁调得起最贵的闭源接口，而是谁敢把过去因为嫌贵而不敢跑的全天候自动化测试、整库重构任务，现在成百倍地铺进日常工作流里。  但站在更现实的视角来看，原帖作者在所有维度上还没有全面压过 Opus 5.5 这一点上说得很诚实，单点基准测试的逆袭不等于全场景通吃，而且所谓下周一美股 AI 板块要暴跌的说法更多是情绪宣泄， 另外 510.3GB 的原始权重想塞进 512GB 内存的机器里，算上系统开销和上下文缓存，不经过量化压缩依然会爆内存，本地部署的实际损耗还得等第一批实测。  靠堆算力把模型喂大只是上半场的蛮力活，靠后训练让模型少说废话多干对活，才是下半场真正拉开成本代差的手艺。