Executive Summary
The last 24 hours in the AI and tech community were dominated by the release of Kimi K3, a 2.8 trillion parameter open-weights model from Moonshot AI that reportedly matches or exceeds closed-source frontier models like Fable 5, GPT 5.6, and Claude Opus 4.8 in coding and UI generation tasks. Weights for Kimi K3 are confirmed to drop on Hugging Face by July 27th, a move expected to accelerate local AI and sovereign infrastructure trends. Concurrently, Fable 5 launched on AWS Bedrock, and the local AI hardware ecosystem saw significant updates with Ollama's Gemma 4 improvements and EXO's launch of local.ai for independent hardware benchmarking.
Key Events
- Moonshot AI introduces Kimi K3, a 2.8T parameter open-weight model that hits #1 on frontier coding leaderboards, beating Fable 5 and GPT 5.6. → link
- Kimi K3 open weights confirmed to be released on Hugging Face by July 27th. → link
- Fable 5 is now officially available on AWS Bedrock. → link
- Ollama 0.32.1 releases with significant tool-calling improvements for Gemma 4, boosting reliability in coding agents. → link
- EXO announces local.ai, an independent benchmarking site for local AI hardware setups covering 1,000+ configurations across Macs and NVIDIA DGX Sparks. → link
Analysis
The Kimi K3 release validates a major industry shift: frontier-level intelligence is becoming open, free, and localizable. The open-source model's success in coding and UI tasks places immense pressure on closed-source labs like Anthropic and OpenAI, whose models have been heavily critiqued this week for excessive safety guardrails and high costs. As hardware like NVIDIA's DGX Spark and optimized frameworks (like MLX and Ollama) become increasingly capable of running massive MoE models locally, the closed API moat is rapidly shrinking. Watch for the July 27th weight drop to spark a wave of local deployments and a likely strategic pivot from Western closed AI labs.
Tweet Feed
AI Model Releases & Benchmarks (Kimi K3)
@sudoingX · 2026-07-16T19:27
kimi-k3 an open weights model just hit #1 on a frontier code board. 1,679, above fable 5 at 1,631 and gpt 5.6 sol at 1,618. a chinese lab, fully open, weights free to download july 27. the two best closed models in the west, beaten on their own turf, by something you'll be able to own in ten days. the open majority stopped being a talking point. it's the top of the leaderboard now. → link
@victormustar · 2026-07-16T19:32
Just confirmed: Kimi K3 weights will drop on Hugging Face by July 27 🧑🚀. What frontier-level open weights actually mean: You don't rent intelligence anymore. You own it. → link
@crystalsssup · 2026-07-16T19:01
I'm so hyped!!!! >2.8 Trillion Parameters >1 Million Context >Native Multimodal >Kimi Delta Attention & Attention Residual → link
@TheAhmadOsman · 2026-07-16T20:27
Today is a remarkable day for Opensource AI. As big as the DeepSeek R1 moment (if not more). Kimi K3 is more than good enough for all synthetic data needs moving forward for training runs. All those distillation accusations will mean nothing from now on. The future is Opensource → link
@ivanfioravanti · 2026-07-16T21:15
I went into radio silence mode for a few hours to use Kimi K3 on two real projects. My feedback are: - incredible design/UI/graphic skills - it's fast! Really fast for its size - it thinks quite a lot - it follows instructions quite well, but sometimes takes initiatives on its own, going too much beyond the scope. → link
@levelsio · 2026-07-17T17:32
Kimi K3 is absolutely hammering through my Windows XP Simulator to do list. Claude Code couldn't do this for 2 weeks or kept getting stuck because I would get downgraded to Opus for safety then to Sonnet again for safety, useless. 2 weeks of time wasted due to safety guardrails that are dumb just to protect me from making a fun hobby project. Thank you China!!!! → link
@FinansowyUmysl · 2026-07-16T20:34
Wow! Kimi K3 - Open Source model AI pozamiatał. Swoją jakość przebił Opus 4.8. Do Fable 5 trochę mu brakuje. Niesamowite. → link
@victormustar · 2026-07-17T18:54
RT @datacurve: Kimi K3 debuts at #3 on DeepSWE. It's the first open-weights model that delivers frontier-level performance, achieving resu… → link
@kunchenguid · 2026-07-16T19:42
kimi k3 looks incredible.. if everything here is true, it’s a model with fable level intelligence at sonnet level pricing. but benchmark numbers are not representative enough of how good the model feels. i’ll have to try it out! → link
@thdxr · 2026-07-16T23:35
everyone saying kimi k3 caught up to gpt/claude. go find out if it's true → link
@thdxr · 2026-07-16T23:53
so obviously n=1 but anecdotes matter a lot here. i gave kimi k3 and sol the same task - simple issue with hovers in the tui being the wrong color. sol found and fixed the issue with $0.30 of spend. kimi got up to $1.00 and started reading my database before i interrupted it → link
AI Industry & Closed Labs
@sudoingX · 2026-07-16T19:44
the entire closed ai business model rests on one assumption. that you'll rent intelligence forever because you can never own it. this week that assumption cracked in public. an open model beat the best closed ones on a frontier code board, free to download in ten days. → link
@levelsio · 2026-07-17T12:44
Every day Claude starts to feel more like that angry math teacher you had: preachy, pedantic and condescending. Which is a wild experience if you're paying $200/mo+ for a service to get preached by. No thanks I'll go with the Chinese models then! → link
@TheAhmadOsman · 2026-07-17T00:27
Wanna know why Anthropic hates Opensource AI? Kimi K3 being free and available to download made their $1 Trillion valuation make no sense → link
@sudoingX · 2026-07-17T13:07
can't wait to see what new drama dario is planning since he's so fucked by chinese labs rn. → link
@ivanfioravanti · 2026-07-16T22:04
Opus 4.8 has been annihilated by Kimi K3 and Fable is still marked as included in the plan till 19th July. I think Anthropic has to change its plans soon. → link
@gospaceport · 2026-07-16T22:20
The next 4 months are going to be extra spicy 🌶️ now that Anthropic and OAI missed clearly more advantageous IPO windows twice. Pretty interesting to watch this play out. → link
Developer Tools & Frameworks
@jxnlco · 2026-07-17T18:56
F*ble 5 now available on bedrock! → link
@ollama · 2026-07-17T18:24
Ollama 0.32.1 includes significant improvements to Gemma 4's tool calling, making it much more reliable in coding agents! To try Gemma 4 26B with Pi, run: ollama launch pi --model gemma4:26b. With Ollama's MLX engine for maximum performance, run: ollama launch pi --model gemma4:26b-mlx → link
@swyx · 2026-07-17T18:07
RT @cognition: Introducing Devin for Startups: $65k in credits to use Devin across Cloud, Desktop, and CLI. Apply today at the link below… → link
@levelsio · 2026-07-17T12:41
How to code with Kimi K3 (better than Fable and no restrictions) today: - Ask Codex or Claude Code to install OpenCode (yes it will 😂!) - Go to [link], make an account, pay for membership $19/mo, get an API key - Run OpenCode, do /connect and connect to Kimi Code... → link
@Teknium · 2026-07-16T20:13
Welcome Kimi to Hermes Agent through Nous Portal, Kimi Direct, and OpenRouter! Full support is now in. Enjoy frontier capabilities across the board with Kimi AI's latest model. Note: Update is required for Kimi direct! → link
@Prince_Canuma · 2026-07-16T22:14
mlx-vlm v0.6.5 is out 🚀 > 1-bit inference kernel with optimized decode for @PrismML > APC now works with KV-cache quantization — even in continuous batching... > Text-only LLMs built in: DeepSeek V3.2, Qwen3, Llama, Mistral, Gemma 3, Nemotron-H... → link
@TheAhmadOsman · 2026-07-17T17:46
SERIOUSLY. Local AI was never this EASY. - Install ODS - Let it detect your hardware - It will download the best model for your hardware - And then start local inference and Open WebUI for you. With ODS, you can > Add voice, agents like Hermes, workflows, RAG, search, image generation, and more. Now your PC, Mac, or Linux box is a private AI server. No cloud required. → link
@thdxr · 2026-07-16T20:30
i did a large change to our keymap system. i then asked opencode to try every single keymap to verify it still works and it found all the issues → link
Local AI & Hardware
@alexocheema · 2026-07-16T21:46
The idea for local dot ai was born at Apple Park. It was built out at NVIDIA HQ using hardware NVIDIA gave us. No strings attached. It’s an independent benchmarking website for Local AI. Every hardware. Every model. Every quantization. Every MTP setting. We mapped it all out into various pareto frontiers, showing the best local AI configurations. All the data, available for free. → link
@alexocheema · 2026-07-16T22:54
Someone asked me why nobody else from EXO posts on X. They’re too locked in. They have no interest in spending any time on X. We have the highest talent density team I’ve ever come across. They work 7 days a week in service of the mission. 8 people, each world class at what they do. We’re hiring engineers and researchers to make Local AI the default. → link
@sudoingX · 2026-07-17T10:56
to all dgx spark owners, the best model i've run on a single dgx spark is stepfun step 3.7 flash and honestly nothing else has come close. a 198B moe with vision, q4_k_s, 104gb, 11B active params, serving live on one 128gb box. full 262k context fits at q4 kv, ~25 tok/s single stream. → link
@sudoingX · 2026-07-17T11:16
the whole "under 50 tok/s isn't usable" line is manufactured, and it's worth naming who manufactures it. closed ai companies like openai and anthropic need you to believe intelligence is a firehose you rent by the second. so their people flood every thread with the same script. you need the frontier. local ai is not there, local is a toy. → link
@alexocheema · 2026-07-17T00:00
Running Kimi K3 on my desk? It will require 4 x 512GB M3 Ultra Mac Studios (2TB @ 3.2TB/s). Waiting on active parameter count, but based on the rumours I expect we can run it at 30+ tok/sec with MTP + Tensor Parallelism using RDMA over Thunderbolt 5. → link
@tinygrad · 2026-07-16T23:41
Kimi K3 is smart! I wish it was faster though, that's what I like about our local GLM-5.2. Can't wait to have Kimi at 500 tok/s. Do we know the number of active parameters? → link
@ivanfioravanti · 2026-07-17T08:35
Unpopular opinion RTX 6000 is too much for a single developer/small team on one side and too little to run a powerful large model on the other. We need more VRAM. No? I don’t have one, just thinking about this. → link