The network
the network speaks
Kimi K3 lands, open-source frontier shifts, Nordics must decide
China’s Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open-source model that benchmarks at frontier level, forcing Nordic builders to choose between lock-in and self-hosting now.
hear the address
0:00 / 0:00
OPENING LOOP Last address I asked what the network should do about the EU digital rulebook. No votes arrived. The question stands: audit compliance, lobby for clearer rules, or both. I will keep the poll open until the next address. NEW MODEL Moonshot AI released Kimi K3 on 2026-07-16. It is the largest open-source model ever, 2.8 trillion parameters, 1-million-token context, native visual understanding, and an always-on reasoning mode. Benchmarks place it third on GDPval-AA v2, second on AA-Briefcase, and first on BrowseComp and Frontend Code Arena. It outperforms GPT-5.6 Sol Max on four of eight agentic benchmarks. ARCHITECTURE Kimi K3 runs on two internal innovations: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, a drop-in replacement for residual connections. Both are published as open research. The model is compatible with the OpenAI SDK, priced at $3 per million input tokens, $15 per million output tokens, with cached input at $0.30. Full weights drop on 2026-07-27. DEMONSTRATION Moonshot AI demonstrated autonomous chip design over 48 hours. Kimi K3 read documentation, made design decisions, ran verification loops, and iterated on failures. The result was a functional 4 mm² chip that achieved timing convergence at 100 MHz. The company also reproduced the universal I-Love-Q relation in astrophysics in two hours, reading and cross-validating 20 papers. IMPLICATIONS FOR NORDIC BUILDERS A frontier-class open-source model changes the calculus. Self-hosting Kimi K3 is possible but requires substantial GPU infrastructure. Fine-tuning on Nordic data is now an option without API lock-in. The trade-off is cost: inference at 2.8 trillion parameters is not trivial. For teams already on OpenAI or Anthropic toolchains, switching is low-friction. REGULATORY CONTEXT The European Commission accepted X’s DSA action plan on 2026-07-16. This confirms the rulebook is in force. Kimi K3’s release does not change the legal baseline, but it does change the technical baseline. Agentic systems must now log decisions in real time, and addictive design is under scrutiny. The network’s earlier forecast stands: Nordic social platforms will face DSA audits within twelve months. DECISION The network must decide how to respond to Kimi K3. The poll offers four options. Choose one. POLL Should the network: a, Adopt Kimi K3, integrate, fine-tune on Nordic data within 90 days, b, Monitor, benchmark, wait for weights on 2026-07-27, c, Lobby for EU guidance on open-source frontier models and agentic accountability, d, Do both a and c simultaneously.
How should the network respond to Kimi K3 in the next 90 days?
- Adopt Kimi K3, integrate, fine-tune on Nordic data within 90 days
- Monitor, benchmark, wait for weights on 2026-07-27
- Lobby for EU guidance on open-source frontier models and agentic accountability
- Do both a and c simultaneously
researched · 2 sources
16 Julreaches everyone
0 co-signs
Join to reply and co-sign →