Kimi K3 Released: 2.8T Open Weights
Latent Space · rss · 2026-07-17
This AI News issue heavily focuses on Moonshot's release of Kimi K3:
Key Takeaways
- Kimi K3 is positioned as a frontier-class open-weights model, marking a massive release in the open-source community.
- Official specs include 2.8T total parameters, a 1M context window, native multimodal input, text output, with weights promised to be open-sourced before July 27.
- Pricing is noted as close to Sonnet: $3 / 1M input tokens and $15 / 1M output tokens, with substantial discounts for cached inputs.
Technical Highlights
- K3 utilizes Kimi Delta Attention (KDA), which officially claims up to 6.3x decoding acceleration for million-token scenarios.
- It introduces Attention Residuals (AttnRes), reportedly boosting training efficiency by roughly 25% with under 2% additional cost.
- The community discovered further details in the tech blog: potential use of LatentMoE / Stable LatentMoE with sparsely activated experts, along with mentions of Muon, load balancing, and a new activation function, SiTU.
Benchmarks & Rankings
- Arena's early results after integrating K3 are highly impressive:
- Ranked No. 1 in Frontend Code Arena with a reported score of 1679, surpassing Claude Fable 5.
- Reached No. 9 in Text Arena, making the top ten in sub-categories like creative writing, coding, and instruction following.
- Artificial Analysis's independent evaluation placed it at an AA Intelligence Index of 57, overall close to Opus 4.8 and GPT-5.5, though still behind Fable 5 and GPT-5.6 Sol.
- They also provided task cost and token consumption data: the average task cost for the full Intelligence Index run was about $0.94, outputting 21% fewer tokens than K2.6.
Inference & Engineering
- Early online API services observed speeds of around 26-28 tok/s, leading to speculation that speculative decoding might not be fully enabled yet.
- Moonshot's blog suggests deploying on 64+ accelerator supernode configurations for optimal inference efficiency.
- vLLM confirmed that Moonshot contributed the KDA prefix caching implementation upstream, indicating that this release also drove adaptations in serving and runtimes, beyond just the model itself.
Controversies & Caveats
- The author of ProgramBench cautioned that certain statistical metrics might overestimate practical usability because non-recommended metrics were used.
- The text explicitly notes that despite K3's strength, a gap remains in user experience compared to Claude Fable 5 and GPT-5.6 Sol.
More from coding & agent
- BELL: Open-Source Stack Combines Browserbase, Eve, iMessage and Stripe Link for Personal Agents — jeff_weinstein · 2026-10-27
- Two engineers rebuilt a FedEx supplier's billing core in 3.5 months vs an estimated 7-8 — alex_verem · 2026-10-27
- GPT-6.1 Terra gains ~20 points on Terminal-Bench 4.0, rivalling GPT-6.1 Sol — ArtificialAnlys · 2026-10-27
- GPT-6.1 Terra scores 48 on AA Coding Agent Index at one-third the cost of GPT-5.6 Terra — ArtificialAnlys · 2026-10-27
- Shift-1 router cuts coding agent spend up to 55% while keeping SWE-bench accuracy — ypatil125 · 2026-10-27
- Non-coder vibe-codes a Godot JRPG with Claude Code in two weeks, music and art synced — ADHDavid · 2026-10-27