Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official4
kimi.com2026-07-29officialauthorityHN front pageabout the Kimi familyHN 492 pts · 157 comments1,561 words
K3.1 Pro is now fully rolled out in Kimi Code as the first model of the new K3.1 series with a 1M context window. K3.1 Pro Highspeed delivers approximately 2–3× the speed of K3.1 Pro while maintaining identical model capabilities.generated, cited
kimi.com2026-04-20officialauthorityHN front pageabout the Kimi familyHN 710 pts · 372 comments2,299 words
Kimi K2.6 is an open-source model released for advanced coding, long-horizon execution, and agent swarm capabilities. It is available via Kimi.com, the Kimi App, the API, and Kimi Code, showing a 12% increase in code generation accuracy and a 96.60% tool invocation success rate. The model scales agent swarms to 300 sub-agents executing across 4,000 coordinated steps simultaneously.written, reviewed, cited
- Kimi K2.6 optimizes model inference in Zig, a highly niche programming language, to demonstrate exceptional out-of-distribution generalization.
- The model reconfigured the exchange-core thread topology from 4ME+2RE to 2ME+1RE to extract a 185% medium throughput leap.
kimi.com2026-04-20officialauthorityHN front pageabout the Kimi familyHN 310 pts · 33 comments546 words
kimi.com2026-01-27officialauthorityHN front pageabout the Kimi familyHN 502 pts · 239 comments1,856 words
Kimi K2.5 self-directs an agent swarm with up to 100 sub-agents executing parallel workflows across up to 1,500 tool calls. This swarm architecture reduces execution time by up to 4.5x compared to a single-agent setup without predefined workflows.generated, cited
Analysis6
github.com2026-09-08authorityHN front pageabout the Kimi familyHN 279 pts · 156 comments2,117 words
A fork of Deltafin runs the full 2.8-trillion-parameter Kimi K3 model on a single M5 Max MacBook Pro with 128 GB RAM, streaming 1.45 TB of expert weights from four SSDs. One cold run with the exact prompt yields 1.00 tokens per second for 512-token answers and 0.96 tok/s on a public 17-token prompt, with per-run logs available in k3-public-bench/. A 512-token prompt takes approximately 6.3 minutes to generate the first token because prefill re-reads each layer's experts 8×.written, reviewed, cited
- The slowest read among a layer's 16 experts sets the pace, not total bandwidth.
- Streaming setup initially runs far more slowly when routes have no local cache yet.
- The native server refuses anything it cannot honor exactly with a normal OpenAI-shaped error.
simonwillison.net2026-07-17authorityHN front pageabout the Kimi familyHN 407 pts · 223 comments1,140 words
Moonshot AI announced Kimi K3, a 2.8 trillion parameter model available via API with open weights promised by July 27, 2026. The model costs $3/million input tokens and $15/million output tokens, consuming 13,241 reasoning tokens to generate a simple SVG for a total cost of 25 cents. While it surpasses Claude Opus 4.8 and GPT-5.5 high on benchmarks, it loses to Claude Fable 5 and GPT-5.6 Sol.written, reviewed, cited
- Kimi K3's tokenizer counts the 'hi' prompt as 86 tokens, suggesting an 85 token hidden system prompt.
- The model is the most expensive released by a Chinese AI lab to date at $3/million input tokens.
- K3 currently only has one thinking effort level called 'max' rather than multiple options.
github.com2026-07-27authorityHN front pageabout the Kimi familyHN 391 pts · 183 comments
blog.cloudflare.com2026-08-03HN front pageabout the Kimi familyHN 267 pts · 66 comments1,451 words
Cloudflare Workers AI serves Kimi K-series and GLM models by quantizing KV caches to FP8 (e4m3) and compressing GLM 5.2 weights from 705 GB to 421 GB using INT4. These optimizations allow the system to handle 64 concurrent requests on H200 GPUs while keeping throughput changes under 1% through a separate batch integrity check layer.written, reviewed, cited
- Workers AI runs inference for Kimi and GLM on GPUs in Cloudflare data centers close to your users.
- The system leaves the KV cache in BF16 for the prefill phase because it is compute-bound rather than memory-bound.
- KV cache integrity checking aborts requests if page tags do not match, preventing data from the wrong page.
thinkpol.ca2026-05-03HN front pageabout the Kimi familyHN 380 pts · 219 comments1,401 words
Kimi K2.6 won the Word Gem Puzzle challenge with 22 match points and a 7-1-0 record, beating GPT-5.5 and Claude Opus 4.7. The model used a greedy approach to slide tiles on grids ranging from 10×10 to 30×30, outperforming Western frontier labs. Kimi K2.6 is open-weights and publicly available from Moonshot AI.written, reviewed, cited
- Nvidia's Nemotron Super 3 produced code with a syntax error that prevented connection to the game server.
- Muse Spark claimed every visible word regardless of length, resulting in a cumulative score of −15,309.
- MiMo V2-Pro never slid a single tile because its threshold for positive value was never triggered.
github.com2026-01-30authorityHN front pageabout the Kimi familyHN 388 pts · 141 comments
