Kimi K3 Drops: Moonshot's 2.8T Parameter Frontier Model Takes on GPT-5.6 and Fable 5
TL;DR
Moonshot AI releases Kimi K3 with 2.8 trillion parameters, 1M context window, and Delta Attention architecture.
Moonshot AI dropped Kimi K3 today, their largest model to date at 2.8 trillion parameters. The release marks a significant acceleration in the company's iteration cycle - just three months after open-sourcing K2.6 - and positions K3 as a direct competitor to frontier models from OpenAI and Anthropic.
What is Kimi K3?#
K3 is Moonshot AI's new flagship reasoning model, built for "agentic coding and knowledge work" according to their documentation. The key specs:
- Parameters: 2.8 trillion (up from ~1T in previous versions)
- Context window: 1 million tokens
- Architecture: Delta Attention, a hybrid linear attention mechanism with Attention Residuals
- Vision: Native multimodal support for images and video
- Thinking mode: Always-on reasoning with configurable effort levels
The model introduces automatic context caching with no manual configuration required, structured JSON output support, and tool integration capabilities including custom tools and dynamic tool loading.
Pricing Reality Check#
K3 comes in at $3 per million input tokens and $15 per million output tokens, with cache hits at $0.30 per million. This is aggressive frontier pricing - roughly matching Anthropic's Sonnet series and sitting just above GPT-5.6 Terra's input rate ($2.50 per million).
But here's where HN commenters raised valid concerns about the real-world economics:
Reasoning efficiency matters more than per-token pricing. As one commenter put it: "If Sol spends 10K reasoning tokens to do something (at $30/1M) vs Kimi K3 that spends 50K reasoning tokens, Sol would win on cost effectiveness."
OpenAI's models are known for reasoning efficiency, and some Claude models like Fable at lower effort settings match that efficiency. K3's actual cost-per-task remains to be seen as independent benchmarks come in.
Tokenizer differences compound pricing gaps. Anthropic's tokenizers encode the same text at higher token counts than OpenAI's. Kimi's tokenization efficiency will affect real-world cost comparisons.
The subscription angle. Moonshot offers subscriptions up to $199/month. Some HN commenters noted that current monthly coding plans from Anthropic and OpenAI often beat pay-per-token pricing for daily coding work unless you're extremely light on usage.
What HN is Saying#
The Hacker News discussion is running hot with 187+ comments at time of writing. The main threads:
Skepticism about positioning claims: Moonshot's quickstart claims K3 "ranks second only to Claude Fable 5 and GPT-5.6 Sol" in overall intelligence. Several commenters pointed out that ranking second to two models means you're third, and that the tech blog hasn't been updated since K2.6 - two releases ago.
DeepSeek comparison: Multiple commenters are comparing K3 unfavorably to DeepSeek V4 on price. DeepSeek's cache pricing sits around $0.003 per million tokens - roughly 100x cheaper than K3. One developer noted: "I've been using it whenever possible as even longer agent sessions cost few cents."
Open weights question: The original quickstart mentioned model weights would be released "in the coming days." That paragraph has since been removed from the documentation, raising questions about whether K3 will actually be open weights like its predecessors.
The AI fatigue contingent: A popular comment requested HN add an "AI filter" button. The replies spawned multiple links to existing filter tools, including Simon Willison's filtered HN and a third-party service that uses AI to ironically filter out AI posts.
Where K3 Actually Fits#
Based on available information, here's a practical read on K3's position:
Best case: K3 delivers on the frontier intelligence claims and the reasoning efficiency is competitive with GPT/Claude models. At that point, it's a genuine third option for production agentic workloads with strong vision capabilities and massive context.
Realistic case: K3 is a capable model that trades blows with Sonnet/GPT-5.5 class models on most tasks, with potentially worse reasoning efficiency that inflates real-world costs. The 1M context window is genuinely useful for large codebases and long documents.
Open weights wildcard: If Moonshot releases the weights as initially suggested, K3 becomes interesting for self-hosted inference despite the model size. A 2.8T MoE model isn't running on consumer hardware, but it's deployable for organizations with the GPU budget.
DeepSeek Shadow#
Several HN commenters mentioned that DeepSeek is expected to release a new model this week. If DeepSeek V5 drops with their characteristic aggressive pricing, K3's launch window gets more crowded.
The broader dynamic at play: Chinese open-weight models are pushing pricing pressure while US labs maintain premium pricing on frontier capabilities. K3 is positioned somewhere in between - frontier ambitions with open-weight origins, but priced like a US frontier model.
What Developers Should Do#
Wait for independent benchmarks. The initial claims are marketing. The real signal comes from lmsys arena rankings, Aider polyglot benchmarks, and production feedback over the next few weeks.
Test the vision capabilities. Native multimodal with video support is still relatively rare. If you're building agents that need to process visual context, K3's vision offering is worth evaluating against GPT-5.6 Vision and Claude's multimodal capabilities.
Watch the open weights situation. If K3 weights do get released, that changes the calculus entirely for organizations that can self-host. Check back on their tech blog and GitHub.
Monitor your actual spend. If you're already using an AI coding subscription (Codex, Claude Code, etc.), compare the effective per-task cost against K3's API pricing before switching. The subscription economics often win for heavy daily use.
K3 is a serious frontier model attempt from a well-funded lab. Whether it justifies the frontier pricing depends on factors we won't know until the community has a few weeks with it.
Continue Reading#
- Cloudflare Runs Kimi and GLM at Scale: FP8 KV Caches, INT4 Weights, and a Cache Safety Net
- GLM 5.2 and the AI Margin Collapse Thesis
- GPT-5.5 Has a 3x Higher Hallucination Rate Than MIT-Licensed GLM-5.2
- Kimi Linear: An Attention Architecture That Outperforms Full Attention
Sources#
Get the next deep dive like this in your inbox
One email a week on News and the rest of the AI dev stack. Free.
Read next on local and open-weight models
Kimi K3 Weights Land on HuggingFace: 2.8T Open Frontier Model You Can Actually Download
Moonshot AI released the full Kimi K3 weights on HuggingFace today - 2.8T parameters, 1M context, native MXFP4 quantization, ~1.63TB download. The HN community reaction, what the license really says, and why this matters for the open-weights AI market.
7 min readGLM 5.2 and the AI Margin Collapse Thesis
Martin Alderson's argument for why open-weights models like GLM 5.2 will compress frontier lab margins is sparking debate on HN. Here is what the thesis actually says, where HN agrees and disagrees, and why it matters for developers choosing models.
7 min readKimi K3 in 10 Minutes: Moonshot AI's 2.8T Open Model, API Setup, Pricing, and Benchmarks
Kimi K3 is the first open-source 3T-class model with a 1M-token context window, native vision, and OpenAI-compatible API. Here is what it does, how to call it, what it costs, and how it benchmarks against Fable 5 and GPT-5.6 Sol.
6 min readNew here? Start with
Technical content at the intersection of AI and development. Building with AI agents, Claude Code, and modern dev tools - then showing you exactly how it works.
