Hasty Briefsbeta

All tags

ai model

15 stories total

Bilingual

Seedance 2.5
2 months ago
- Seedance 2.5 generates up to 30 seconds of high-quality audio-video in a single pass, supports multi-round extensions, and improves shot transitions and scene changes for coherent long-form storytelling.
- Multimodal referencing is fully upgraded, allowing up to 30 images, 10 videos, and 10 audio clips as inputs, with enhanced capabilities for clay render, motion, and creative references.
- Editing is more precise with timestamp-level control, green screen, camera perspective, and reference-based editing, supporting professional film and advertising workflows.
- The model is integrated into education (e.g., historical visualization) and industry (e.g., synthetic data for robotics and autonomous driving), unlocking real-world productivity.
- Seedance 2.5 is available on Jimeng AI and Doubao Pro, with API access via BytePlus ModelArk coming soon.
DeepSeek-AI/DeepSeek-V4-Flash-0731
2 months ago
- DeepSeek-V4-Flash-0731 can be used with Transformers via pipeline or loading model directly, with code examples provided.
- It is also deployable using vLLM (pip install or Docker) with DSpark speculative decoding enabled via a specific flag.
- SGLang serves the model via pip or Docker, and the Docker Model Runner command is provided.
- Benchmark results show DeepSeek-V4-Flash-0731 outperforms previous versions and is competitive with proprietary models like Opus-4.8.
- The model uses a custom chat template with a dedicated encoding folder, and supports three reasoning effort levels: low, high, max.
- Local deployment recommendations include sampling parameters (temperature=1.0, top_p=0.95 for agentic tasks) and a max output length of 384K tokens.
- The model is licensed under MIT License, and citation information is provided.
Inkling-Small
2 months ago
- Inkling-Small is an efficient open-weights model that achieves comparable performance to Inkling at a quarter of its size.
- It is a Mixture-of-Experts transformer with 276B total parameters (12B active), trained on NVIDIA GB300 NVL72 systems.
- Features include native reasoning over audio and images, variable thinking effort, a 1M token context window, and strong benchmark performance.
- Inkling-Small surpasses Inkling on reasoning and agentic tasks while being more efficient, and matches or exceeds other models in its weight class.
- Full weights are released, and the model is available for fine-tuning on Tinker and chat via Tinker Playground.
- Improvements in training include a better pre-training data mix, post-training with on-policy distillation, and extended agentic coding RL.
- Multimodal capabilities are native, using an encoder-free architecture for audio and images, with strong performance on visual and audio benchmarks.
- Safety and epistemics are well-handled, with calibration, instruction following, and resistance to censorship matching Inkling's performance.
Introducing Claude Opus 5
2 months ago
- Anthropic launched Claude Opus 5, described as thoughtful and proactive, offering frontier intelligence close to Claude Fable 5 at half the price.
- Opus 5 leads the Artificial Analysis leaderboard, is priced the same as Opus 4.8, and offers a 'fast mode' at double the cost.
- In a test, Opus 5 proactively wrote a computer vision pipeline to reconstruct a 3D FreeCAD model from raw pixels when unable to view a drawing directly.
- Opus 5 improved at finding cybersecurity vulnerabilities due to general capability gains, but was intentionally not trained on exploitation, distinguishing it from Mythos 5.
- Anthropic published a prompting guide for Claude Opus 5, and Thariq Shihipar wrote on context engineering for Claude 5 models.
- The author mentioned kayaking with sea otters and noted a pelican model issue with missing bicycle wheels.
Kimi-K3 Releases on HuggingFace 7/27
2 months ago
- Moonshot AI's Kimi-K3 is the next generation open frontier model, scheduled for release on July 27, 2026.
- It is the world's first open 3T-class model, focusing on long-horizon coding, knowledge work, and reasoning.
- Key features include a new architecture with Kimi Delta Attention and Attention Residuals, native agentic capabilities (tool calling, browsing, multi-step planning), and an extended context window for repository-scale code understanding.
- The model will have open weights released on the same page, and users can sign up for a single notification email when weights become public.
- Over 1,820 users are already waiting for the release.
Apertus 1.5, Swiss open-weight, open-source, open training data model
2 months ago
- Apertus 1.5 is an update to the 8B and 70B models, adding image understanding, thinking mode, longer context, better instruction-following, and improved tool use.
- The models are fully open source with open weights, data, values, and training details, released on July 24, 2026.
- Key enhancements include native image understanding, optional thinking mode for reasoning, a 262,144-token context window, improved adherence to instructions, and better tool integration.
- The 70B model can interpret multimodal inputs like technical drawings, and a technical report with benchmarks will be published soon.
Claude Opus 5
2 months ago
- Claude Opus 5 is a new model from Anthropic, offering near-frontier intelligence at half the price of Claude Fable 5.
- It achieves state-of-the-art results on coding and knowledge work benchmarks like Frontier-Bench and GDPval-AA, though lags behind Mythos 5 on cybersecurity.
- Opus 5 is designed for everyday use, with improved efficiency and cost-effectiveness compared to its predecessor Opus 4.8.
- It excels in software engineering, knowledge work, and scientific research tasks, often surpassing other models at lower cost.
- The model demonstrates strong agency and thoroughness, such as writing its own computer vision pipeline to solve a task and fixing bugs missed by human patches.
- Early-access customers report significant improvements in areas like financial modeling, legal work, code review, and agentic workflows.
- Opus 5 is highly aligned with safety principles, showing low rates of deceptive behavior and susceptibility to misuse.
- It includes safeguards for cybersecurity and biology, with automatic fallbacks and a Cyber Verification Program for enterprises.
- Opus 5 is available today at $5 per million input tokens and $25 per million output tokens, with a Fast mode option.
Agent Security Is a Systems Problem
2 months ago
- Agent security should be approached as a systems problem, treating the AI model as an untrusted component and enforcing security invariants at the system level.
- Increasing model robustness alone is insufficient; techniques from systems security must complement existing efforts.
- Core principles from decades of systems security research provide a foundation for designing agentic systems with predictable guarantees.
- Analysis of eleven real-world attacks on agents shows that systems principles could have prevented these attacks if realized.
- Research challenges must be addressed to implement these systems security principles in agents.
Kimi K3: Open Frontier Intelligence
2 months ago
- Introduction of Kimi K3: a 2.8T-parameter open model with native vision and a 1-million-token context window.
- Architecture based on Kimi Delta Attention and Attention Residuals, with efficient scaling and 16 active experts out of 896.
- Strong performance in long-horizon coding, including kernel optimization, compiler development, and game dev with visual reasoning.
- Capabilities in chip design, research automation, and knowledge work, such as generating interactive reports and analyzing data.
- Availability on various platforms (Kimi.com, Work, Code, API) with pricing details and enterprise options.
- Benchmarks show competitive results in coding and multimodal tasks, though trailing top proprietary models like Claude Fable 5 and GPT 5.6 Sol.
- Limitations include sensitivity to thinking history and excessive proactiveness, requiring careful handling in applications.
Grok 4.5
3 months ago
- Grok 4.5 is released in collaboration with SpaceXAI as the most intelligent model, designed for tasks beyond software engineering.
- It can handle difficult, long-running tasks in various fields like software engineering, data science, finance, and legal work by creatively using tools.
- Cursor subscription plans include significant model usage with double usage for the first week and added safeguards for cybersecurity.
- Grok 4.5 is a mixture-of-experts model trained with trillions of tokens from Cursor data, learning from software and developer-agent interactions.
- Training data was broadened to include high-quality STEM tasks, research papers, and knowledge work for proficiency across domains.
- Reinforcement learning was applied in realistic environments to teach problem investigation, tool usage, error recovery, and result verification.
- A distributed agent system scaled environment construction, using previous models to accelerate progress on difficult problems.
- Grok 4.5 is available on multiple platforms with pricing at $2/M input tokens and $6/M output tokens, and a fast variant at higher rates.
- Grok 4.5 and Composer 2.5 will both be supported, with future models planned for the Composer size.
SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence
3 months ago
- Cognition launches SWE-1.7, a frontier-level AI model for software engineering at lower cost, advancing cost-performance.
- SWE-1.7 achieves 42.3% on FrontierCode 1.1, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual.
- Training improvements include preserving entropy via top-p sampling replay, multi-cluster infrastructure for scalability, and fault tolerance.
- Self-compaction enables long-horizon tasks by summarizing working state, with alternating length penalty to balance efficiency.
- High-quality data curation ensures task difficulty and prevents cheating, enhancing model alignment and trustworthiness.
- SWE-1.7 shows condensed chain-of-thought reasoning and thorough codebase exploration compared to base model Kimi K2.7 Code.
Tencent Hy3
3 months ago
- The text provides instructions on how to use the tencent/Hy3 model with various libraries, inference providers, notebooks, and local apps.
- Detailed examples are given for using Hy3 with Transformers, vLLM, SGLang, and Docker, including code snippets for installation, server setup, and API calls.
- Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters, developed by Tencent Hy Team, offering strong capabilities in agent tasks, reasoning, and long-context handling.
- The model shows improvements in stability, knowledge accuracy, and multi-turn intent tracking, with reduced hallucination and error rates based on internal evaluations.
- Deployment recommendations include using vLLM or SGLang on 8 GPUs with specific configurations, and the model is open-sourced under Apache License 2.0, available on multiple platforms.
China's LongCat-2.0 Becomes the Biggest AI Model Without Nvidia Chips
3 months ago
- Meituan released LongCat-2.0, a 1.6 trillion parameter open-source LLM, trained end-to-end on domestic Chinese hardware without NVIDIA chips.
- It's the industry's first trillion-parameter model to complete both training and inference on domestic hardware, marking a major milestone for China's tech self-reliance.
- The model uses Huawei's Atlas-950 SuperPoDs and HCCL software, mirroring NVIDIA's setup, and shows strong performance but still trails top global models like GPT-5.5.
- The launch demonstrates frontier-scale training is viable on Chinese hardware, potentially narrowing the gap with Western systems faster than expected.
- Challenges remain, including a less mature software ecosystem and memory limits compared to NVIDIA's banned H800 chips.
Claude Sonnet 5
3 months ago
- Claude Sonnet 5 is the most agentic Sonnet model, capable of planning, tool use, and autonomous operation.
- It offers performance close to Opus 4.8 at lower prices, with improvements in reasoning, tool use, coding, and knowledge work over Sonnet 4.6.
- Safety assessments show it has a lower rate of undesirable behaviors than Sonnet 4.6 and is safer in agentic contexts, but it has a lower ability to perform cybersecurity tasks compared to Opus models.
- It is available across all plans with introductory pricing of $2 per million input tokens and $10 per million output tokens until August 31, 2026, after which it increases to $3 and $15 respectively.
- Feedback from early access partners highlights its improved agentic capabilities, such as completing complex tasks, self-checking outputs, and efficient performance in coding and automation.
- The model includes cyber safeguards enabled by default, similar to those in Claude Opus 4.7 and 4.8, to block dangerous cyber usage.
Liquid AI releases a 230M model optimized for phones, Raspberry Pi, and robots
3 months ago
- LFM2.5-230M is a small, fast model for fine-tuning and deployment in agentic workflows, with fast inference on various devices including cloud GPUs and CPUs.
- It was pre-trained on 19T tokens and post-trained via supervised fine-tuning, direct preference optimization, and multi-domain reinforcement learning, balancing capabilities and adaptability.
- The model was tested on a humanoid robot for skill selection, turning natural-language instructions into structured plans using NVIDIA's SONIC framework.
- Benchmarks show LFM2.5-230M competes with or outperforms larger models in knowledge, instruction following, data extraction, and tool use tasks.
- It supports fast inference across ecosystems like llama.cpp, MLX, vLLM, and ONNX, with optimized CPU and GPU performance for low latency and high throughput.
- LFM2.5-230M is open-weight, available on Hugging Face, and designed for large-scale data extraction or lightweight on-device workloads, but not for reasoning-heavy tasks.

firstprev0next

ai model

Seedance 2.5

DeepSeek-AI/DeepSeek-V4-Flash-0731

Inkling-Small

Introducing Claude Opus 5

Kimi-K3 Releases on HuggingFace 7/27

Apertus 1.5, Swiss open-weight, open-source, open training data model

Claude Opus 5

Agent Security Is a Systems Problem

Kimi K3: Open Frontier Intelligence

Grok 4.5

SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence

Tencent Hy3

China's LongCat-2.0 Becomes the Biggest AI Model Without Nvidia Chips

Claude Sonnet 5

Liquid AI releases a 230M model optimized for phones, Raspberry Pi, and robots