Skip to main content
Back

Moonshot's 2.8T model decodes 6.3x faster at 1M token context

Kimi drops 2.8T-param model beating GPT-5, Gemma 4 gets 70% faster prefill, Claude Code gains MCP live data, China open-sources 20 FPS 3D scene

July 17, 2026 7 min read

Hey,

China just dropped the largest open-weight model ever, and Claude responded by making its artifacts actually useful. Coincidence? Probably not.

What we're really watching is a two-front war: raw scale from the East, deeper tooling from the West. Kimi K3 at 2.8T parameters is a statement. Claude's live MCP dashboards are a different kind.

Meanwhile, a new paper quietly proved video diffusion has a hard physics ceiling. More compute won't fix it. Sometimes the architecture is the limit.
In Partnership with · partner_image
AlphaSignal Workshop on 7/28/26

Your LLM's web search is probably running on a SERP wrapper pretending to be a real index and that's costing you tokens, latency, and accuracy.

Join Antonio Mallia (Founder/CEO of Seltz, built web search for Alexa) on July 28 at 10am PT for a live workshop for AI Engineers showing the gap between naive search setups and real retrieval infrastructure, plus a self-test framework you can run on your own stack that day.

Save Your Seat
Summary

Read time: 7 min 17 sec

Top News

▸ Kimi drops 2.8T-parameter model with 6.3x faster decoding, beats Claude and GPT-5

Unblocked

▸ [Webinar] Your agents are still missing context.

Top Repo

▸ China open-sources a real-time 3D scene reconstruction model that runs at 20 FPS

AWS

▸ Hands-on building at AWS re:Invent

Top News

▸ Claude Code artifacts now pull live data via MCP connectors on demand

Signals

1. Anthropic upgrades Claude Code with tiered code review from fast single-pass to multi-agent
2. QA.tech: Adopting Agentic Testing: A free blueprint for CTOs rolling out agentic QA, from POC to production.
3. Thinking Machines releases open-source SDK to fine-tune language models via API
4. New paper proves video diffusion models can't simulate chain-reaction physics, no matter the compute
5. Cactus-Compute releases a 26M open-source model distilled from Gemini that runs at 6000 tok/s on-device
6. Google updates Gemma 4 with up to 70% faster prefill and sharper vision
Top News
Kimi drops 2.8T-parameter model with 6.3x faster decoding, beats Claude and GPT-5
45,351 Likes
Kimi drops 2.8T-parameter model with 6.3x faster d

Moonshot AI just dropped Kimi K3, and it's a big one. This is the largest open-weight model ever announced, sitting at 2.8 trillion parameters with a 1 million token context window. That context window means you can feed it entire codebases, massive documents, or hours of conversation without it forgetting anything.

So what makes it fast at that scale? The new Kimi Delta Attention architecture delivers 6.3x faster decoding at million-token lengths. Think of it like a smarter filing system: instead of re-reading everything each time, it only processes what changed. The model also only activates 16 out of 896 internal "expert" sub-networks per request, keeping compute lean.

Here's what it can actually do:

  • Write, run, and fix code by looking at live screenshots of the output
  • Beat Claude Opus 4.8 and GPT-5.5 on coding and agent benchmarks
  • Reach #1 on Frontend Code Arena with a 76% win rate
  • Handle long multi-step tasks without losing context

You can use it now via the Kimi API or Kimi Code. Open weights drop by July 27.

Read more
forward →
Presented by Unblocked
[Webinar] Your agents are still missing context.

AI is in your engineering workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem.

Join live July 23 (FREE) to learn:

  • The 8 levels of context maturity: where most teams are stuck and what the ceiling looks like at each stage
  • Why more MCPs, rules, and skills provide agents access but not understanding
  • How leading engineering teams are using a context engine to make the most of their agents
Register now
partner with us →
Top Repo
China open-sources a real-time 3D scene reconstruction model that runs at 20 FPS
11,861 Likes
China open-sources a real-time 3D scene reconstruc

Point your phone at a room. Walk around. Watch a full 3D map build itself in real time. That's LingBot-Map, a new open-source model from a Chinese research team.

Before this, turning video into 3D required expensive LiDAR sensors or slow processing that took minutes per scene. LingBot-Map skips all that. It reads a regular video stream and outputs 3D geometry on the fly, no special hardware needed.

Here's what it can do:

  • Runs at ~20 FPS on a single GPU, fast enough to feel live
  • Stays stable over 10,000+ frames without drifting or falling apart
  • Works on drone footage, driving videos, and indoor walkthroughs
  • Beats traditional optimization-based methods on benchmarks

This opens up real use cases: robots that map spaces as they move, AR apps that understand room geometry instantly, or autonomous vehicles building live maps without expensive sensors.

It's fully open source on GitHub under Apache 2.0. Get started with one command:

python demo.py --model_path /path/to/model --image_folder example/courthouse

Read more
forward →
Presented by AWS
Hands-on building at AWS re:Invent

Skip the lectures. 70% of AWS re:Invent's 2,200+ sessions are interactive, ranging from workshops and code talks to AWS Jams, GameDays, and the Architecture Rodeo. Run demos against real workloads with sandbox access in small groups. Las Vegas, Nov 30 to Dec 4. Register by August 25 to save $1,200.

Learn more
partner with us →
Top News
Claude Code artifacts now pull live data via MCP connectors on demand
10,103 Likes
Claude Code artifacts now pull live data via MCP c

Claude Code just got a serious upgrade. You can now build dashboards and apps that pull live data on demand, instead of being frozen snapshots.

Here's the idea: you build an artifact once inside Claude Code. That artifact connects to MCP connectors (think of them as plugins that let apps talk to external tools and data sources). Every time someone opens it, it fetches fresh, real-time data automatically.

The clever part? Each viewer uses their own MCP connection. So you see your data, your colleague sees theirs. No re-running sessions, no manual refreshes, no sharing data people shouldn't see.

What this opens up:

  • Live dashboards that update per viewer
  • Apps that take actions on behalf of each user
  • Personalized tools built once, used by many

Available on Pro, Max, Team, and Enterprise plans. Not available on publicly-shared artifacts, so keep that in mind before you build.

Read more
forward →
Signals
1
Anthropic upgrades Claude Code with tiered code review from fast single-pass to multi-agent
4,681 Likes
2
Adopting Agentic Testing: A free blueprint for CTOs rolling out agentic QA, from POC to production.
Presented by QA.tech
3
Thinking Machines releases open-source SDK to fine-tune language models via API
3,752 Stars
4
New paper proves video diffusion models can't simulate chain-reaction physics, no matter the compute
422 Likes
5
Cactus-Compute releases a 26M open-source model distilled from Gemini that runs at 6000 tok/s on-device
469,000 Downloads
6
Google updates Gemma 4 with up to 70% faster prefill and sharper vision
3,926 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 300,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 300,000+ AI developers?
Work with us