--- title: "Kimi-K3 Technical Report [pdf]" url: https://daily.dev/posts/kimi-k3-technical-report-pdf--fcozqb5y4 source_url: https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf type: article source: "Hacker News" published: 2026-07-27T15:53:29.399Z updated: 2026-07-29T19:56:30.803Z tags: ["ai", "llm"] reading_time: 1 upvotes: 2 comments: 0 language: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Kimi-K3 Technical Report [pdf] **[Hacker News](https://daily.dev/sources/hn)** · 1 min read · 2 upvotes · 0 comments ## Summary MoonshotAI has released the technical report for Kimi-K3, their latest frontier AI model, hosted on GitHub. The repository is titled 'Open Frontier Intelligence' and contains the K3 technical report PDF detailing the model's architecture and capabilities. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Community take How the wider developer community reacted, aggregated from 1 discussion and 108 comments across hackernews (as of 2026-07-29). **TL;DR:** The community is broadly excited about Kimi K3's open-weight release and its implications for the AI ecosystem, but there is significant debate around the license restrictions (not truly OSI-compliant), whether open models are 'decel' or 'accel', and the practical economics of self-hosting a frontier-scale model. **Sentiment:** 45% positive · 35% mixed · 20% skeptical **The case for** - Chinese labs like Moonshot/DeepSeek have demonstrably contributed major research advances (MoE, reasoning models) that even closed Western labs have copied, undermining the 'fast-follower only' narrative. - The economics of self-hosting on a GB300 rack work out to under $0.60/million output tokens for a frontier model, making on-prem deployment viable for large enterprises. - Open weights cap inference margins and let every lab worldwide start from the frontier rather than spending billions to catch up independently. - Accompanying open-source infra repos (MoonEP, AgentEnv, FlashKDA) compound the value beyond just the weights. **The pushback** - The license is not OSI-compliant ('modified MIT') — it restricts commercial use above $20M revenue or 100M MAU, making it 'downloadable' rather than truly open source. - Training data provenance is opaque and likely includes copyrighted material without licensing, a cost that is often glossed over. - The multi-teacher distillation methodology raises questions about whether logprobs from closed frontier models (e.g., OpenAI, Anthropic) were used, which may have licensing implications. - Self-hosting at this scale still requires dedicated ops staff and significant capital, which is non-trivial for most organizations. **By community** - hackernews (mixed): Enthusiastic about the release and its open-science implications, but heavily debating the license's OSI compliance, the decel/accel framing, and the practical realities of self-hosting a 1.5TB frontier model. **Hottest debate:** Whether open-weight models accelerate or decelerate AI progress — specifically whether they undercut closed-lab funding/research or democratize the frontier for everyone. **Open questions** - Is a torrent available for the 1.5TB weights download? - What were the pre-training token counts and total FLOPs consumed? - Did the multi-teacher distillation use logprobs from closed frontier models like GPT or Claude, and is that permitted under those APIs? - Will Moonshot actually enforce the commercial license clauses, given past non-enforcement precedents? **Highlights** > Back of the envelope calculation (could be off, correct me if I am) If you are a large enough company that spends million+ on inference a month, it makes sense to buy a GB300 rack ($6M on top range from what I could find) which has 20.7 TB. Since the model is mixed trained (MXFP4), you would need less than 10% of the rack's memory to serve the full model. Aggregate HBM bandwidth: 576 TB/s. You can run over 6000 parallel agentic workflows (each with ~100k context on average) at ~30 tok/s. Assuming the annual amortization+electricity at $1.5M/year and about 50% average annual utilization, you get less than 60 cents (USD) per million output token, for a frontier model with plenty of capacity to share, all your data never leaving premises and well over an order of magnitude cheaper! As long as a company believes that the openweight models will continue to get more capable and 'AI is here to stay', this model provides the first solid footing for a decision to just buy a rack. > — [GodelNumbering on hackernews · 2 comments](https://news.ycombinator.com/item?id=49072173) > License: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE > If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose. + the existing 100 million monthly active users, or more than 20 million US dollars for commercial products clause > — [fahrradflucht on hackernews · 5 comments](https://news.ycombinator.com/item?id=49071313) > only if you only get your news from main stream business press and Big Lab propaganda channels There's no chance K3 is a distill of Fable, it came out way too soon after the limited fable release to be feasbile. If you look at all of the top ML conferences, chinese labs contribute way more to advances in ML than "Open"AI and Anthropic: https://www.reddit.com/r/TheMachineGod/comments/1pi4q7f/pape... This K3 release just helped every other lab on the planet stay in the race by making it possible for them to build on top of it, placing them at the frontier starting line instead of having to spend billions of their own dollars and risking it all to attempt to catch up. The open source contributions I linked to above will move the whole field forward and reduce the costs of training and inference for everyone. Open science compounds on it self, every new advancement pushes the field forwards and opens up new grounds for future improvements. > — [m\_ke on hackernews · 1 comments](https://news.ycombinator.com/item?id=49072450) > It's a gradient almost, with some steps. So far, I think you could categorize every single released so far as one of: - Proprietary - No access beyond remote endpoints - Downloadable - You can run it, but there are restrictions and training data/code isn't public and/or under FOSS license, nor are the weights under a FOSS license - Open weights - The weights are under a FOSS license and downloadable without restrictions, but not all of training code/data is public or under a FOSS license - Open source - The model architecture, weights, training code and data are all under a FOSS license, and you can freely download and use them without any restrictions Few models hit the mark for "open source model", many models people call "open weights" would go under "downloadable" here, which personally I think would be accurate. > — [embedding-shape on hackernews](https://news.ycombinator.com/item?id=49072787) > I think it's pretty hard to hold that worldview: Anthropic couldn't ship a reasoning model until they copied DeepSeek R1's homework, and they've all copied DS-style super-sparse MoEs at this point too. > — [reissbaker on hackernews · 1 comments](https://news.ycombinator.com/item?id=49072492) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49070985) · 171 points · 108 comments ## Similar posts on daily.dev - [Kimi-K2.5/tech\_report.pdf at master · MoonshotAI/Kimi-K2.5](https://daily.dev/posts/kimi-k2-5-tech-report-pdf-at-master-moonshotai-kimi-k2-5-cdol9l0nd) · Hacker News · 2 upvotes · 1 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/kimi-k3-technical-report-pdf--fcozqb5y4)