Tag: large language model

  • Google Gemini 3.8 Flash: A New AI Model for Autonomous Coding and AI Agents

    Google Gemini 3.8 Flash: A New AI Model for Autonomous Coding and AI Agents

    Google has released Gemini 3.8 Flash, its latest AI model designed to power autonomous coding agents and handle complex enterprise tasks with minimal human intervention. Available since September 2, 2026, this model is positioned as Google’s most capable Flash variant yet, targeting long-horizon software development workflows.

    Massive Context for Complex Codebases

    Gemini 3.8 Flash supports a context window of up to 1 million tokens and can generate up to 64,000 output tokens. It accepts text, images, audio, and video inputs. Developers can adjust the model’s effort level to low, medium, or high, allowing teams to balance quality, cost, and speed according to their specific tasks.

    This large context is especially valuable for software agents that need to process entire codebases, including multiple files, tests, logs, and configuration details in a single workflow. The model can handle complex multi-file refactors and deterministic tool tasks effectively.

    Google Antigravity now uses Gemini 3.8 Flash as its default agent model, and its software development kit defaults to it as well. This integration enables agents to plan tasks, make code changes, call tools, and review results within a unified workflow.

    Built-in Tools for Autonomous Action

    The API provides Gemini 3.8 Flash with access to built-in tools such as function calls, code execution, file search, URL context, and structured output. Computer use remains in preview. These tools are essential for agents that need to act on software, gather information, inspect files, and deliver clear results.

    Benchmark Results Highlight Long-Task Strengths

    Google reports a 73.7% score for Gemini 3.8 Flash on DeepSWE v1.1, a benchmark for long-horizon software tasks, compared to 65.3% for Gemini 3.7 Flash. On Terminal-Bench 2.1, an agentic terminal coding benchmark, the model achieved 89.4%, up from 85.8% for its predecessor.

    These results demonstrate the model’s focus on tasks requiring multiple actions rather than single code answers. The model can take extra analysis steps and make repeated tool calls for harder tasks, though this may increase token usage, especially at higher effort levels. Lower effort settings can reduce costs when deep analysis isn’t required.

    Additional benchmarks include 61.4% on the Vals Finance Agent v2 benchmark, a 10% pass rate on the Harvey Legal Agent Benchmark, and 54.9% on HLE-Verified, indicating broader applicability across finance, law, science, and other expert domains.

    Competitive Pricing for Production Workflows

    Gemini 3.8 Flash launched at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens, with these rates guaranteed through December 31, 2026. After that, standard rates will increase to $1.50 per 1 million input tokens and $7.50 per 1 million output tokens.

    This pricing is significant for agent systems that make many model calls per task. The true measure becomes the cost of completing a task, not just the price of a single response. However, lower token costs don’t eliminate the need for robust safeguards, including tests, code review, access limits, sandbox controls, and approval steps for serious software work.

    Expanding the AI Agent Ecosystem

    Google’s broader Gemini 3.8 release includes Gemini 3.8 Flash Cyber, a separate model for trusted defenders that can find software flaws and create patches. The Fairwind Program provides access to selected government groups, critical infrastructure operators, and software maintainers.

    Additionally, two Gemini 3.8 Live models were released on September 15 for real-time voice agents, expanding the family beyond software work into live agent systems with audio input and output.

    Gemini 3.8 Flash represents a shift from simple code assistance toward software agents capable of handling longer tasks, using tools, inspecting results, and continuing work autonomously. With its 1 million-token context, adjustable effort, low price, and strong long-task performance, Google has created a model aimed at the practical side of autonomous AI. The real test now goes beyond benchmark scores: whether these agents can complete real software work with fewer errors, fewer retries, and less human help.

    FAQs

    1. What is Gemini 3.8 Flash?
    Gemini 3.8 Flash is Google’s latest Flash model, designed for complex reasoning, autonomous AI agents, software development, and enterprise workloads.

    2. What is the context window of Gemini 3.8 Flash?
    It supports a context window of up to 1 million tokens, allowing agents to work with large amounts of project and reference information.

    3. Can Gemini 3.8 Flash perform autonomous coding?
    Yes. It can use tools, inspect files, modify code, execute commands, and evaluate results, supporting multi-step software-development workflows.

    4. How much does Gemini 3.8 Flash cost?
    The launch pricing is $0.75 per 1 million input tokens and $3.75 per 1 million output tokens, with standard rates scheduled to increase from January 1, 2027.

    5. What makes Gemini 3.8 Flash different from a basic coding assistant?
    It is designed for longer workflows where an agent can plan, take multiple actions, use external tools, test its work, and continue iterating rather than simply generating code.

  • DeepSeek’s V4.1 Flash: 9 Key Facts About Its 552B-Parameter AI Model

    DeepSeek’s V4.1 Flash: 9 Key Facts About Its 552B-Parameter AI Model

    DeepSeek has introduced its latest artificial intelligence system, V4.1 Flash, a 552-billion-parameter model designed to deliver powerful general-purpose capabilities while keeping computational demands in check. Below are the nine essential details about this new release.

    1. Mixture-of-Experts Architecture: V4.1 Flash uses a Mixture-of-Experts (MoE) design that balances high-level capability with efficient computing, making it suitable for complex workloads without excessive resource use.
    2. 552 Billion Total Parameters: With 552 billion parameters, this model ranks among the largest AI systems built for advanced general-purpose applications, giving it substantial capacity for learning and reasoning.
    3. 8 Billion Active Parameters on Input: During input processing, only about 8 billion parameters are activated, significantly reducing computational requirements while still handling complex tasks effectively.
    4. 16 Billion Active Parameters During Generation: When producing responses, the model activates roughly 16 billion parameters, providing extra capacity for generating higher-quality outputs and completing nuanced instructions.
    5. One-Million-Token Context Window: V4.1 Flash supports a context window of up to one million tokens, enabling it to process extremely long documents and maintain coherence across lengthy interactions.
    6. Native Vision Capabilities: The model includes built-in multimodal support, allowing it to understand visual information alongside text for a broader range of AI applications.
    7. New Causal Encoder-Decoder Architecture: DeepSeek introduces a fresh architecture that separates input understanding from generation, improving efficiency during demanding inference tasks.
    8. Reduced KV Cache Requirements: The design significantly lowers KV-cache memory consumption, which improves deployment efficiency across large-scale AI infrastructure.
    9. Optimized for Faster Inference: V4.1 Flash is built around efficient inference, targeting lower computational overhead while maintaining strong performance in coding, reasoning, and agentic workloads.

    DeepSeek’s V4.1 Flash demonstrates a clear focus on balancing scale with practicality, making it a notable entry in the evolving landscape of high-parameter AI models.

  • Moonshot AI Opens Kimi K3 for Public Download, Challenging OpenAI and Anthropic

    Moonshot AI Opens Kimi K3 for Public Download, Challenging OpenAI and Anthropic

    Moonshot AI has officially released its latest large language model, Kimi K3, for public download. The move allows developers worldwide to access, deploy, customize, and experiment with the model, marking a significant step in the global AI race.

    The Chinese AI startup gained international attention after Kimi K3 demonstrated competitive reasoning, coding, and multilingual performance against leading AI models from OpenAI and Anthropic. Reports indicate that the model’s capabilities surprised industry observers, approaching performance levels associated with advanced OpenAI systems across several benchmark evaluations.

    Kimi K3’s strong reasoning and coding results have positioned it as another notable competitor challenging Anthropic’s latest large language models. The public release includes open model weights, encouraging transparency, customization, fine-tuning, and community-driven improvements for developers and AI researchers.

    Organizations can integrate Kimi K3 into applications, automate workflows, build AI assistants, and accelerate software development with flexible deployment options. The availability of Kimi K3 expands choices beyond proprietary AI systems, potentially accelerating innovation, reducing costs, and encouraging broader adoption across industries.

    The release reflects intensifying global competition as companies rapidly develop powerful foundation models with stronger reasoning, efficiency, and accessibility for enterprises.

  • Inkling AI: Mira Murati’s Open-Weight LLM Challenges ChatGPT and Claude

    Inkling AI: Mira Murati’s Open-Weight LLM Challenges ChatGPT and Claude

    Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has unveiled its first in-house AI model, Inkling. Designed as a response to the growing demand for open-weight AI systems, Inkling gives developers and enterprises greater control over how large language models are trained, adapted, and deployed.

    Unlike closed AI models such as ChatGPT from OpenAI and Claude from Anthropic, which are primarily accessed through hosted products or APIs, Inkling’s weights are available for outside developers to download, modify, and fine-tune. This open-weight approach allows organizations to customize the model for their specific needs, rather than relying on a controlled environment.

    What is Inkling?

    Inkling is a 975 billion-parameter LLM with a one-million-token context window. However, it does not use all 975 billion parameters for every task. Thanks to its mixture-of-experts architecture, it activates around 41 billion parameters per task, making the model more efficient. The model was trained from scratch on 45 trillion tokens of text, image, audio, and video, and it now generates textual outputs such as code, structured data, and formatted diagrammatic artifacts.

    According to Thinking Machines, Inkling is a generalist model developed for agentic tasks, coding, reasoning, and structured outputs. The company states, “That breadth matters for customization and real-world use: different users need models that can adapt to very different workflows, not just excel on benchmarks.” The model can also be configured to adjust its “thinking effort,” allowing it to trade off speed and reasoning depth depending on the use case.

    How Inkling Differs from ChatGPT and Claude

    The primary distinction is openness and customization. Closed frontier models like ChatGPT and Claude are accessed through controlled apps or APIs and must be used in a controlled environment. Inkling, by contrast, is an open-weight base model that enterprises can tune using Thinking Machines’ model customization platform, Tinker.

    This makes Inkling more attractive for organizations that want to create domain-specific AI systems with their own processes and information. However, it also means companies take on more responsibility for safety, alignment, and security after fine-tuning the model.

    Performance and Limitations

    Thinking Machines notes that Inkling performs well for coding and reasoning but acknowledges that it is “not the strongest overall model available today, open or closed.” Instead, the company positions Inkling as the first step in a broader model family. Alongside the main model, it also showcased Inkling-Small, a lighter model with only 12 billion active parameters that offers lower costs and faster response times.

    Inkling represents a significant shift toward more open and customizable AI, giving enterprises new options for building tailored AI solutions.

  • Two LLMs Team Up to Help Robots Interpret Vague Instructions and Prioritize What Matters

    Two LLMs Team Up to Help Robots Interpret Vague Instructions and Prioritize What Matters

    Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a new method that uses two large language models (LLMs) to help robots understand ambiguous instructions and focus on key details. The approach, called Masked Inverse Reinforcement Learning (Masked IRL), reduces the amount of demonstration data needed to teach a robot by nearly five times, while improving the robot’s ability to infer unspoken user preferences.

    Traditional robot training often requires either extensive physical demonstrations or detailed written instructions. Masked IRL automates the process: first, one LLM clarifies ambiguous prompts (e.g., turning “stay close” into “stay close to the surface of the table”) by comparing a user’s demonstration trajectory to the shortest possible path. Then a second LLM evaluates the environment and “masks” irrelevant details – such as a person leaning on a table – while highlighting critical ones like obstacles to avoid. The robot then uses these prioritized details to generate a safe motion plan.

    In experiments, the system correctly identified unstated user preferences up to 15 percent more often than comparable baselines. Real-world tests showed a robotic arm successfully moving a coffee mug around a laptop, wiping a table while “staying close” to it, and handing a user a bag of chips while “staying away” from both the person and the table – all after fewer than 50 kinesthetic demonstrations.

    The team plans to enhance Masked IRL with camera input, allowing robots to visually focus on relevant objects in dynamic environments. The work was supported by the Tata Group via the MIT Generative AI Impact Consortium Award and the Department of Defense, and will be presented at the 2026 IEEE International Conference on Robotics and Automation.