Kimi K3
Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight Mixture-of-Experts model with native vision and a 1-million-token context window, built on Kimi Delta Attention and Attention Residuals. On the API it always reasons, with configurable reasoning effort.
Released
July 16, 2026
Type
multimodal
License
open-weight
Kimi K3 License
Context
1,048,576 tokens
Pricing
Input
$3.00 / 1M tokens
Output
$15.00 / 1M tokens
List price on the developer's own pages, checked . Input = standard (cache-miss) input. Cached input $0.30; cache write $3.00 (TTL 5min) or $6.00 (TTL 1h). Prices exclude taxes.
Capabilities
Architecture
Parameters: 2.8T total, 104B activated (MoE; 16 of 896 experts per token)
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Kimi research index (dates) (opens in a new tab)
- Kimi K3 tech blog (opens in a new tab)
- Kimi API: Model inference pricing (opens in a new tab)
- Kimi API: Platform changelog (opens in a new tab)
- Hugging Face: moonshotai/Kimi-K3 model card (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Kimi K3 and Moonshot AI
MarkTechPost
Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB
Z.ai has released GLM-Legal, a legal AI model designed for tasks such as contract review, evidence analysis, and compliance research. The model is built on Z.ai's GLM-5.3-Flash architecture and has achieved a 19.2% all-pass rate on Harvey’s Legal Agent Benchmark (LAB). This rate is nearly three times higher than its base model's 6.7% pass rate. GLM-Legal is a domain-specialized version of Z.ai's GLM-5.3-Flash model, which is natively multimodal and can process text, images, video, and files, outputting text. The base model features 320 billion total parameters with 18 billion active parameters, utilizing a mix of sparse and linear attention to reduce serving costs for long contexts. The model underwent a two-stage training process. The first stage involved mid-training on expanded legal corpora to enhance reasoning from facts to conclusions. The second stage included post-training with reinforcement learning across tens of thousands of legal task environments, with the agent harness co-evolving with the model. Performance metrics indicate GLM-Legal achieved a 19.17% Harvey LAB task pass rate, ranking seventh out of 79 models on Vals AI, at a cost of $0.15 per test. While its performance on Vals AI’s Legal Research Bench was 41.83%, below its base model's 45.19%, Z.ai states GLM-Legal offers frontier-class legal agent scores at approximately 1/128 the cost of Claude Fable 5. Access to GLM-Legal is currently limited, with its API available only through application. Public weights are not provided, and the model runs exclusively on Z.ai’s hosted API.
MarkTechPost
Meet Together Link: A Free CLI That Runs Open Models Like Kimi K3 and GLM 5.3 Inside Claude Code, Codex, and OpenCode
Together AI has released Together Link, a free, MIT-licensed Command Line Interface (CLI) tool that is currently in beta. This new tool enables developers to connect their existing coding agents with open-source AI models that are hosted on the Together AI platform. The supported coding agents include Claude Code, Claude Desktop, Codex, ChatGPT Desktop, OpenCode, and Pi. The core concept behind Together Link is to allow users to retain their familiar coding harnesses and workflows while switching to more cost-effective open models, thereby reducing expenses. The tool is immediately deployable, with a straightforward installation process requiring a single command on macOS or Linux, and only necessitates a Together API key for operation. As it is in beta, Together AI has indicated that commands, routing mechanisms, and the list of supported models may be subject to change. Together Link addresses a common inefficiency in development workflows where coding agents are frequently directed to use the same premium, closed-source model for all tasks, regardless of complexity. This can range from minor code corrections to extensive code rewrites. In its launch announcement, the Together AI team highlighted that engineering organizations can incur monthly costs ranging from tens of thousands to millions of dollars for using closed-source models. Together AI argues that open-source models have significantly narrowed the performance gap with their proprietary counterparts. For demanding coding tasks, models like Kimi K3 and GLM 5.3 are recommended. For more routine tasks, GLM 5.3 Flash and DeepSeek V4.1 Flash are suggested. The installation of Together Link is designed to be simple, requiring users to execute a single command to download and set up the tool. The installer automatically includes Bun if it is not already present on the system and places the necessary commands in the ~/.local/bin directory. Users can then launch a graphical interface by running the 'togetherlink' command or initiate specific tools directly, such as 'togetherlink claude', 'togetherlink codex', 'togetherlink opencode', or 'togetherlink pi'. Shorter aliases, like 'tclaude', are also functional. According to the official documentation, Together Link does not require any local proxy or daemon to be running. Instead, each coding agent communicates directly with Together AI's hosted gateway. Terminal agents are provided with temporary per-launch configurations that are automatically removed once the session concludes. For applications like Claude Desktop and ChatGPT Desktop, separate, reversible profiles are utilized, with commands like 'togetherlink chatgpt off' available to revert these changes. The default configuration for Together Link includes a virtual auto-router for sessions. This auto-router analyzes the initial task of each session to determine the most appropriate model. The stated objective is to direct quick fixes to fast and low-cost models, while more complex or challenging problems are routed to models with advanced capabilities. The auto-router's functionality is further enhanced by API key integration. For instance, with an Anthropic API key, the router can dynamically switch between Anthropic's Opus 5.5 model and GLM 5.3. In scenarios where an Anthropic API key is not provided, the routing will proceed without this specific integration, potentially utilizing other available models based on the session's requirements and the user's configuration.
MarkTechPost
Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads
Reflection AI has introduced Beam, its inaugural open-weight model, on March 18, 2024. Beam is a sparse Mixture-of-Experts (MoE) architecture boasting a total of 501 billion parameters, with 23 billion active parameters per token. This model is specifically engineered for coding, reasoning, and agentic workloads. According to the Reflection AI team, Beam is positioned to directly compete with larger open models such as GLM 5.2, while demonstrating a significant advantage in inference compute efficiency, using 3 to 4 times less compute on reasoning benchmarks. Currently, Beam is not available for self-hosting but is undergoing final red-teaming. Early access is being managed through a waitlist on the Reflection platform. Beam functions as a general agent model, developed from the ground up by Reflection AI, with a primary focus on enterprise-level coding and agentic applications. Reflection AI frames Beam as a significant advancement in the Western open-weight model landscape. The research team acknowledges that while Kimi K3 may still lead in raw capability, Beam's key selling proposition is its efficiency during inference. Users can adjust a "reasoning effort" parameter, where lower settings yield concise answers and higher settings enable more extensive reasoning for complex tasks. This allows teams to tailor the model's effort to the specific difficulty of a task and their available compute budget. The pretraining of Beam involved an extensive dataset of 23.8 trillion tokens, sourced from the web, public repositories, and proprietary licensed datasets. The Reflection AI team reported that their curation process removed approximately 95% of raw internet tokens, while retaining about 1.8 trillion high-quality tokens that might have been discarded by conventional filtering methods. The model's architecture integrates local and global attention mechanisms with finely routed experts. Its load balancing strategy builds upon DeepSeek-V3's auxiliary-loss-free balancing, incorporating a cosine decay for expert-bias updates. This approach ensured that the most utilized expert experienced a load of only 1.04 times the average by the conclusion of pretraining. Throughout the 52 layers, residual norms were maintained within bounds through depth-based scaling, SandwichNorm, attention gating, and FP32 residual accumulation. The entire pretraining process was completed in under four weeks, utilizing 6,144 NVIDIA GB300 NVL72 GPUs. The reported goodput reached 92.3% near the end of the training period, with nine semi-automatic rewinds implemented. During mid-training, the effective context window was extended to 1 million tokens.
The Hacker News
OpenAI Disrupts Reasoning Extraction Campaign Linked to Moonshot AI Associates
OpenAI announced on Wednesday that it has successfully identified and disrupted a coordinated campaign aimed at illicitly extracting protected reasoning from its artificial intelligence (AI) models. The company stated that a "core cluster of the activity," which began in the first week of July, has been attributed to individuals associated with Moonshot AI, a Chinese AI company headquartered in Beijing. This campaign represents a significant effort to bypass security measures and gain unauthorized access to the proprietary reasoning processes embedded within OpenAI's advanced AI systems. The disruption of this campaign highlights the ongoing challenges in safeguarding AI model integrity and intellectual property. AI models, particularly large language models (LLMs) like those developed by OpenAI, are trained on vast datasets and involve complex architectures and reasoning capabilities that represent substantial research and development investments. The extraction of this "protected reasoning" could allow competitors or malicious actors to replicate or reverse-engineer these capabilities without undertaking the extensive resources required for independent development. This practice, often referred to as "model distillation" or "model extraction," can lead to the creation of unauthorized or counterfeit AI models that may lack the safety and ethical guardrails implemented by the original developers. OpenAI has not provided specific details regarding the technical methods used in the extraction campaign or the exact nature of the "protected reasoning" being targeted. However, the attribution to Moonshot AI suggests a potential international dimension to these efforts. Moonshot AI, as a Beijing-based entity, operates within a rapidly evolving and competitive global AI landscape. The company's involvement, as alleged by OpenAI, points to the strategic importance of AI model capabilities and the lengths to which some entities might go to acquire them. OpenAI's proactive disruption of this campaign underscores its commitment to protecting its intellectual property and maintaining the security and integrity of its AI technologies. The company's statement implies that further details may be forthcoming as its investigation progresses, but the immediate focus has been on neutralizing the threat and informing the public about the nature of the activity. This incident also raises broader questions about AI security and the governance of AI development globally. As AI capabilities become increasingly sophisticated and integrated into various sectors, the protection of the underlying models and their reasoning processes becomes paramount. The ability to extract and replicate these capabilities could have significant implications for market competition, national security, and the responsible deployment of AI. OpenAI's action serves as a warning to other AI developers and a signal to the broader AI community about the persistent threats to AI model security. The company's efforts to combat such campaigns are crucial for fostering trust and ensuring the continued innovation and ethical development of artificial intelligence technologies.
MarkTechPost
Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens
Fireworks AI has released Ember-1, a specialized model developed by Fireworks Research through post-training Moonshot AI's open-weight Kimi K3. This new model is designed to produce shorter reasoning traces while preserving task accuracy, a feat achieved by fundamentally altering the model's reasoning process rather than simply adjusting inference settings. According to a release post from Fireworks, Ember-1 delivers the quality of Kimi K3 with approximately 40% fewer tokens. Currently, Ember-1 is available as a Research Preview exclusively through the Fireworks serverless API. Fireworks has not released the model's weights, training code, or specific training algorithms, making self-hosting impossible at this time. The development of Ember-1 addresses a significant challenge with current reasoning models, where they can expend over 90% of generated tokens on internal reasoning processes. This inefficiency becomes particularly costly in multi-turn agentic workloads, where prior reasoning is replayed to the model in each subsequent turn. The context window grows roughly quadratically with the number of turns, leading to long reasoning traces from early turns being re-read and re-billed on every later call. Fireworks stated that customers desired Kimi K3's coding capabilities at a reduced cost, but lowering K3's reasoning effort setting resulted in an unacceptable loss of quality. Consequently, the Fireworks team focused on training the model to reason more efficiently. Fireworks Research constructed Ember-1 by retaining useful self-reflection behaviors, such as revisiting assumptions or responding to feedback, while eliminating redundant reasoning and unproductive loops. The training dataset encompasses a broad range of tasks, including mathematics, coding, instruction following, conversation, search, tool use, and software engineering. This dataset includes both standalone problems and extended multi-step interactions. The training process utilized task and environment feedback to guide on-policy planning and learning. The Fireworks team conducted over 50 training experiments and more than 200 evaluations. They also developed new, unpublished training algorithms. All training operations were performed on Fireworks' own infrastructure. This innovation by Fireworks AI aims to make advanced reasoning models more cost-effective for applications that require extensive sequential processing or agentic behavior. By reducing the token overhead associated with internal thought processes, Ember-1 could enable more complex and longer-running AI agent tasks without a proportional increase in computational cost. The focus on maintaining accuracy while reducing token count is a critical step towards more efficient and scalable AI deployments, particularly in competitive cloud API markets where cost per token is a key metric for adoption. The model's availability as a research preview suggests that Fireworks AI is seeking feedback and further refinement before a broader commercial release.