AI Briefing#kimi-k3
The latest AI and developer news about #kimi-k3, with the original source and a short summary.
The latest AI and developer news about #kimi-k3, with the original source and a short summary.
Self-hosting Kimi K3 costs roughly $0.25 per million tokens, significantly below the $1.13 API price, indicating healthy margins rather than unsustainable subsidies.
In 2,160 simulated outage scenarios, Claude Sonnet 5.5 and GPT-6.1 Sol emerged as leaders 1.7 times more often than Gemini 3.1 Pro, while agents listed first in the system prompt issued plans three times more frequently.
The deployment uses 4-bit quantization and pipeline parallelism to keep model weights static on-chip while streaming activations between wafers.
The $30/month subscription grants $80 in monthly credit and supports coding agents like fx, Pi, and Cline with Zero Data Retention enabled by default.
Pro teams can subscribe for $30 per month to receive $80 in shared credit for coding agents using GLM 5.3 and Kimi K3.
The models collaborated in a shared session to draft a five-point regulatory framework covering prior approval, liability, and cyber defense funding.
Fireworks AI has released 'Ember-1' as a research preview, an inference-optimized model based on Kimi K3 that reduces token usage by approximately 35–40% without quality loss in benchmarks and real-traffic tests.
Ember-1 is a research preview model built on Kimi K3 that reportedly generates 40% fewer tokens than its base model.
Fireworks Research releases Ember-1, a specialized model built on Kimi K3 that delivers comparable performance with 40% fewer tokens, now available as a Research Preview.
AWS updated key services including the GA launch of Kimi K3 on Bedrock, the release of Corretto 27, and performance improvements for EC2 T8i.
Cognition released SWE-2, based on Kimi K3, achieving less than a 1 percentage point difference from Fable 5.1 on the FrontierCode benchmark while reducing costs by 64%.
Redwood Research defined 'NLS depth' and distinguished the depth increase in Kimi-K3 using Gated DeltaNet from the overestimation in DeepSeek-V4-Pro, which does not use DeltaNet, due to Sinkhorn-Knopp operations.
vLLM v0.28.0 adds large-scale performance optimizations and inference capabilities for the Kimi-K3 and DeepSeek V4 models.
vLLM v0.28.0 introduces major performance optimizations and new inference features for Kimi-K3 and DeepSeek V4.
Harvey has post-trained the Kimi K3 model for long-running legal tasks.
AirLLM has been updated to enable Kimi K3 (2.8T) to run on 4GB VRAM via sparse MoE model streaming.
The emergence of open-weight models capable of offensive security research has made a defensive response urgent.
The average price per token on AI Gateway fell by 13.6% in July, and DeepSeek rose to second place in usage.
Moonshot AI joins the competition by releasing Kimi K3 as an open-weight model.
Unsloth has released the massive GGUF quantized files for Kimi K3.
Kimi K3 has been officially integrated into GitHub Copilot and is now available for use.
WASTE ran the 2.78T Kimi K3 via disk streaming on a 64GB MacBook Pro.
Moonshot released the high-performance open model Kimi K3 for free, shaking up the cost structure of sovereign AI.
WASTE ran the 2.78 trillion-parameter Kimi K3 on 64GB of memory.