#kimi

Fireworks says routing between Kimi K3 and Fable 5 reached 93% accuracy across roughly 1,030 agentic tasks. The HN debate focused on a bigger claim: single-model deployments are becoming economically wasteful.

The policy fight is no longer just about model benchmarks. Axios reports that U.S. officials have revisited tools such as Entity List threats, security advisories, procurement pressure, and hosting liability rules as cheaper Chinese open-weight models gain enterprise traction.

HN focused less on the launch framing and more on the pressure Kimi K3 puts on model economics: a 2.8T open model with a 1M-token context is expensive, capable, and hard to ignore.

Copilot now has its first selectable open-weight model. GitHub says Kimi K2.7 Code starts in VS Code for Pro tiers, with Business and Enterprise admins required to enable it by policy.

A LocalLLaMA user built a 768GB RAM system using discontinued Intel Optane Persistent Memory from the secondhand market, running the 1-trillion-parameter Kimi K2.5 model locally at over 4 tokens per second.

Chinese AI startup Moonshot AI raised $2 billion led by Meituan at a $20 billion valuation, bringing total capital to $3.9 billion over six months — making it China's most heavily funded LLM startup.

Why it matters: Moonshot is turning “agent swarm” from a demo phrase into an execution claim with real scale numbers. The Kimi post says one run can coordinate 300 sub-agents across 4,000 steps and return 100-plus files instead of chat transcripts.

HN read Kimi K2.6 as a test of whether open-weight coding agents can last through real engineering work. The 12-hour and 13-hour coding cases drew attention, while commenters immediately pressed on speed, provider accuracy, and benchmark realism.

LocalLLaMA cared about this eval post because it mixed leaderboard data with lived coding-agent pain: Opus 4.7 scored well, but the author says it felt worse in real use.

Cloudflare says Workers AI has made Kimi K2.5 3x faster for agent workloads. The technical change pushed p90 time per token from roughly 100 ms to 20-30 ms and raised peak input-token cache hit ratios from 60% to 80% with heavy internal users.

A Reddit thread surfaced Kimi's AttnRes paper, which argues that fixed residual accumulation in PreNorm LLMs dilutes deeper layers. The proposed attention-based residual path and its block variant aim to keep the gains without exploding memory cost.