Thinking Machines 新模型领跑西方开放权重 AI

Yana Iskayeva/Moment via Getty Images
Thinking Machines 发布了一款颇具竞争力的新开放模型 Inkling 2 Small。
最具竞争力的开放模型大多出自中国。
此次发布之前,法国实验室 Mistral 刚推出了擅长编程的开放模型 Le Chonk。
由前 OpenAI CTO Mira Murati 创办的小型 AI 初创公司 Thinking Machines Lab 发布了其最新模型 Inkling 2 Small——据第三方测试机构 Artificial Analysis 评价,该模型有望撼动中国模型的领先地位。严格来说,它目前仍属专有模型,但 Thinking Machines 表示将于下周开放其权重。
过去几年,中国企业在开放模型领域一直领跑。DeepSeek 和 Moonshot 的模型经常位居各类榜单前列,能力足以与 OpenAI 和 Anthropic 那些价格高得多的闭源模型相抗衡。
Inkling 2 Small 能做什么
Thinking Machines 的这款最新模型拥有 5940 亿参数,支持图像输入,但仅输出文本。该公司特别提到了其在网页与游戏设计方面的能力,并表示在测试中不仅评估了模型生成的代码,也评估了它解决编程问题的思路。
该公司表示:“最终得到的模型不仅在编程方面更强,也更容易在真实的智能体循环中使用。”
Inkling 2 Small 在 Artificial Analysis Intelligence Index 上得分 31,这一综合测试“用于追踪迈向通用人工智能的进展”,汇总了涵盖编程、科学和通用类别的基准测试得分。31 分高于同级别模型的平均水平,与 Gemini 3.8 Flash 持平。作为参照,Gemini 4 Carbon 和 Claude Opus 5.5 分别以 60 分和 59 分位居该指数榜首。
Thinking Machines 表示将于下周发布该模型的权重,这种做法正逐渐成为惯例。鉴于围绕 AI 的整体安全担忧加剧,AI 公司会先花几周时间,与网络安全团队和安全审查人员一起测试即将开放的模型,然后再发布权重,而不是一上来就放出权重。
法国 AI 实验室 Mistral 最近也以同样的分阶段流程发布了其 Le Chonk 模型——它目前在编程能力上已跟上了 OpenAI 的 GPT-6.1 Terra,这对西方开发的开放模型来说十分罕见。
Artificial Analysis 表示:“按当前排名,[Thinking Machines] 发布权重的计划将使其成为中国以外开发的得分最高的开放权重模型。”
但很少有模型能做到面面俱到。
“Inkling 2 Small 在智能水平上处于领先模型之列,与价位相近的其他模型相比定价也算合理。但它的单任务速度明显偏慢,而且输出非常冗长,”Artificial Analysis 表示。该机构还称,Inkling 2 Small 每百万输入 Token 0.30 美元的定价“略显昂贵”,因为该级别的中位成本为 0.24 美元。
不过,在 Artificial Analysis 的评测者看来,该模型相较于前代的进步依然突出:Inkling 2 Small 的自动化能力跃升 41%,编程能力提升 20%。但大多数模型都是在某些领域或特定基准测试上表现优异,因此很难断定哪一款是整体上绝对最强的编程模型。
Artificial Analysis 表示,Inkling 2 Small 在准确率方面也有所提升。
其评测报告称:“在 AA-Omniscience v1.1 上,幻觉率从 Inkling Small 的 61% 降至 27%,准确率从 32% 升至 34%,尝试回答率从 77% 降至 53%。”
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Thinking Machines’ new model takes the lead in western open-weight AI)
Yana Iskayeva/ Moment via Getty Images
Thinking Machines released a competitive new open model, Inkling 2 Small.
Most competitive open models come out of China.
The release follows French lab Mistral’s Le Chonk open model, which excels at coding.
Thinking Machines Lab, a small AI startup from ex-OpenAI CTO Mira Murati, has released its latest model, Inkling 2 Small — and according to third-party tester Artificial Analysis, it’s set to give Chinese models a run for their lead. It’s technically proprietary until now, but Thinking Machines said it will release its weights next week.
Chinese companies have led the pack on open models over the last few years. Models from DeepSeek and Moonshot regularly top leaderboards and rival the capabilities of much more expensive, closed models from OpenAI and Anthropic.
What Inkling 2 Small does
Thinking Machines’ latest model has 594 billion parameters and supports image inputs but outputs only text. The company noted its web and game design capabilities and that it evaluated the model’s approach to coding problems alongside the code it generated during testing.
“The result is a model that is both stronger at coding and easier to use inside real agent loops,” the company said.
Inkling 2 Small scored a 31 on the Artificial Analysis Intelligence Index, a composite test “for tracking progress toward artificial general intelligence” that combines benchmark scores spanning coding, science, and general categories. A score of 31 is above average for models in its class and is tied with Gemini 3.8 Flash. For context, Gemini 4 Carbon and Claude Opus 5.5 top the index at 60 and 59, respectively.
Thinking Machines said it will release the model’s weights next week, a practice that is becoming customary. Given the heightened security concerns around AI in general, AI companies spend a few weeks testing the to-be-open model with cybersecurity teams and safety reviewers before releasing the weights, rather than leading with the weights out of the gate.
French AI lab Mistral recently released its Le Chonk model via the same staggered process — it now keeps pace with OpenAI’s GPT-6.1 Terra in coding abilities, which is rare for western-made open models.
“[Thinking Machines’] plan to release its weights would make it the highest-scoring open weights model developed outside China based on current rankings,” Artificial Analysis said.
But rarely is a model the full package.
“Inkling 2 Small is amongst the leading models in intelligence and reasonably priced when comparing to other models of similar price. It’s also notably slow per task and very verbose,” Artificial Analysis said. The organization also called Inkling 2 Small’s pricing “somewhat expensive” at $0.30 per one million input tokens, since the median cost for its class is $0.24.
Still, the model’s progress over its predecessor stands out to Artificial Analysis reviewers: Inkling 2 Small jumped 41% in automation capabilities and 20% in coding abilities. However, most models excel in certain areas or on specific benchmarks, making it hard to deem one or another the outright best coding model overall.
Inkling 2 Small also made some gains in accuracy, Artificial Analysis said.
“On AA-Omniscience v1.1, the hallucination rate falls from Inkling Small’s 61% to 27%, while accuracy rises from 32% to 34% and the attempt rate falls from 77% to 53%,” its review reported.
Please accept cookies to view the comments powered by Disqus.
Radhika Rajkumar is a senior editor at ZDNET based in New York City. She covers AI, specializing in safety, privacy and security, policy, education, and synthetic media. She also leads ZDNET's newsletter strategy. Radhika holds a Masters in Creative Publishing and Critical Journalism from The New School. See full bio
原文为英文,正文由自动抓取并保留原文。
原文发布于 ZDNet(Thinking Machines’ new model takes the lead in western open-weight AI),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。