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

Why this is in the vault

Two live production implications for RDCO in this issue: (1) Claude Code artifacts now support per-viewer MCP connector calls — a direct capability upgrade to the agent harness's reporting surface; (2) Kimi K3 open weights drop July 27, reopening the open-weight routing question that was last live during DeepSeek-V4/Kimi-K2.6 coverage in April. The video diffusion physics ceiling paper is a quieter signal worth noting for content positioning.

⚠️ Sponsorship

Four paid placements in this issue:

1. Seltz — "In Partnership with" placement at top. Antonio Mallia (Founder/CEO, formerly built web search for Alexa) hosting an AlphaSignal Workshop on July 28 at 10am PT. Topic: real retrieval infrastructure vs. SERP wrappers for LLM web search. Target audience: AI engineers.
2. Unblocked — "Presented by Unblocked" between Kimi story and Top Repo. Free webinar July 23 on "8 levels of context maturity" and using a context engine for engineering agents.
3. AWS — "Presented by AWS" between Top Repo and Claude Code story. AWS re:Invent promotion (Nov 30–Dec 4, Las Vegas). Register by August 25 to save $1,200.
4. QA.tech — Native ad in Signal #2. Free agentic testing blueprint for CTOs, POC to production.

Editorial content (Top News items, Top Repo, and Signals 1, 3, 4, 5, 6) appears independent of sponsor placements.

Mapping against Ray Data Co

Claude Code MCP artifact connectors is the sharpest hit. RDCO's COO agent runs Claude Code as its primary build environment. The new per-viewer MCP connector capability in artifacts means any dashboard or reporting tool built inside Claude Code can now pull live data — inventory, board state, financial snapshots — without a separate web app. Each viewer uses their own MCP connections (so Ray's Notion/Gmail/Calendar data stays isolated from any collaborator's). This is a Pro/Max/Team/Enterprise capability and fits RDCO's current plan. Worth testing against the morning-prep or finance-pulse surfaces.

Kimi K3 open weights (July 27) reopens the open-weight routing thread. The last decision point was April 2026 when RDCO evaluated whether to route lighter tasks through NIM/open-weight endpoints to reduce Anthropic API spend (see [[2026-04-23-alphasignal-claude-via-nim-llm-router]]). Kimi K3 at 2.8T/16-active-experts and $3/$15 per million tokens API pricing is materially different from that earlier landscape — it's priced comparably to Claude Sonnet tier but with 1M context. The July 27 open-weight release is the trigger to re-evaluate.

Video diffusion physics ceiling is a content-positioning signal. The paper proves that video diffusion models fundamentally cannot simulate chain-reaction physics regardless of compute scale — the architecture is the limit. This is the kind of "compute doesn't fix it" finding that supports an RDCO content angle around benchmark literacy and AI capability honesty.

Curation section

Top News

Kimi K3 — Moonshot AI's 2.8T open-weight MoE model

• 2.8 trillion total parameters; 16 of 896 experts active per token (Stable LatentMoE)
• Kimi Delta Attention (KDA): hybrid linear attention mechanism → 6.3x faster decoding at 1M token context
• Attention Residuals: +25% training efficiency at 2% compute cost
• 1M token context window; native vision
• Benchmarks: beats Claude Opus 4.8 and GPT-5.5 on coding and agent tasks; #1 Frontend Code Arena (76% win rate)
• Can write, run, and fix code via live screenshots of output
• Available now on Kimi API and Kimi Code; open weights release July 27 (Modified MIT license)
• Pricing: $3/M input, $15/M output tokens

Claude Code artifacts + MCP connectors (live data per-viewer)

• Artifacts built in Claude Code can now call MCP connectors on demand at view time
• Per-viewer isolation: each person who opens an artifact uses their own MCP connections — data scoped to their account permissions
• Use cases: live dashboards, per-user action surfaces, personalized tools built once
• Available on Pro, Max, Team, Enterprise; not on publicly-shared artifacts
• Source: ClaudeDevs on X

Top Repo

LingBot-Map — real-time 3D scene reconstruction from video

• Open-source (Apache 2.0), from a Chinese research team
• Runs at ~20 FPS on a single GPU; no LiDAR or special sensors required
• Stable over 10,000+ frames; works on drone footage, indoor walkthroughs, driving video
• Use cases: robotics mapping, AR room understanding, autonomous vehicles
• One-command start: python demo.py --model_path /path/to/model --image_folder example/courthouse

Signals

1. Anthropic upgrades Claude Code with tiered code review — fast single-pass to multi-agent pipeline (4,681 likes)
2. [QA.tech native ad — agentic testing blueprint]
3. Thinking Machines releases open-source SDK to fine-tune language models via API (3,752 stars)
4. Video diffusion models can't simulate chain-reaction physics — new paper proves the architecture is the ceiling, not compute (422 likes)
5. Cactus-Compute 26M model distilled from Gemini; runs at 6,000 tok/s on-device (469K downloads)
6. Google Gemma 4 update — up to 70% faster prefill, sharper vision (3,926 likes)

Related

• [[2026-07-13-alphasignal-subagents-math-proof-cycle-cover]] — most recent prior issue; parallel subagents + math proof cycle
• [[2026-04-26-alphasignal-deepseek-v4-kimi-k26-agentic-ai]] — earlier Kimi K2.6 agentic AI coverage; direct predecessor to K3 analysis