What we are adding: K3 is now available via the Telnyx Inference API, hosted on GPUs that we own and operate.
This matters due to the size of this model. A 2.8T model needs dedicated infra to serve well. Because we own and operate the GPUs, we can contorl throughput. with no inter-provider hops and no cloud tenant, latency is minimized.\nWe have GPUs in each of the US, EU, APAC, and MENA. Inference runs in the region you pick, with zero data retention. We do not store prompts or completions after the response returns.\nBecause we own the infra, the per-token price reflects the cost of running the model, not the cost of renting someone else's plus their margin.
We have not benchmarked K3 ourselves yet. Moonshot's own numbers and early third-party evaluations put it at frontier level for coding and agentic work, trailing only Claude Fable 5 and GPT 5.6 Sol on aggregate. Full breakdown in their blog.
Pricing on Telnyx: $2.70/1M input tokens, $13.50/1M output tokens, $0.27/1M cached input tokens. Prompt caching enabled by default. Served via an OpenAI-compatible endpoint so you can test easily.
Model ID: [MODEL_ID] API: https://api.telnyx.com/v2/ai/chat/completions Docs: https://developers.telnyx.com/docs/inference Technical blog (Moonshot): https://www.kimi.com/blog/kimi-k3"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi K3 Now Available via Telnyx Inference API"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://telnyx.com/release-notes/kimi-k3-telnyx-inference"}},"_tags":["story","author_fionaattelnyx","story_49076505"],"author":"fionaattelnyx","children":[49076506,49076546,49076656,49078033,49078159,49078391,49078615,49078703,49079557,49080071,49080336,49080461,49080508,49080563,49080676,49080698,49080779,49082020,49082170,49082339,49082363,49082487,49082929,49105217],"created_at":"2026-07-27T22:46:28Z","created_at_i":1785192388,"num_comments":88,"objectID":"49076505","points":132,"story_id":49076505,"story_text":"Moonshot AI released open weights for Kimi K3 today and it's live on Telnyx Inference. The architecture and Moonshot's own benchmarks are in their technical blog. This post is about running it on Telnyx.
What we are adding: K3 is now available via the Telnyx Inference API, hosted on GPUs that we own and operate.
This matters due to the size of this model. A 2.8T model needs dedicated infra to serve well. Because we own and operate the GPUs, we can contorl throughput. with no inter-provider hops and no cloud tenant, latency is minimized.\nWe have GPUs in each of the US, EU, APAC, and MENA. Inference runs in the region you pick, with zero data retention. We do not store prompts or completions after the response returns.\nBecause we own the infra, the per-token price reflects the cost of running the model, not the cost of renting someone else's plus their margin.
We have not benchmarked K3 ourselves yet. Moonshot's own numbers and early third-party evaluations put it at frontier level for coding and agentic work, trailing only Claude Fable 5 and GPT 5.6 Sol on aggregate. Full breakdown in their blog.
Pricing on Telnyx: $2.70/1M input tokens, $13.50/1M output tokens, $0.27/1M cached input tokens. Prompt caching enabled by default. Served via an OpenAI-compatible endpoint so you can test easily.
Model ID: [MODEL_ID] API: https://api.telnyx.com/v2/ai/chat/completions Docs: https://developers.telnyx.com/docs/inference Technical blog (Moonshot): https://www.kimi.com/blog/kimi-k3","title":"Kimi K3 Now Available via Telnyx Inference API","updated_at":"2026-09-03T21:41:33Z","url":"https://telnyx.com/release-notes/kimi-k3-telnyx-inference"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"noumenon1111"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"Intel has spent the last decade chasing moonshots: AI accelerators, Optane, RealSense, Thunderbolt branding. All while slowly bleeding out cultural relevance. Meanwhile, gamers, devs and power users are quietly revolting against walled gardens and bloated software. Windows is being shoehorned into an "agentic OS," consoles are doubling down on subscription models, and users are mourning the loss of simplicity and control.
When was the last time you were unpleasantly surprised by a big, required update when you try to play a game?
Intel has a once-in-a-decade chance to blow the doors off the console wars and make Linux mainstream. But it requires something Intel hasn't shown in years: restraint and aggression.
Here's the play:
1. Build the box: Take Panther Lake. Stick it in a boring little box that fits under the TV. No RGB. No gimmicks. Just raw performance in a console-size form factor.
2. Ship SteamOS: Boot it straight into Big Picture mode. No custom UI. No Intel launcher. No "ecosystem." The heavy lifting is already done: Steam, Proton, Vulkan. Valve did the hard work already, Intel just needs to leverage it.
3. Sell at cost: Price it like a weapon. Undercut consoles. Forget margins; this is about relevance. Get it on the shelf at Best Buy. Intel Arc becomes a household name overnight. Gamers stop asking "Can it run Crysis?" and start asking "Does it run on Arc?"
4. Walk away: No subscriptions. No proprietary APIs. No marketing fluff or Intel trying to be relevant with cheeky references to the past. Don't try to own the platform, and DON'T ALLOW FEATURE CREEP!! Just set the grenade, pull the pin, and let the market do the rest.
This play works expertly, because gamers are fed up, Valve already proved Linux gaming works, the open ecosystem will catch the giants Sony and Microsoft sleeping. Nintendo will just keep doing what it's always done. Intel Arc gains legitimacy in the one place where driver optimization matters more than CUDA.
IF Intel does this, here's what happens:
1. Sony and Microsoft scramble to defend their turf. Expect rushed subscription perks, price cuts, and more "exclusive content." However, the damage is done. The idea of an open, console-like PC (crucially, WITHOUT PC branding) becomes mainstream.
2. Valve wins big. SteamOS becomes the de facto standard living room OS. Proton development accelerates. Linux gaming stops being a niche and becomes a cultural norm.
3. NVIDIA feels the heat. Intel Arc suddenly matters, and NVIDIA can't ignore a competitor selling hardware at cost. Expect aggressive driver optimizations and maybe even a Linux-first marketing push.
4. Linux adoption is already exploding, more than past years. 2026 might really be the much mythologized year of desktop Linux. But if Intel ships a little Linux box into every other living room in America, OEMs and enterprise takes notice.
This isn't just a play for games. It's a real cultural reset.
However, Intel won't do it. It's not that it can't, as such, but it goes against two of their biggest anti-traits:
1. It requires restraint, which Intel doesn't have. They must NOT build an ecosystem or chase subscriptions.
2. It requires aggression, which Intel doesn't have. They MUST price it at cost and market it like a rebellion.
I really don't want or need any credit at all. But I love Intel. My inner child loves Intel. After working for them for a few years, even though my whole team was laid off last year, I want to see them succeed and return to cultural relevance. If this message reaches ONE person who has convincing power in the right meeting, hey, maybe it will happen."},"title":{"matchLevel":"none","matchedWords":[],"value":"Tell HN: Intel could blow up the Console Wars if it had the guts"}},"_tags":["story","author_noumenon1111","story_46607497","ask_hn"],"author":"noumenon1111","children":[46607961,46608047,46609982,46613306,46616773,46618583,46624718,46732830],"created_at":"2026-01-13T20:33:25Z","created_at_i":1768336405,"num_comments":14,"objectID":"46607497","points":9,"story_id":46607497,"story_text":"Intel has spent the last decade chasing moonshots: AI accelerators, Optane, RealSense, Thunderbolt branding. All while slowly bleeding out cultural relevance. Meanwhile, gamers, devs and power users are quietly revolting against walled gardens and bloated software. Windows is being shoehorned into an "agentic OS," consoles are doubling down on subscription models, and users are mourning the loss of simplicity and control.
When was the last time you were unpleasantly surprised by a big, required update when you try to play a game?
Intel has a once-in-a-decade chance to blow the doors off the console wars and make Linux mainstream. But it requires something Intel hasn't shown in years: restraint and aggression.
Here's the play:
1. Build the box: Take Panther Lake. Stick it in a boring little box that fits under the TV. No RGB. No gimmicks. Just raw performance in a console-size form factor.
2. Ship SteamOS: Boot it straight into Big Picture mode. No custom UI. No Intel launcher. No "ecosystem." The heavy lifting is already done: Steam, Proton, Vulkan. Valve did the hard work already, Intel just needs to leverage it.
3. Sell at cost: Price it like a weapon. Undercut consoles. Forget margins; this is about relevance. Get it on the shelf at Best Buy. Intel Arc becomes a household name overnight. Gamers stop asking "Can it run Crysis?" and start asking "Does it run on Arc?"
4. Walk away: No subscriptions. No proprietary APIs. No marketing fluff or Intel trying to be relevant with cheeky references to the past. Don't try to own the platform, and DON'T ALLOW FEATURE CREEP!! Just set the grenade, pull the pin, and let the market do the rest.
This play works expertly, because gamers are fed up, Valve already proved Linux gaming works, the open ecosystem will catch the giants Sony and Microsoft sleeping. Nintendo will just keep doing what it's always done. Intel Arc gains legitimacy in the one place where driver optimization matters more than CUDA.
IF Intel does this, here's what happens:
1. Sony and Microsoft scramble to defend their turf. Expect rushed subscription perks, price cuts, and more "exclusive content." However, the damage is done. The idea of an open, console-like PC (crucially, WITHOUT PC branding) becomes mainstream.
2. Valve wins big. SteamOS becomes the de facto standard living room OS. Proton development accelerates. Linux gaming stops being a niche and becomes a cultural norm.
3. NVIDIA feels the heat. Intel Arc suddenly matters, and NVIDIA can't ignore a competitor selling hardware at cost. Expect aggressive driver optimizations and maybe even a Linux-first marketing push.
4. Linux adoption is already exploding, more than past years. 2026 might really be the much mythologized year of desktop Linux. But if Intel ships a little Linux box into every other living room in America, OEMs and enterprise takes notice.
This isn't just a play for games. It's a real cultural reset.
However, Intel won't do it. It's not that it can't, as such, but it goes against two of their biggest anti-traits:
1. It requires restraint, which Intel doesn't have. They must NOT build an ecosystem or chase subscriptions.
2. It requires aggression, which Intel doesn't have. They MUST price it at cost and market it like a rebellion.
I really don't want or need any credit at all. But I love Intel. My inner child loves Intel. After working for them for a few years, even though my whole team was laid off last year, I want to see them succeed and return to cultural relevance. If this message reaches ONE person who has convincing power in the right meeting, hey, maybe it will happen.","title":"Tell HN: Intel could blow up the Console Wars if it had the guts","updated_at":"2026-03-05T23:23:01Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vinhnx"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"VT Code is a Rust CLI/TUI coding agent for AST-aware edits (Tree-sitter, ast-grep). Multi-provider routing with failover and caching (OpenAI, Anthropic, Gemini, DeepSeek, xAI, OpenRouter, Z.AI, Moonshot; Ollama locally). Policy-gated tools, workspace boundaries, Zed ACP integration. Config-first via vtcode.toml; reproducible model/constant metadata in the repo.
Try it: cargo install vtcode; vtcode
Code: https://github.com/vinhnx/vtcode"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: VT Code \u2013 AST-aware Rust agent for terminal (Tree-sitter/AST-grep)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/vinhnx/vtcode"}},"_tags":["story","author_vinhnx","story_45617591","show_hn"],"author":"vinhnx","created_at":"2025-10-17T15:01:22Z","created_at_i":1760713282,"num_comments":0,"objectID":"45617591","points":6,"story_id":45617591,"story_text":"VT Code is a Rust CLI/TUI coding agent for AST-aware edits (Tree-sitter, ast-grep). Multi-provider routing with failover and caching (OpenAI, Anthropic, Gemini, DeepSeek, xAI, OpenRouter, Z.AI, Moonshot; Ollama locally). Policy-gated tools, workspace boundaries, Zed ACP integration. Config-first via vtcode.toml; reproducible model/constant metadata in the repo.
Try it: cargo install vtcode; vtcode
Code: https://github.com/vinhnx/vtcode","title":"Show HN: VT Code \u2013 AST-aware Rust agent for terminal (Tree-sitter/AST-grep)","updated_at":"2026-03-05T22:49:16Z","url":"https://github.com/vinhnx/vtcode"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"arlanrakh"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"I built an AI version of Paul Graham, fully powered by the Nia API.
* Nia gives coding agents accurate context by indexing entire codebases, documentation, and packages. It fixes hallucinations by letting agents retrieve real source information instead of guessing. Developers ship faster because AI can read and understand their actual project. This is a different use case of Nia but apparently it also works.
You can chat with it and ask any question because it has access to Nia\u2019s knowledge base, which indexed all of his personal essays. The agent is able to call multiple tools that directly use Nia\u2019s API:
\u2022 NiaWebSearch - searches the web\n\u2022 searchEssays - semantic search over all essays\n\u2022 browseEssays - shows the full tree of essays\n\u2022 listDirectory - lists essays in a path\n\u2022 readEssay - reads full essay content\n\u2022 grepEssays - regex pattern search\n\u2022 getSourceContent - retrieves full source by identifier
Models:\nanthropic/claude-sonnet-4.5\nmoonshotai/kimi-k2-thinking\nxai/grok-4-fast-reasoning\nalibaba/qwen3-vl-thinking
It\u2019s fully free and open source.
Try it: https://www.paulgraham-nia.com/\nCode: https://github.com/nozomio-labs/paulgraham-ai"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai"],"value":"Show HN: AI Paul Graham"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.paulgraham-nia.com/"}},"_tags":["story","author_arlanrakh","story_46180304","show_hn"],"author":"arlanrakh","children":[46181142,46185347,46190254],"created_at":"2025-12-07T09:11:49Z","created_at_i":1765098709,"num_comments":5,"objectID":"46180304","points":4,"story_id":46180304,"story_text":"I built an AI version of Paul Graham, fully powered by the Nia API.
* Nia gives coding agents accurate context by indexing entire codebases, documentation, and packages. It fixes hallucinations by letting agents retrieve real source information instead of guessing. Developers ship faster because AI can read and understand their actual project. This is a different use case of Nia but apparently it also works.
You can chat with it and ask any question because it has access to Nia\u2019s knowledge base, which indexed all of his personal essays. The agent is able to call multiple tools that directly use Nia\u2019s API:
\u2022 NiaWebSearch - searches the web\n\u2022 searchEssays - semantic search over all essays\n\u2022 browseEssays - shows the full tree of essays\n\u2022 listDirectory - lists essays in a path\n\u2022 readEssay - reads full essay content\n\u2022 grepEssays - regex pattern search\n\u2022 getSourceContent - retrieves full source by identifier
Models:\nanthropic/claude-sonnet-4.5\nmoonshotai/kimi-k2-thinking\nxai/grok-4-fast-reasoning\nalibaba/qwen3-vl-thinking
It\u2019s fully free and open source.
Try it: https://www.paulgraham-nia.com/\nCode: https://github.com/nozomio-labs/paulgraham-ai","title":"Show HN: AI Paul Graham","updated_at":"2026-03-05T23:09:23Z","url":"https://www.paulgraham-nia.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"marsninja"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"I built this tool (kimi-fetch) to download the full 1.5TB moonshotai/Kimi-K3 from Hugging Face, one file at a time, and survive being stopped and restarted for any reason (Ctrl-C, kill -9, power loss).
If anything happens you can run it again and it resumes midfile until all 100 plus files are complete and verified with a sha256 hash.
The tool itself is a 68kb standalone binary that runs on mac and linux (wsl if you are on windows). And when you run it ./kimi-fetch it will immediately start downloading k3 from hungingface with max restart-ability. And will verify everything it downloads. I'll likely add a feature in a future release to have it automatically create 3 copies after for triple modular redundancy via a flag, but users can do that themselves for now.
I created this because I couldn't find a focused tool just for this type of thing, and I want to get this done with then next 24 hrs, because there is no telling what the government might do.
Its easy to install via a single curl command. And you simply have ot run ./kimi-fetch and it createa Kimi-K3 folder and gets going. Would love feedback."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Easiest tool (65kb binary) to download Kimi K3 fast (while we can)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/marsninja/kimi-fetch"}},"_tags":["story","author_marsninja","story_49073376","show_hn"],"author":"marsninja","children":[49074333,49077309],"created_at":"2026-07-27T18:05:47Z","created_at_i":1785175547,"num_comments":2,"objectID":"49073376","points":4,"story_id":49073376,"story_text":"I built this tool (kimi-fetch) to download the full 1.5TB moonshotai/Kimi-K3 from Hugging Face, one file at a time, and survive being stopped and restarted for any reason (Ctrl-C, kill -9, power loss).
If anything happens you can run it again and it resumes midfile until all 100 plus files are complete and verified with a sha256 hash.
The tool itself is a 68kb standalone binary that runs on mac and linux (wsl if you are on windows). And when you run it ./kimi-fetch it will immediately start downloading k3 from hungingface with max restart-ability. And will verify everything it downloads. I'll likely add a feature in a future release to have it automatically create 3 copies after for triple modular redundancy via a flag, but users can do that themselves for now.
I created this because I couldn't find a focused tool just for this type of thing, and I want to get this done with then next 24 hrs, because there is no telling what the government might do.
Its easy to install via a single curl command. And you simply have ot run ./kimi-fetch and it createa Kimi-K3 folder and gets going. Would love feedback.","title":"Show HN: Easiest tool (65kb binary) to download Kimi K3 fast (while we can)","updated_at":"2026-07-29T02:08:40Z","url":"https://github.com/marsninja/kimi-fetch"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vinhnx"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"I\u2019ve been building VT Code, a Rust-based terminal coding agent that combines semantic code intelligence (Tree-sitter + ast-grep) with multi-provider LLMs and a defense-in-depth execution model. It runs in your terminal with a streaming TUI, and also integrates with editors via ACP and a VS Code extension.
* Semantic understanding: parses your code with Tree-sitter and does structural queries with ast-grep.
* Multi-LLM with failover: OpenAI, Anthropic, xAI, DeepSeek, Gemini, Z.AI, Moonshot, OpenRouter, MiniMax, and Ollama for local\u2014swap by env var.
* Security first: tool allowlist + per-arg validation, workspace isolation, optional Anthropic sandbox, HITL approvals, audit trail.
* Editor bridges: Agent Conext Protocol supports (Zed); VS Code extension (also works in Open VSX-compatible editors like Cursor/Windsurf).
* Configurable: vtcode.toml with tool policies, lifecycle hooks, context budgets, and timeouts.
GitHub: https://github.com/vinhnx/vtcode"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: VT Code \u2013 Rust TUI coding agent with Tree-sitter and AST-grep"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/vinhnx/vtcode"}},"_tags":["story","author_vinhnx","story_45842578","show_hn"],"author":"vinhnx","children":[45842717],"created_at":"2025-11-07T01:14:04Z","created_at_i":1762478044,"num_comments":2,"objectID":"45842578","points":3,"story_id":45842578,"story_text":"I\u2019ve been building VT Code, a Rust-based terminal coding agent that combines semantic code intelligence (Tree-sitter + ast-grep) with multi-provider LLMs and a defense-in-depth execution model. It runs in your terminal with a streaming TUI, and also integrates with editors via ACP and a VS Code extension.
* Semantic understanding: parses your code with Tree-sitter and does structural queries with ast-grep.
* Multi-LLM with failover: OpenAI, Anthropic, xAI, DeepSeek, Gemini, Z.AI, Moonshot, OpenRouter, MiniMax, and Ollama for local\u2014swap by env var.
* Security first: tool allowlist + per-arg validation, workspace isolation, optional Anthropic sandbox, HITL approvals, audit trail.
* Editor bridges: Agent Conext Protocol supports (Zed); VS Code extension (also works in Open VSX-compatible editors like Cursor/Windsurf).
* Configurable: vtcode.toml with tool policies, lifecycle hooks, context budgets, and timeouts.
GitHub: https://github.com/vinhnx/vtcode","title":"Show HN: VT Code \u2013 Rust TUI coding agent with Tree-sitter and AST-grep","updated_at":"2026-03-05T22:58:50Z","url":"https://github.com/vinhnx/vtcode"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pouzy"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"I've been discussing with a few people and we were wondering in what way has the YC portfolio changed over the years.
Is there a shift from consumer-facing possible moonshots (Airbnb, etc.) to more "world problem solving" ideas, like genetics, big data, renewable energies... ?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What ideas is YC now mostly looking for? Has it changed over the years?"},"url":{"matchLevel":"none","matchedWords":[],"value":""}},"_tags":["story","author_pouzy","story_9937567","ask_hn"],"author":"pouzy","children":[9937736],"created_at":"2015-07-23T18:29:11Z","created_at_i":1437676151,"num_comments":1,"objectID":"9937567","points":2,"story_id":9937567,"story_text":"I've been discussing with a few people and we were wondering in what way has the YC portfolio changed over the years.
Is there a shift from consumer-facing possible moonshots (Airbnb, etc.) to more "world problem solving" ideas, like genetics, big data, renewable energies... ?","title":"Ask HN: What ideas is YC now mostly looking for? Has it changed over the years?","updated_at":"2024-09-19T22:08:41Z","url":""},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kemyd"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"Hey HN,
We built a different kind of AI benchmark for UI generation.
Instead of static leaderboards or curated screenshots, you can watch multiple models generate the same design live, side-by-side, and decide which output is actually better.
Under the hood, we call AI models from Anthropic (Opus), OpenAI (GPT), Google (Gemini), and Moonshot AI (Kimi).
Each model generates a real, editable project using Tailwind CSS (not screenshots or canvas exports). You can export it for Next.js, Laravel (Blade), Symfony (Twig), WordPress, or plain HTML.
What we noticed building this:
* Popular benchmarks don't reflect UX/UI quality. For a different prompt, one model is better than another (that's why live comparison on a single screen matters).
* Some models overuse wrappers/div soup. Some hallucinate layout constraints.
* Kimi likes Cyrillic, even if all other models won't use it for the same prompt.
The interesting part wasn't ranking models. It was making their outputs easier for humans to compare visually.
Short demo:\nhttps://www.youtube.com/watch?v=RCTZlvqMQdc
Curious whether this feels more useful than traditional leaderboard-style AI benchmarks.
Happy to answer technical questions.
Example for HN:
Prompt: Redesign the Hacker News website for 2030, including sample entries that could realistically appear on the platform in that year.
Results: https://shuffle.dev/ai-design/Tjjy7XAFMq25AI
Previews:
Opus: \nhttps://shuffle.dev/preview/d6d5ba4eeede381cee7e30c697f010c7...
GPT: \nhttps://shuffle.dev/preview/f050359977c1d6dc6c8fc104a24b83c3...
Gemini: \nhttps://shuffle.dev/preview/eab78f9748a6d8ccecb94a8b0390f044...
Kimi: \nhttps://shuffle.dev/preview/394bb596a8efa50342db4dc88c5f9fab..."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai"],"value":"Show HN: Real-Time AI Design Benchmark"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai"],"value":"https://shuffle.dev/ai-design"}},"_tags":["story","author_kemyd","story_47135831","show_hn"],"author":"kemyd","created_at":"2026-02-24T11:35:21Z","created_at_i":1771932921,"num_comments":0,"objectID":"47135831","points":2,"story_id":47135831,"story_text":"Hey HN,
We built a different kind of AI benchmark for UI generation.
Instead of static leaderboards or curated screenshots, you can watch multiple models generate the same design live, side-by-side, and decide which output is actually better.
Under the hood, we call AI models from Anthropic (Opus), OpenAI (GPT), Google (Gemini), and Moonshot AI (Kimi).
Each model generates a real, editable project using Tailwind CSS (not screenshots or canvas exports). You can export it for Next.js, Laravel (Blade), Symfony (Twig), WordPress, or plain HTML.
What we noticed building this:
* Popular benchmarks don't reflect UX/UI quality. For a different prompt, one model is better than another (that's why live comparison on a single screen matters).
* Some models overuse wrappers/div soup. Some hallucinate layout constraints.
* Kimi likes Cyrillic, even if all other models won't use it for the same prompt.
The interesting part wasn't ranking models. It was making their outputs easier for humans to compare visually.
Short demo:\nhttps://www.youtube.com/watch?v=RCTZlvqMQdc
Curious whether this feels more useful than traditional leaderboard-style AI benchmarks.
Happy to answer technical questions.
Example for HN:
Prompt: Redesign the Hacker News website for 2030, including sample entries that could realistically appear on the platform in that year.
Results: https://shuffle.dev/ai-design/Tjjy7XAFMq25AI
Previews:
Opus: \nhttps://shuffle.dev/preview/d6d5ba4eeede381cee7e30c697f010c7...
GPT: \nhttps://shuffle.dev/preview/f050359977c1d6dc6c8fc104a24b83c3...
Gemini: \nhttps://shuffle.dev/preview/eab78f9748a6d8ccecb94a8b0390f044...
Kimi: \nhttps://shuffle.dev/preview/394bb596a8efa50342db4dc88c5f9fab...","title":"Show HN: Real-Time AI Design Benchmark","updated_at":"2026-03-05T23:36:15Z","url":"https://shuffle.dev/ai-design"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ScioNos"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"Hi HN!
Moonshot AI released Kimi K2.5 today, and we integrated it on RouterLab within hours.
Why this matters:
*Open source beats proprietary:*\n\u2022 Kimi K2.5: 50.2% on HLE (Humanity's Last Exam)\n\u2022 GPT-5: 41.7%\n\u2022 Claude 4.5: 32.0%
First time an open-source model beats GPT-5 on expert-level reasoning.
*Agent Swarm architecture:*\n\u2022 Orchestrates up to 100 parallel agents\n\u2022 1,500 simultaneous tool calls\n\u2022 4.5x faster than sequential execution\n\u2022 Autonomous task decomposition
Example: "Analyze 50 competitors and create a report"\n\u2192 Creates 50 research agents\n\u2192 Parallel execution\n\u2192 Compiled report in minutes
*Technical specs:*\n\u2022 1T parameters (32B activated)\n\u2022 384 experts MoE\n\u2022 INT4 quantization native\n\u2022 256k context window\n\u2022 Open weights on Hugging Face
*Benchmarks:*\n\u2022 HLE (reasoning): 50.2% (GPT-5: 41.7%)\n\u2022 BrowseComp (web nav): 60.2% (GPT-5: 54.9%)\n\u2022 SWE-Bench (coding): 71.3%\n\u2022 VideoMMMU: 86.6%
*Pricing:*\n\u2022 $0.60/$3.00 per 1M tokens\n\u2022 5x cheaper than GPT-5
We're a Swiss company (Eyelo SA, founded 1983) and integrated Kimi K2.5 on RouterLab with Swiss/German hosting for GDPR compliance.
OpenAI-compatible API, migration is 2 lines of code:
```python\nclient = OpenAI(\n base_url="https://routerlab.ch/v1",\n api_key="your-key"\n)\nI wrote a technical analysis:\nhttps://medium.com/@comeback01/kimi-k2-5-the-agent-swarm-rev...
Try it: https://routerlab.ch/blog/kimi-k2-5 (14-day free trial)
Happy to answer technical questions about Agent Swarm, MoE architecture, or deployment!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Kimi K2.5 (Agent Swarm, beats GPT-5) now on RouterLab (Swiss hosting)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://routerlab.ch/blog/kimi-k2-5"}},"_tags":["story","author_ScioNos","story_46781155","show_hn"],"author":"ScioNos","created_at":"2026-01-27T15:21:01Z","created_at_i":1769527261,"num_comments":0,"objectID":"46781155","points":1,"story_id":46781155,"story_text":"Hi HN!
Moonshot AI released Kimi K2.5 today, and we integrated it on RouterLab within hours.
Why this matters:
*Open source beats proprietary:*\n\u2022 Kimi K2.5: 50.2% on HLE (Humanity's Last Exam)\n\u2022 GPT-5: 41.7%\n\u2022 Claude 4.5: 32.0%
First time an open-source model beats GPT-5 on expert-level reasoning.
*Agent Swarm architecture:*\n\u2022 Orchestrates up to 100 parallel agents\n\u2022 1,500 simultaneous tool calls\n\u2022 4.5x faster than sequential execution\n\u2022 Autonomous task decomposition
Example: "Analyze 50 competitors and create a report"\n\u2192 Creates 50 research agents\n\u2192 Parallel execution\n\u2192 Compiled report in minutes
*Technical specs:*\n\u2022 1T parameters (32B activated)\n\u2022 384 experts MoE\n\u2022 INT4 quantization native\n\u2022 256k context window\n\u2022 Open weights on Hugging Face
*Benchmarks:*\n\u2022 HLE (reasoning): 50.2% (GPT-5: 41.7%)\n\u2022 BrowseComp (web nav): 60.2% (GPT-5: 54.9%)\n\u2022 SWE-Bench (coding): 71.3%\n\u2022 VideoMMMU: 86.6%
*Pricing:*\n\u2022 $0.60/$3.00 per 1M tokens\n\u2022 5x cheaper than GPT-5
We're a Swiss company (Eyelo SA, founded 1983) and integrated Kimi K2.5 on RouterLab with Swiss/German hosting for GDPR compliance.
OpenAI-compatible API, migration is 2 lines of code:
```python\nclient = OpenAI(\n base_url="https://routerlab.ch/v1",\n api_key="your-key"\n)\nI wrote a technical analysis:\nhttps://medium.com/@comeback01/kimi-k2-5-the-agent-swarm-rev...
Try it: https://routerlab.ch/blog/kimi-k2-5 (14-day free trial)
Happy to answer technical questions about Agent Swarm, MoE architecture, or deployment!","title":"Show HN: Kimi K2.5 (Agent Swarm, beats GPT-5) now on RouterLab (Swiss hosting)","updated_at":"2026-03-05T23:26:50Z","url":"https://routerlab.ch/blog/kimi-k2-5"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nateb2022"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Related: Kimi-K3 Technical Report [pdf] - https://news.ycombinator.com/item?id=49070985"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi-K3 on HuggingFace"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://huggingface.co/moonshotai/Kimi-K3"}},"_tags":["story","author_nateb2022","story_49065752"],"author":"nateb2022","children":[49065757,49065868,49066164,49066165,49066200,49066201,49066235,49066240,49066343,49066351,49066354,49066398,49066399,49066492,49066495,49066509,49066536,49066552,49066672,49066733,49066752,49066917,49066981,49066990,49067146,49067204,49067206,49067228,49067369,49067492,49067498,49067501,49067514,49067628,49067657,49067699,49067738,49068616,49068905,49068959,49069140,49069214,49069344,49069392,49069841,49070385,49070397,49070490,49070499,49070527,49070707,49070732,49070758,49070833,49070851,49070887,49070898,49070962,49070979,49070982,49071072,49071410,49072873,49073251,49073254,49073269,49073424,49073571,49073615,49073821,49075078,49075205,49075315,49076001,49076242,49076281,49078395,49078596,49080388,49120431],"created_at":"2026-07-27T06:18:10Z","created_at_i":1785133090,"num_comments":544,"objectID":"49065752","points":1382,"story_id":49065752,"story_text":"Related: Kimi-K3 Technical Report [pdf] - https://news.ycombinator.com/item?id=49070985","title":"Kimi-K3 on HuggingFace","updated_at":"2026-09-04T18:26:22Z","url":"https://huggingface.co/moonshotai/Kimi-K3"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nekofneko"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://moonshotai.github.io/Kimi-K2/thinking.html"}},"_tags":["story","author_nekofneko","story_45836070"],"author":"nekofneko","children":[45836876,45837113,45837162,45837175,45837190,45837253,45837271,45837282,45837352,45837459,45837481,45837483,45837492,45837573,45837602,45837617,45837700,45837982,45838017,45838227,45838252,45838443,45838612,45838641,45838900,45838932,45839591,45839656,45839713,45839944,45840335,45840934,45841202,45841372,45842447,45842517,45842727,45843056,45843222,45843278,45843373,45843397,45843813,45844816,45845544,45845878,45846342,45847305,45854485,45855642],"created_at":"2025-11-06T15:06:06Z","created_at_i":1762441566,"num_comments":427,"objectID":"45836070","points":936,"story_id":45836070,"title":"Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model","updated_at":"2026-04-12T12:38:59Z","url":"https://moonshotai.github.io/Kimi-K2/thinking.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nekofneko"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi K2.7-Code: open-source coding model with better token efficiency"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://huggingface.co/moonshotai/Kimi-K2.7-Code"}},"_tags":["story","author_nekofneko","story_48502347"],"author":"nekofneko","children":[48502638,48502661,48502859,48503148,48503161,48503169,48503278,48503345,48503463,48503528,48503581,48503595,48503625,48503638,48503855,48504011,48504348,48504565,48505405,48505456,48505949,48506499,48507096,48507538,48507923,48514054,48537804,48538566],"created_at":"2026-06-12T10:42:50Z","created_at_i":1781260970,"num_comments":240,"objectID":"48502347","points":463,"story_id":48502347,"title":"Kimi K2.7-Code: open-source coding model with better token efficiency","updated_at":"2026-08-08T09:22:44Z","url":"https://huggingface.co/moonshotai/Kimi-K2.7-Code"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vinhnx"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Related: Kimi-K3 on HuggingFace - https://news.ycombinator.com/item?id=49065752"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi-K3 Technical Report [pdf]"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf"}},"_tags":["story","author_vinhnx","story_49070985"],"author":"vinhnx","children":[49071129,49071185,49071249,49071277,49071313,49071435,49071483,49071507,49071516,49071529,49071548,49071555,49071564,49071662,49071908,49072022,49072105,49072109,49072173,49072745,49074458,49074612,49074700,49074819,49075045,49082130],"created_at":"2026-07-27T15:23:45Z","created_at_i":1785165825,"num_comments":183,"objectID":"49070985","points":391,"story_id":49070985,"story_text":"Related: Kimi-K3 on HuggingFace - https://news.ycombinator.com/item?id=49065752","title":"Kimi-K3 Technical Report [pdf]","updated_at":"2026-09-03T21:29:33Z","url":"https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vinhnx"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi K2.5 Technical Report [pdf]"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://github.com/MoonshotAI/Kimi-K2.5/blob/master/tech_report.pdf"}},"_tags":["story","author_vinhnx","story_46826597"],"author":"vinhnx","children":[46828650,46828776,46828903,46829045,46829170,46829272,46829352,46829977,46830275,46831013,46831104,46831371,46831419,46831485,46831757,46831790,46832278,46833409,46833693,46834096,46834169,46834336,46835609,46836341,46836819,46837404,46838184,46838830],"created_at":"2026-01-30T16:43:50Z","created_at_i":1769791430,"num_comments":141,"objectID":"46826597","points":388,"story_id":46826597,"title":"Kimi K2.5 Technical Report [pdf]","updated_at":"2026-06-11T09:47:42Z","url":"https://github.com/MoonshotAI/Kimi-K2.5/blob/master/tech_report.pdf"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ConteMascetti71"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://github.com/MoonshotAI/Kimi-K2"}},"_tags":["story","author_ConteMascetti71","story_44543508"],"author":"ConteMascetti71","children":[44544036,44544300,44548861],"created_at":"2025-07-12T17:26:24Z","created_at_i":1752341184,"num_comments":2,"objectID":"44543508","points":352,"story_id":44543508,"title":"Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model","updated_at":"2025-10-27T09:17:14Z","url":"https://github.com/MoonshotAI/Kimi-K2"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"GaggiX"},"title":{"matchLevel":"none","matchedWords":[],"value":"Attention Residuals"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://github.com/MoonshotAI/Attention-Residuals"}},"_tags":["story","author_GaggiX","story_47458595"],"author":"GaggiX","children":[47459229,47459405,47459484,47459736,47460036,47460324,47461658,47462031,47462439,47464269,47465573,47465749,47466481,47466652,47467134],"created_at":"2026-03-20T18:23:41Z","created_at_i":1774031021,"num_comments":34,"objectID":"47458595","points":241,"story_id":47458595,"title":"Attention Residuals","updated_at":"2026-08-08T08:54:44Z","url":"https://github.com/MoonshotAI/Attention-Residuals"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"blackcat201"},"title":{"matchLevel":"none","matchedWords":[],"value":"Kimi Linear: An Expressive, Efficient Attention Architecture"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["moonshot","ai"],"value":"https://github.com/MoonshotAI/Kimi-Linear"}},"_tags":["story","author_blackcat201","story_45766937"],"author":"blackcat201","children":[45768006,45768363,45768708,45769528,45769733,45769758,45771056,45772041,45772203,45772588,45774379],"created_at":"2025-10-31T00:07:36Z","created_at_i":1761869256,"num_comments":47,"objectID":"45766937","points":217,"story_id":45766937,"title":"Kimi Linear: An Expressive, Efficient Attention Architecture","updated_at":"2026-07-28T19:20:53Z","url":"https://github.com/MoonshotAI/Kimi-Linear"}],"hitsPerPage":50,"nbHits":156,"nbPages":4,"page":0,"params":"query=Moonshot+AI&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":10,"processingTimingsMS":{"_request":{"queue":38,"roundTrip":20},"afterFetch":{"format":{"highlighting":1,"total":1},"merge":{"mergeLoop":{"total":1},"total":1},"total":1},"fetch":{"query":5,"scanning":2,"total":8},"total":10},"query":"Moonshot AI","serverTimeMS":50}