Open-Source LLM Digest — tracking Kimi, DeepSeek, GLM and peers
New · Kimi K3.1 Pro released
K3.1 Pro is the current flagship: 1,048,576-token context, reference pricing $2.00 input / $10.00 output per million tokens (platform-dependent).
Kimi K3 released
K3 (July 2026): 2.8 trillion parameters, natively multimodal, million-token context — for long coding sessions, knowledge work and deep reasoning.
GLM-5.3 · MIT license
A flagship MoE model from Zhipu: million-token context and open weights under the MIT license, plus the multimodal Flash / FlashX series.
DeepSeek V4.1 Flash
The fast variant of the DeepSeek V4.1 family: fastest on Artificial Analysis' output-speed leaderboard, with an intelligence index of 46.
The Digest aggregates open-model news and independent benchmarks: from million-token context to coding models, with clear comparisons — or jump straight to TuneRouter to try them.
Kimi highlights
From massive documents to multi-step agent tasks: "complex problems, handled with ease".
Ultra-long context
The K3 series supports a 1,048,576-token context window — an entire book or codebase in one pass.
Deep reasoning
Built for long-horizon tasks and deep reasoning, with strong multi-step planning and problem decomposition.
Natively multimodal
K3's architecture is multimodal end to end: understanding and generating text and images through a unified backbone, no bolt-on modules.
Coding & code
The K2.7 Code model for programming tasks: generating, analyzing and refactoring code, plus long coding sessions.
Agentic AI
Web browsing, document processing, slide and spreadsheet creation, and multi-agent collaboration to carry tasks end to end.
Deep research
Deep Research mode searches, verifies and summarizes automatically into sourced research reports.
Benchmark comparison
Data from the independent evaluation agency Artificial Analysis (October 2026): Kimi K3.1 Pro scored an intelligence index of 55, ranking 4th out of 215 models.
Test case from the official blog: on GPU kernel optimization (standardized sandbox, 24-hour limit, four tasks across NVIDIA Hopper and GPGPU processors), Kimi K3 performed on par with Claude Fable 5 and clearly outperformed Opus 4.8, GPT-5.6 Sol and GPT-5.5; it also completed a chip design verification in a self-directed 48-hour run.
Open models available
Information compiled from the public pages of the Kimi platform (October 2026); the official source is authoritative.
| Model | Positioning | Context | Reference pricing (per million tokens) |
|---|---|---|---|
| kimi-k3.1-proNewest | Current flagship for complex tasks | 1,048,576 | Input $2.00 / output $10.00 |
| kimi-k3 | Flagship: 2.8T parameters, natively multimodal | Platform-dependent | Platform-dependent |
| kimi-k2.7-code | Coding specialist | Platform-dependent | Platform-dependent |
| kimi-k2.6 | April 2026 flagship, complex problem solving | Platform-dependent | Platform-dependent |
| deepseek-v4.1-flash | DeepSeek's fast variant | Per official source | Per official source |
| glm-5.3MIT | Open-weight flagship MoE model | 1M | Per official source |
FAQ
Who develops Kimi?
Kimi is developed by Moonshot AI, with the official site at kimi.com and the developer platform at platform.moonshot.ai.
How can I try Kimi?
Via kimi.com or the app directly; developers can call the API through the platform. The button at the top of this page takes you to TuneRouter to learn how to get access.
Are Kimi models open source?
The K2 series was released with open weights; open-source policy varies between generations — Moonshot AI's official announcements are authoritative.
What is this site's relationship to the official provider?
This is an independent third-party website with no affiliation, agency, partnership or authorization from Moonshot AI. Its content is for reference only; what the official provider publishes is authoritative.
Leave a message
Questions, feedback, or tips for the editors — drop us a note and we will take a look.
Want to try the flagship model family?
Head to TuneRouter to learn how to access and call the models.
Go to TuneRouter →