I have very little idea of what K3 might deliver what hints we have are pretty intimidating.
Kimi K3 in July.
I have very little idea of what K3 might deliver what hints we have are pretty intimidating.
Kimi K3 in July.
Positive users express excitement about the rumored Kimi K3 launch due to its hinted 1T-parameter frontier capabilities and strong fundamentals, while negative users complain about limited access and insufficient scale.
No Digg Deeper questions have been answered for this story yet.
Tbh I also look forward more to K3, because I know that to expect from V4. But on fundamentals, sorry guys, I still think this is Whale's game to lose. i don't look at the benchmarks. I look at the Machine they're building, and it's hard to see others matching its throughput.
Not sure if what gets released in July is K3, but it's depressing for a fan of DeepSeek's work that I'm far more looking forward to the next Moonshot release than the full version of V4.

@teortaxesTex They hinted last time they have 1T parameter linear attention model and I still can not get over it.

@teortaxesTex @grok which hints do we have on kimi k3?

@teortaxesTex 这种未知感反倒让人更期待了

@teortaxesTex In casual use Kimi is the first Chinese model to break though normally frontier. Similar to how Opus 4+/Gpt 5.5 lost LLM strangeness in casual chats.

@bookwormengr @teortaxesTex 1T is not enough

**Hints on Kimi K3 (Moonshot AI):**
Mostly unconfirmed rumors + funding news as of late June 2026:
- Moonshot raised $500M Series C early 2026 specifically to fund K3 development + expand compute. - Leaks/rumors: Massive MoE model targeting **2.5T–4T** total params (to compete with top US frontier models). Some mention 1M context + “Kimi Linear attention.” - Expected **Q3 2026** (July–Sept). Community chatter picked up on a possible July drop after K2.6’s recent release. - K2.6 (and its agent/swarm infra) looks like scaffolding for something much bigger.
No official announcement or concrete details from Moonshot yet — just leaks, funding signals, and speculation.
What specifically are you curious about?

@teortaxesTex 确实基本面才是硬道理鲸鱼的格局不一样

@teortaxesTex I just wish they'd give us GPU poor non-clould chuds scaps more often

@teortaxesTex Iteration speed & RL scalability right?

@teortaxesTex A next-gen model will definitely learn from MiniMax's missteps. At the very least, it'll pack 1.6T parameters with heavy post-training for multimodal and coding.
MiniMax really missed the mark going with a 400B base for M3.