AI Agent News
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Latest industry news
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Key events timeline
OpenClaw erupts on GitHub
OpenClaw hits global GitHub Top 10 in 10 days, outpacing the Linux kernel star growth
Meta acquires Manus for $2B
Meta acquires Manus AI for $2B, locking in the general-purpose Agent race
DeepSeek-V3 open-sourced
The value king, at just 5% of GPT-4 cost
Manus goes viral overnight
The world's first general-purpose AI Agent draws unprecedented attention
OpenAI Deep Research
OpenAI ships a deep-research Agent that generates professional reports in one click
MCP Servers pass 500
The MCP ecosystem erupts — 500+ servers built in 3 months
DeepSeek-R1 stuns the world
Open-source reasoning model at just 3% of OpenAI cost, reshaping the global AI landscape
MCP protocol born
Anthropic releases the Model Context Protocol, the de facto standard for Agent interfaces
Claude Computer Use
Anthropic lets AI directly control the computer screen for the first time, opening a new paradigm
Replit Agent full-stack automation
Natural language to a shipped product, aimed at non-engineers
Cursor ARR passes $100M
The fastest-growing SaaS ever, the new king of AI coding tools
Claude 3.5 tops SWE-bench
The strongest coding AI, bug-fixing at a junior engineer level
Devin launches
The world's first autonomous AI software engineer, able to complete full coding tasks on its own
Baidu Open-Sources Unlimited-OCR: Constant KV Cache Enables End-to-End Long Document Parsing, Sets New SOTA on OmniDocBench
Baidu recently open-sourced the Unlimited-OCR model, with a total of 3B parameters (500M activated), achieving a comprehensive score of 93.92% on OmniDocBench v1.6, setting a new end-to-end OCR SOTA. The core innovation is Reference Sliding Window Attention (R-SWA), which compresses the decoder KV cache from linear growth to a constant, enabling single forward inference to transcribe dozens of pages without increasing latency or memory usage as output length grows. ## Technical Core: R-SWA and DeepEncoder - **R-SWA Mechanism**: Each generated token simultaneously attends to all reference tokens (visual tokens and prompt tokens) and the most recent 128 output tokens. Visual tokens do not participate in state updates, avoiding feature degradation. The KV cache size is constant, equal to the reference segment length plus the sliding window width, and does not grow with sequence length. - **DeepEncoder**: Inherits the encoder from DeepSeek-OCR, compressing a 1024×1024 PDF image into 256 visual tokens (16× compression). Supports two modes: Base (fixed resolution) and Gundam (dynamic resolution). - **Model Architecture**: 3B total parameter MoE with 500M activated parameters; all attention layers are replaced with R-SWA. Trained for 4000 steps from a DeepSeek-OCR checkpoint using 2 million OCR samples (single-page:multi-page = 9:1). Multi-page samples are randomly generated with 2-50 pages and sequence length up to 32K. ## Performance: Comprehensive SOTA - **OmniDocBench v1.5**: Overall score 93.23%, a 6.22 percentage point improvement over DeepSeek-OCR (87.01%). Text edit distance dropped from 0.073 to 0.038, formula CDM rose from 83.37 to 92.61, table TEDS rose from 84.97 to 90.93, and reading order edit distance dropped from 0.086 to 0.045. - **OmniDocBench v1.6**: Overall score 93.92%, end-to-end SOTA. - **Long Document Test**: Edit distance of 0.0572 for 20-page documents, and 0.1069 for 40+ page documents (Distinct-35 reaches 96.90%). The team notes that errors in 40+ page documents mainly stem from resolution limitations of DeepEncoder's multi-page mode, not R-SWA. - **Efficiency**: TPS on OmniDocBench reaches 5580 (DeepSeek-OCR: 4951). When outputting 6144 tokens, TPS is 7847 (DeepSeek-OCR: 5822), with advantages growing as output length increases. ## Open Source and Impact - Model weights and code have been open-sourced on GitHub and HuggingFace. - R-SWA is designed as a general decoding scheme applicable to long-output tasks such as ASR and translation. The team plans to validate its transferability in the next steps. - The technical director of the paper is credited as "YY", speculated by the industry to be Wei Haoran (former core author of DeepSeek OCR, now departed), who previously led GOT-OCR2.0 and the DeepSeek-OCR series. - Baidu's combination of PaddleOCR's industrial foundation with cutting-edge research is expected to drive OCR evolution from single-page recognition to full-book understanding.
Weibo Open-Sources 3B Small Model VibeThinker, Matching Trillion-Parameter Models in Verifiable Reasoning
Sina Weibo team recently open-sourced VibeThinker-3B, a dense reasoning model with only 3 billion parameters, achieving astonishing performance on verifiable reasoning tasks such as math competitions and programming, rivaling top models with hundreds of billions or even trillions of parameters. ## Core Performance Data - **AIME26**: Score 94.3, improved to 97.1 with test-time scaling strategy CLR. - **IMO-AnswerBench**: 76.4 standalone, 80.6 with CLR, on par with DeepSeek V3.2 (78.3, 671B params), GLM-5 (82.5, 744B params), and Kimi K2.5 (81.8, 1T params). - **LiveCodeBench v6**: Pass@1 of 80.2. - **Latest LeetCode Unseen Weekly Contests**: 123 out of 128 problems solved on first submission (96.1% pass rate) from April 25 to May 31, 2026. - **IFEval**: 93.4. ## Training Method VibeThinker-3B is built on Qwen2.5-Coder-3B using an upgraded Spectrum-to-Signal post-training pipeline, including: - **Curriculum Two-Stage SFT**: First stage covers general abilities like math, programming, STEM reasoning; second stage focuses on high-difficulty long-span samples, using diversity exploration distillation to retain multiple effective solution paths. - **Multi-Domain Reasoning Reinforcement Learning**: Reuses MGPO strategy, training sequentially on math, programming, and STEM tasks with a single 64K long context window. - **Offline Self-Distillation**: Selects high-quality trajectories from RL checkpoints, prioritizing trajectories that are correct but not yet well learned by the model based on learning potential scores, distilled into a unified student model. - **Instruction Reinforcement Learning**: Improves controllability over user prompts, using rule-based verifiers and scoring reward models for format-sensitive and open-ended instructions respectively. ## Parameter Compression-Coverage Hypothesis The research team proposes that different capabilities depend on parameter scale in different ways: verifiable reasoning (e.g., math, programming) is a highly compressible parameter-intensive capability with clear task structure and reliable feedback signals, allowing small models to approach frontier performance; while open-domain knowledge and general conversation rely on large parameter scale to cover facts and concepts. This hypothesis suggests small and large models are complementary, not substitutive. ## Open Source and Limitations The model is open-sourced on Hugging Face, GitHub, and ModelScope. The team explicitly states that the model performs poorly in domains requiring general knowledge, with its strengths concentrated on verifiable reasoning tasks.
Zhipu GLM-5.2 Released and Open-Sourced: First in Code Arena Among Globally Available Models, 1M Context Long-Form Programming Leap
On June 17, 2025, Zhipu AI officially released and open-sourced its flagship model GLM-5.2 under the MIT license, supporting 1 million token context. It achieved first place among globally available models in the Code Arena front-end development blind test (second only to the unreleased Claude Fable 5) and first place globally in Design Arena. ## Core Capabilities and Benchmark Performance - **Programming & Agents**: On the FrontierSWE benchmark, GLM-5.2 trails Claude Opus 4.8 by only ~1% and leads GPT-5.5 by ~1%; on PostTrainBench, it surpasses Opus 4.7 and GPT-5.5, second only to Opus 4.8; on the SWE-Marathon ultra-long task benchmark, it ranks second, behind the Opus series. On the standard programming benchmark Terminal-Bench 2.1, it scores 81.0 (GLM-5.1: 63.5), and on SWE-bench Pro, it scores 62.1 (GLM-5.1: 58.4). - **Long Context**: Supports 1 million token context, excelling on long-form programming benchmarks. The team reduced single-token FLOPs by approximately 2.9x via the IndexShare mechanism (shared indexer every 4 layers) and optimized the inference engine to improve long-context throughput. - **Thinking Effort Levels**: Introduces effort level control, allowing users to balance capability and computational cost. The Max level further extends programming capabilities. ## Architecture Innovations - **IndexShare Dynamic Sparse Attention**: A lightweight indexer is shared every 4 Transformer layers, reducing indexer computation overhead by 3/4, significantly lowering computation at 1M context. - **Improved MTP (Multi-Token Prediction)**: IndexShare is also applied in MTP layers, increasing speculative decoding acceptance length by 20%. ## Open Source and Ecosystem - Model weights are publicly available on HuggingFace and ModelScope, supporting inference frameworks such as transformers, vLLM, and SGLang. - On launch day, full adaptation was completed with domestic computing platforms including Huawei Ascend, T-Head, Moore Threads, Cambricon, Kunlunxin, Muxi, Haiguang, and Biren. - Hugging Face provided 6 hours of free global computing power for a domestic model for the first time. ## Reactions - Elon Musk replied on X that China might reach the Fable level by Q1 2027; Zhipu founder Tang Jie responded, "It won't take that long," sparking heated discussion among overseas netizens. - Silicon Valley VC a16z co-founder Marc Andreessen and Perplexity CEO Aravind Srinivas also took note. - US open-source model researcher Nathan Lambert commented that GLM-5.2 surpasses Google Gemini in agent capabilities, marking the first time a Chinese model has entered the global "Big Three" (Claude, OpenAI, Zhipu). ## Impact and Significance The release of GLM-5.2 marks the first time a Chinese open-source large model has reached the level of world-class closed-source models in agent programming and long-form complex tasks, breaking the competitive landscape previously dominated by models like Claude and GPT. Its open-source strategy (MIT license, no regional restrictions) provides global developers with a cutting-edge model that can be freely used, advancing AI programming from single-shot output to long-term engineering capabilities.
DeepSeek Image Recognition Goes Fully Live: Can Grade Homework but Misidentifies Its Own Founder
## Event Overview Just before the 2025 Dragon Boat Festival, DeepSeek fully rolled out its image recognition mode on its official platform, with the mobile app updated simultaneously. Previously in gray-scale testing, the feature is now available to all users. Tests by multiple media outlets and users show that DeepSeek's performance on visual understanding tasks is mixed: it can handle some complex tasks (e.g., grading elementary school math papers), but frequently makes errors in scenarios such as person identification, handwritten text recognition, and logical reasoning—even failing to recognize its own founder, Liang Wenfeng. ## Key Test Results ### Person Identification: Can't Recognize the Boss, but "Knows" Liu Qiangdong - **Liang Wenfeng (DeepSeek Founder)**: In multiple tests, DeepSeek misidentified him as Zhang Xiaolong, Su Hua, Wang Xiaochuan, Robin Li, etc., with different answers each time. - **Jensen Huang**: Correctly identified, but misjudged the douzhi (fermented bean drink) in his hand as milk; after switching to deep thinking mode, correctly inferred it was douzhi through reasoning. - **Liu Qiangdong**: Identified with certainty, sticking to the correct answer even when users tried to mislead. - **Lei Jun**: After uploading the image, the model prompted that it "may violate usage guidelines" and refused to identify. ### Homework Grading: Performs Well, but Logic Has Flaws - Tested a fourth-grade math exam (including image-based questions); DeepSeek successfully recognized and graded it, correctly identifying wrong answers. - However, in an average height calculation problem, the model recognized the original average height (140.2 cm) and new average height (139 cm), but incorrectly judged Honghong's height as 139 cm (option B), while the correct answer should be below 139 cm (option A). ### Handwritten Text Recognition: Low Accuracy - Tested with messy Chinese characters (including horizontal line interference, stroke粘连, and typos); 4 out of 7 characters were misidentified, indicating significant room for improvement in real-world handwritten text recognition. ### Other Visual Tasks - **Clock Recognition**: Misjudged a time around 6:04:50 as 6:00:50, and insisted on it. - **Artifact Recognition**: Correctly identified it as Mughal Empire style and provided detailed analysis of craftsmanship, but could not find the specific origin. - **Find Matching Socks**: Failed to correctly identify the identical pair (correct answer: first row, third sock and third row, second sock). - **Piano Chord Recognition**: Uploaded a real photo of piano playing and asked "what chord is being played"; DeepSeek gave a wrong answer (correct answer: ACE). ## Reactions and Comparisons - **User Feedback**: Many users on social media found that DeepSeek also misidentified well-known figures like He Tongxue. Some users joked that it "talks nonsense with a straight face." - **Competitor Comparison**: - **Doubao**: Performed more accurately and faster on tasks like person identification, homework grading, and clock recognition. - **Gemini 3.5 flash, GPT 5.5**: Both gave wrong answers on the piano chord recognition task. - **Claude Sonnet 4.6**: Directly refused to answer that task. - **Opus 4.8**: Reflected and gave correct reasoning on a Tank Battle logic problem, standing out. ## Impact and Next Steps - **Technical Questions**: Developers are curious whether this mode is related to DeepSeek 4.1, whether it uses native multimodal technology, and when the multimodal API will be released. DeepSeek multimodal team researcher Xiaokang Chen did not respond. - **Industry Observation**: The launch of DeepSeek's image recognition feature fills a gap in multimodal capabilities, but current accuracy is low, especially in person identification and handwritten text recognition. It is expected that technical documentation or optimizations may follow.
Zhipu GLM-5.2 Open-Sourced and Tops Coding Leaderboard, with Significant Frontend Improvements
Zhipu AI fully released the GLM-5.2 model on June 13, 2025. It ranks second globally on the LMSYS Code Arena: Frontend leaderboard, behind only Claude Fable 5, making it the top open-source model. It also achieved first place globally in Design Arena. GLM-5.2 supports a 1M context, excels in long-range tasks, and is open-sourced under the MIT license. ## Leaderboard Performance and Evaluation - **Code Arena: Frontend**: GLM-5.2 ranks second, scoring 29 points higher than Claude Opus 4.8 Thinking. It is second in the React sub-leaderboard and fourth in the HTML sub-leaderboard, and ranks first in subcategories such as Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Goods, and Games & Simulations. - **Design Arena**: First place globally, demonstrating the model's aesthetic and design capabilities. - **Eight authoritative benchmarks**: Strong performance, though specific scores were not detailed in the report. ## Frontend Capability Improvements Multiple independent evaluators noted that GLM-5.2's frontend capabilities have seen a qualitative leap over its predecessors (GLM-5.0/5.1). Typical test cases include: - **Cyberpunk version of Along the River During the Qingming Festival**: GLM-5.0 produced rough results, while GLM-5.2 generated complete building and object forms, though the cyberpunk feel was strong but the Qingming flavor slightly inferior to Opus 4.8. - **Infinite text adventure game**: GLM-5.0 had layout collapse, while GLM-5.2 had normal layout and cool animation effects. - **Gomoku (Five-in-a-Row) game**: GLM-5.0 had poor design sense; GLM-5.2 optimized the board, background, and color scheme. - **Neon Runner**: GLM-5.2 produced strong depth perception in the scene, supporting double jumps and explosion effects. - **3D Solar System**: GLM-5.2 used abstract line design, which, while not realistic, had a sense of design. Evaluators believe that GLM-5.2 has been specifically trained for frontend tasks, and its strong design sense may lead to stylistic convergence, but overall results are close to or even locally surpass Opus 4.8. ## Long Context and Engineering Capabilities GLM-5.2 supports a 1M context and performs outstandingly in real engineering tasks: - **Full codebase understanding**: In the Appsmith project, GLM-5.2 accurately outlined the architecture, identified coupling points, and provided a refactoring roadmap, with coverage depth superior to CodeX. - **Cross-file bug tracking**: In the OpenWebUI project, GLM-5.2 located the link issue between SSE fragmentation and backend parsing, providing both frontend and backend fixes. - **New feature addition**: In OpenWebUI, adding a "session summary export to Markdown" feature, GLM-5.2 split the implementation into five layers, and all 38 backend tests passed. - **Multi-task delivery**: In building a UK student accommodation industry research package, GLM-5.2 output a complete folder at once, including charts, reports, and scripts. Evaluators note that the 1M context is suitable for complex tasks like full codebase understanding, cross-file bug tracking, and long-term refactoring, but may lead to over-engineering in simple tasks. ## Ecosystem and Tools Zhipu also launched ZCode (zcode.z.ai), an agent development kit similar to Claude Code and OpenAI Codex, supporting Windows and macOS. Evaluations show that GLM-5.2 generates significantly better UI results in ZCode than in Claude Code, likely due to ZCode's engineering optimizations. New ZCode users can use it for free for 5 days, and subscribers enjoy a 150% quota. ## Industry Impact The open-sourcing and leaderboard performance of GLM-5.2 mark the first time a domestic model has entered the "Big Three" of AI programming (Claude, OpenAI, Zhipu), pushing Google Gemini out of the top ranks. Against the backdrop of restricted access to Claude Fable 5 due to safety controversies, Zhipu emphasizes that "frontier intelligence should belong to everyone," promoting the open-source ecosystem. ## Limitations and Outlook Despite significant frontend improvements, GLM-5.2 still lags behind Opus 4.8 in overall capabilities, especially in processing time, depth of thought, and first-attempt accuracy. Evaluators suggest avoiding the 1M context for simple tasks to maintain efficiency. The model API will be available next week, and the open-source version follows the MIT license.
Anthropic Fable 5 Ban: From Launch to Global Takedown in Four Days, First-Ever AI Model Export Controls
On June 9, 2026, Anthropic released its most powerful model, Claude Fable 5, along with Mythos 5 for security teams. Just four days later, the U.S. Commerce Department imposed export controls on both models citing national security, forcing Anthropic to take them down globally. This marks the first time an AI model itself has been subject to export controls, sending shockwaves through the industry. ## Timeline: 72 Hours from Launch to Ban - **June 9**: Anthropic launches Fable 5, claiming top-tier performance across nearly all AI benchmarks, surpassing GPT-5.5 and Gemini 3.1 Pro. - **June 11 evening**: Amazon CEO Andy Jassy, Anthropic's largest investor, reports to the White House that Fable 5 has a jailbreak vulnerability that could bypass safety guards to access Mythos's underlying cybersecurity capabilities. At least five other companies subsequently express similar concerns to the government. - **June 12 morning**: Emergency White House meeting with the Treasury Secretary, cybersecurity chief, and other senior officials. Amazon's report is sent to the NSA for review. - **June 12 afternoon**: The White House calls Anthropic CEO Dario Amodei three times, demanding the vulnerability be fixed or the model voluntarily taken down. Amodei refuses, arguing the jailbreak is narrow and other models have similar issues. - **June 12, 5:21 PM**: Commerce Secretary Howard Lutnick issues a formal export control letter, banning access to Fable 5 and Mythos 5 outside the U.S. and by foreign nationals within the U.S. Anthropic cuts global access within 90 minutes. ## Core Controversy: Technical Issue or Communication Failure? ### Reality of the Jailbreak Vulnerability Anthropic insists the jailbreak is a "narrow potential jailbreak," not a general one, and that models like GPT-5.5 can achieve similar results. But the government views Anthropic's response as dismissive and lacking urgency to fix it. ### Communication and Ideological Clash Axios reports the ban stems from ideological conflict and communication breakdown. Anthropic published a blog post denying Amazon's findings and hired a cybersecurity expert considered a "radical opponent" by the government. The expert was later publicly celebrated by a former official fired by Trump, seen as a provocation. A government official said: "Anthropic has done a terrible job trying to talk to the government. They speak a different language." ### Amazon's "Betrayal" Amazon, Anthropic's largest investor (over $13 billion total), proactively reported the issue, shocking the industry. However, reports indicate Amazon was not the only whistleblower. ## Reactions and Impact ### Cybersecurity Experts Demand Unban On June 14, dozens of top U.S. cybersecurity experts (including Alex Stamos, Bruce Schneier) published an open letter demanding the ban be lifted. Key arguments: - Mythos is not a unique weapon; GPT-5.5 can do similar things. - The ban hurts defenders (code auditing capabilities), not attackers. - Other countries are catching up fast; the ban weakens U.S. defenses. ### Failed Negotiations On June 15, Anthropic executives flew to Washington for talks but failed to convince the government to lift controls. The Commerce Department offered a conditional concession: if the jailbreak issue is resolved, consumer access to Fable 5 could be restored. But "perfectly preventing jailbreaks" is technically nearly impossible. ### Industry Impact - **OpenAI Dragged In**: Anthropic claims GPT-5.5 can also achieve similar jailbreaks, trying to pull OpenAI into the controversy. Reports suggest GPT-5.6 may be delayed as a result. - **Precedent Set**: This is the first time an AI model has been treated as a strategic asset under export controls. The White House says controls are unlikely to extend to other companies, but the industry sees it as the beginning of a "licensing system." - **Developer Ecosystem Hit**: Before Fable 5 went offline, developers had already used it for heavy-duty projects (e.g., rewriting a DOS game in 30 minutes, building a 3D Hogwarts). The planned Developer Day was forced to switch to Opus 4.8. ### Other Events - **Subscription Restriction Reversed**: Anthropic rescinded a ban on third-party programming calls using Claude Code subscription quotas, allowing apps like OpenClaw to continue using subscriptions. - **Class Action Lawsuit**: On the same day, Anthropic was hit with a class action lawsuit over the Max plan's actual usage being far lower than advertised (Max 20x actually only 6-8x). - **Zhipu GLM-5.2 Fully Released**: On the same day Fable 5 was banned, Zhipu announced GLM-5.2 is open to all subscribers and promised open-source, emphasizing that "frontier intelligence should belong to everyone." ## Deeper Significance: A Turning Point in AI Regulation This event marks a shift in AI regulation from hardware (chips) to software (the model itself). Anthropic CEO Dario Amodei once compared AI's danger to nuclear weapons, but when the government asked him to shut down the system, he refused, leading to export controls. As Axios commented: "This is not a technical issue; it's a restructuring of power." When model capabilities reach a certain threshold, the state will not leave ultimate power in private hands. The Fable 5 ban may be just the prelude to the era of AI regulation.
Brazilian Municipal IT Company's Rio 3.5 Model Exposed as Wrapper of Domestic Nex and Qwen
An IT company under the Rio de Janeiro city government recently open-sourced a 397B-parameter large model named Rio 3.5, which achieved SOTA results on multiple benchmarks, even surpassing Alibaba's Qwen 3.7 Plus, causing a stir in the community. However, the Nex-AGI team initiated by Shanghai Chuangzhi Institute quickly released evidence showing that Rio 3.5 is actually a fusion of Nex N2 Pro and Qwen 3.5. ## Evidence of Wrapping The Nex team provided two key pieces of evidence on GitHub: - **Model Self-Identification Test**: After removing Rio's hardcoded system prompts, the model was asked "Who are you?" 120 times. 79% of responses claimed to be "Nex", 73% mentioned "Nex-AGI", and never identified as "Rio". The model could also verbatim recite Nex's institutional introduction, such as "Large Model Ecosystem Alliance" and "Shanghai Chuangzhi Institute". - **Weight Analysis**: Each of Rio's 60 layers was verified, showing that every weight tensor precisely lies on the line between Nex and Qwen, with a stable mixture ratio of approximately 0.57 Nex plus 0.43 Qwen. The collinearity cos_fit ranged from 0.984 to 0.993, with statistical deviations of thousands of standard deviations, ruling out the possibility of independent training coincidence. ## Aftermath Faced with irrefutable evidence, the Rio team removed the model from HuggingFace, leaving only an apology statement acknowledging the use of Nex and Qwen to build the model, but claiming that "the wrong version without final distillation was uploaded". The Nex team emphasized that the open-source community must adhere to the baseline of attribution and acknowledgment. ## Similar Cases This is not the first time a domestic model has been wrapped: - **Cursor Composer 2** (March 2025): A self-proclaimed self-developed code model was exposed to have API request paths containing "kimi", later admitting to using Kimi through the Fireworks AI platform, but without mentioning it at release. - **Stanford Llama3-V** (2024): Claimed to train a model surpassing GPT-4V for only $500, but its code and weights were nearly identical to Tsinghua's MiniCPM-Llama3-V 2.5, even reproducing the same recognition errors on unpublished "Tsinghua Bamboo Slips". These incidents show that wrapping domestic models is common across universities, enterprises, and official institutions.
Kimi K2.7 Code Released: Enhanced Code and Agent Capabilities with 30% Token Reduction
Moonshot AI recently released Kimi K2.7 Code, the first code-specialized model in the K2 series, now open-sourced on HuggingFace. The model features comprehensive upgrades in code generation, agent execution, and long-horizon tasks, with average token consumption reduced by approximately 30% compared to its predecessor K2.6. However, some users report severe quota limitations affecting the actual experience. ## Core Capability Improvements - **Code Benchmarks**: Kimi Code Bench v2 score increased by 21.8% (50.9→62.0), Program-Bench by 11% (48.3→53.6), and MLS Bench Lite by 31.5% (26.7→35.1), the latter approaching GPT-5.5's 35.5. - **Agent Benchmarks**: Kimi Claw 24/7 Bench improved by 9.3% (42.9→46.9), MCP Atlas by 9.5% (69.4→76.0), and MCP Mark Verified by 11.4% (72.8→81.1), partially surpassing Claude Opus 4.8 (76.4) in tool-calling scenarios. - **Long-Horizon Task Optimization**: Mitigated the "overthinking" issue, reducing average token consumption by 30% and improving success rates for long-cycle complex tasks. ## Real-World Performance - **Physics Simulation**: In scenarios like black holes and water wave rendering, K2.7 Code produces realistic outputs, with water wave rendering outperforming GPT-5.5 and Claude Opus 4.8. - **Game Development**: When generating an HTML version of Super Mario, K2.7 Code can produce a playable first level, but character and map details remain abstract, with a noticeable gap compared to Claude Fable 5. - **Frontend Tasks**: In 9 frontend examples, standalone tests show good results, but batch execution leads to lazy behavior and inconsistent output quality. ## Pricing and Quota Controversy - **Pricing**: Standard input at 6.5 RMB/1M tokens, output at 27 RMB/1M tokens, cache input at 1.3 RMB/1M tokens, consistent with K2.6. - **Quota Limits**: Multiple users report that the weekly quota for the basic Code plan is exhausted after a few tests, resulting in numerous 429 (rate limit) and 402 (insufficient quota) errors, severely impacting development workflows. One user noted, "Running a single example used 63% of my weekly quota." - **Usage Requirement**: Thinking mode must be enabled; disabling it causes API errors or fallback to K2.6. ## Architecture and Deployment - **Model Architecture**: Continues the MoE design with 1T total parameters, 32B activated parameters, 384 experts with 8 selected per step, 1 shared expert, and a context length of 256K tokens. The vision component uses a MoonViT encoder (400M parameters), supporting image and video input. - **Open Source and Deployment**: Released under a Modified MIT License, supports deployment via vLLM, SGLang, and KTransformers, with native INT4 quantization. - **High-Speed Version Preview**: A 6x faster version will launch on June 15, with output speeds of approximately 180 tokens/s (typical scenarios) and up to 260 tokens/s for short contexts, priced at 2x the standard version. ## Industry Rankings - Second overall in ErdosBench, trailing only Claude Fable 5 max. - First among open-source models in SWE-bench and Terminal-Bench 2.1, third in Vibe Code Bench, and second in ProgramBench. - In Weco's independent research task evaluation, K2.7 Code ranked fifth with a score of 0.747, and first in machine learning engineering specialization.
Brazil's Rio Open-Source Model Rio 3.5 Accused of Being a 'Wrapper' Around Chinese Models, Removed After Evidence Surfaces
The large language model Rio 3.5 Open 397B, open-sourced in June 2025 by IplanRIO, an IT company under the Rio de Janeiro city government, achieved SOTA results on multiple benchmarks, drawing community attention. However, within 24 hours of release, the Nex-AGI team, initiated by the Shanghai Institute of Intelligent Science and Technology, publicly accused it of being a 'wrapper,' claiming Rio 3.5 was essentially a hybrid of Nex N2 Pro and Alibaba's Qwen 3.5. The Nex team provided two key pieces of evidence: first, after removing hardcoded system prompts, the model self-identified as 'Nex' with 79% probability and could recite Nex-specific institutional introductions; second, mathematical analysis of the model's 60-layer weights showed that every tensor fell precisely on the line connecting Nex and Qwen weights, with a stable mixing ratio of approximately 0.57:0.43, collinearity cos_fit ranging from 0.984 to 0.993, and statistical deviations of thousands of standard deviations, ruling out coincidence from independent training. Faced with irrefutable evidence, the Rio team issued an apology on the HuggingFace page, admitting to using Nex and Qwen to build the model but claiming they 'uploaded an incorrect version that had not undergone final distillation,' and subsequently removed the model. This incident is not isolated; similar cases such as Cursor being accused of wrapping Kimi and the Stanford team wrapping MiniCPM have frequently occurred, highlighting the importance of model attribution and credit in the open-source community.