Kimi K2

Evidence (19)

• 7Hacker News·9/27/2026model_release
  Fireworks Research launches Ember-1

  Fireworks Research introduced Ember-1, a specialized model built on Kimi K3 that aims to deliver similar answer quality with about 40% fewer tokens. The company says it achieved this through training that reduces unnecessary reasoning, and validated the model in benchmarks, live customer A/B tests, and internal workloads.

• 7Ars Technica AI·9/15/2026research
  Open AI models nearly match frontier models at much lower cost

  A Mozilla report says the performance gap between frontier AI models from US companies and top open-weights models from Chinese companies has narrowed to 4.4 months. It also finds that open models are often the better default for routine work, while closed frontier models mainly justify their premium for expert work, long-context tasks, and heavy retrieval.

• 8Hacker News·9/10/2026model_release
  Cognition launches SWE-2 coding model

  Cognition introduced SWE-2, its newest coding model, claiming it improves the cost-performance frontier and reaches 50.0% on FrontierCode 1.1 Main while being significantly cheaper than competing models. The company says SWE-2 scales reinforcement learning to a multi-trillion-parameter regime and is now available in Devin Desktop and CLI, with rollout to Devin Web and Fusion underway.

• 6Hacker News·9/8/2026research
  Kimi K3 runs at 1 token per second on a MacBook Pro via streamed SSDs

  The post demonstrates running the Kimi K3 2.8T model locally on a MacBook Pro at about 1 token per second. It uses four SSDs to stream model data, highlighting an experimental approach to making a very large model usable on consumer hardware.

• 7The New Stack·9/1/2026product_launch
  Agent-built apps make persistence the new hard problem

  The article argues that AI agents that can build, deploy, and maintain full applications create a new infrastructure challenge: keeping state durable while keeping compute ephemeral. Using Moonshot AI’s Kimi as a production example, it says the real issue is not performance but the economics of persistence across millions of mostly idle agent-created apps.

• 6The New Stack·8/26/2026product_launch
  Claude Desktop can now use Qwen, DeepSeek and Kimi models through Ollama

  Ollama has added a new integration that lets Claude Desktop connect to third-party models served through Ollama, including local and cloud-hosted options. The update restores a feature that previously broke after a Claude Desktop change and now allows users to choose models such as Qwen, DeepSeek, Kimi, and GLM inside Anthropic’s app.

• 6Hacker News·8/23/2026product_launch
  Fable shifts AI coding toward cheaper model routing

  The article argues that Fable’s high cost and operational constraints are pushing developers to rethink AI coding workflows, with more emphasis on context engineering, harnesses, and routing tasks to cheaper models like GLM 5.2. It suggests that falling inference prices may not fully restore a single-largest-model workflow because better tooling and model-specific tradeoffs now matter more.

• 7Hacker News·8/23/2026breakthrough
  GLM-5.3 tops real-world LLM benchmark at lower cost

  A Hacker News post highlights a benchmark comparing LLMs on 28 practical tasks, with GLM-5.3 reportedly becoming the first model to clear all five categories at 100%. The write-up says it outperformed Anthropic and OpenAI models on this suite while costing about one-fifth as much, though it had a slower median time-to-first-token than some alternatives.

• 4Hacker News·8/23/2026security
  AI models help root an Amazon Fire HD 10 tablet

  A user reported spending $266 on AI model usage to help find and validate a root exploit for an Amazon Fire HD 10 tablet used as a kiosk. The effort centered on a known Mali GPU kernel vulnerability, CVE-2022-38181, which had been patched by Amazon in later firmware but remained exploitable on the user's version. The write-up describes how multiple AI models assisted with analysis after conventional methods and months of manual investigation had stalled.

• 5Hacker News·8/22/2026research
  NanoGPT Speedrun Frontier Charts AI Agent Progress

  This Hacker News item highlights a benchmark-style chart tracking how different AI agents close the gap toward a human record over time. It compares multiple named models and agent systems by the share of the record gap they have closed across several days of agent runtime. The post appears focused on relative performance and speed in a coding or task-completion speedrun setting rather than a single product launch.

• 7Hacker News·8/13/2026product_launch
  Netlify Adds OpenRouter Models to AI Gateway and Agent Runners

  Netlify has partnered with OpenRouter to let projects use a wider range of models through its AI Gateway, while also expanding the coding models available in Agent Runners. The company says it tested identical prompts across multiple models to compare output quality, functionality, and credit efficiency for common site-building tasks. The report highlights newer open models such as Kimi K3, GLM 5.2, and DeepSeek V4, along with Netlify’s own AXIS evaluation tooling.

• 8Ars Technica AI·8/7/2026breakthrough
  ByteDance trains a giant AI model to challenge top US labs

  ByteDance is reportedly training an AI model with up to 10 trillion parameters, potentially putting it in the same class as Anthropic’s most advanced systems. The model is still in pre-training, and the final size may change before any fine-tuning or release.

• 7Hacker News·8/2/2026research
  Running Kimi K3 on AMD MI355X with Better Performance per Dollar Than B300

  The post reports benchmark results for serving the Kimi K3 model on AMD MI355X GPUs, claiming better performance per dollar than Nvidia B300 in this setup. It also describes a ROCm-side bug in speculative decoding infrastructure and notes that a small PyTorch fix was needed to stabilize the scheduler.

• 2Simon Willison·8/2/2026open_source
  Simon Willison publishes July 2026 newsletter preview

  Simon Willison posted an update announcing his sponsors-only monthly newsletter, with a preview of the June edition and a list of topics covered. The note highlights recent discussions of model behavior, several new model releases, and renewed interest in MCP, but it is primarily a newsletter announcement rather than a major product or research release.

• 4Simon Willison·7/31/2026research
  Simon Willison Discusses Open-Weight AI Models

  Simon Willison joined Bryan Cantrill and Adam Leventhal on the Oxide and Friends podcast to discuss a week of major AI news, centered on open-weight models competing with proprietary frontier systems. The conversation also covered AI security incidents, open-weight policy letters, and several broader digressions, with a note that some topics were already outdated by the time of publication.

• 5Hacker News·7/31/2026open_source
  Run Kimi K3 with 29 GB of RAM at 0.50 tokens per second

  A GitHub project shows a way to run the Kimi K3 model locally using about 29 GB of RAM, albeit at a slow 0.50 tokens per second. The Hacker News discussion focuses on the practicality of running a large model on modest hardware and the tradeoff between memory usage and inference speed.

• 7Hacker News·7/31/2026research
  Moonshot's Kimi reportedly runs on a 20,000-GPU Alibaba cluster

  A Hacker News discussion cites a report that Moonshot's Kimi is using a large Nvidia GPU cluster hosted by Alibaba. The post highlights the scale of the compute infrastructure behind the model, which suggests significant training or inference capacity.

• 5Hacker News·7/29/2026model_release
  Kimi K3 Gets a 256K Context Version

  Kimi Code has introduced k3-256k, a 256K-context variant of its Kimi K3 model. The update keeps the same results within the 256K window while using less quota than the 1M-context version, with guidance on how to switch models safely in coding tools.

• 6Hacker News·7/29/2026research
  Self-hosting Kimi K3 adds cost but improves task success

  A benchmark-style comparison tested self-hosted GPUs, rented hardware, and commercial APIs on 64 real coding tasks. The article finds that self-hosting can cost about 20% more than renting while achieving about 20% better task resolution in some setups, but matching top frontier-model performance requires very large GPU nodes.