Kimi K3 now available for public download
Moonshot's 2.8-trillion-parameter model is downloadable directly, following its initial launch.
AI Daily Recap
20 articles · 6 categories
The finishable daily brief
What happened in AI — Jul 27, 2026
Monday, Jul 27, 2026
20 articles · 6 categories
read top to bottom · then stop
In 30 seconds
- Moonshot's Kimi K3 (2.8T params) went public with day-0 vLLM and Modal serving support, though a HackerNoon portability audit disputes its launch-week benchmark scores.
- Belay and a new OpenCode static verifier both launched to catch unsafe agent tool calls before they execute.
- philschmid's EvoCode-Bench tests coding agents across 227 sequential rounds, showing single-turn scores overstate reliability.
- A coding agent refactored a 750,000-line codebase over three days with zero human code review, running 31 verification passes to catch 201 errors.
- Chinese models now handle nearly a third of enterprise LLM tokens at one-tenth the cost of US rivals, per a widely syndicated report, even as DeepSeek paused a new funding round.
- Anthropic named Cognizant a Global Premier Partner after training over 30,000 of its associates on Claude.
Moonshot's 2.8-trillion-parameter Kimi K3 became publicly downloadable today, and vLLM and Modal already have day-0 serving support live — vLLM ships hybrid KDA prefix caching and DSpark speculative decoding, Modal pairs the model with a custom-trained DFlash speculator. A HackerNoon portability audit pushed back on the launch-week hype, arguing K3's headline benchmark scores don't hold up once you account for how it was evaluated.
Two new tools target unsafe agent tool calls directly: Belay adds a local firewall for coding agents, and a static verifier built on the "Guardians of the Agents" formal-verification paper checks OpenCode's tool calls before they run. Separately, a widely syndicated report put Chinese models at nearly a third of enterprise LLM tokens at one-tenth the cost of US rivals, even as DeepSeek paused a new funding round.
Kimi K3 Goes Public, and the Skepticism Starts Immediately 4 items
Moonshot's 2.8-trillion-parameter Kimi K3 is now publicly downloadable with day-0 vLLM and Modal serving support, but a portability audit is already challenging its launch-week benchmark claims.
A portability audit questions Kimi K3's launch-week scores
HackerNoon argues K3's headline benchmark numbers don't hold up once you account for how the model was evaluated.
vLLM ships day-0 Kimi K3 serving support
vLLM's Kimi K3 support includes hybrid KDA prefix caching and DSpark speculative decoding, tuned for both NVIDIA and AMD GPUs.
Kimi K3 lands on Modal with a custom speculator
Modal's Kimi K3 deployment pairs the 2.8T-parameter model with a custom-trained DFlash speculative-decoding model.
Coding Agents Take On Bigger, Less-Supervised Work 3 items
A case study shows a coding agent refactoring a 750,000-line codebase over three days with no human code review, while GitHub frames its own harness, not any single model, as the unit that makes agentic workflows reliable.
A coding agent refactors a 750k-LOC app with zero human code review
The agent rebuilt a core system invariant in three days, running 31 verification passes and correcting 201 errors before shipping with no human review.
GitHub: the harness, not the model, is what makes agentic coding reliable
GitHub's Copilot workflow guidance argues a consistent prototype-plan-implement-review harness matters more than chasing each new model release.
An open-source resume-screening agent pipeline
Hyre chains an LLM pipeline (PDF-to-markdown, layout analysis, data extraction, scoring) to turn a resume into a numeric evaluation.
New Tools Aim to Catch Unsafe Agent Actions Before They Run 4 items
Two new builder tools, a local firewall and a static verifier, check agent tool calls before execution, while a new benchmark shows single-turn evals miss the multi-turn regressions that actually break agents.
Belay: a local firewall for AI coding agents
Belay intercepts and filters coding-agent actions locally, before they reach the file system or shell.
A static verifier blocks unsafe tool calls in OpenCode
Built on the "Guardians of the Agents" formal-verification paper, the plugin statically checks OpenCode's tool calls before execution.
EvoCode-Bench tests agents across 227 sequential rounds
Single-turn benchmark scores overstate reliability; the real failure mode is regressions across a persistent workspace, not missing features.
NVIDIA and industry partners launch the Open Secure AI Alliance
The alliance frames open-source security practices as infrastructure for AI safety, extending the model that secured cloud and telecom software.
Platform Teams Build Serving and Data Layers Purpose-Built for Agents 4 items
Netflix detailed its in-house Triton/vLLM serving stack, LangChain built a sub-second full-text search index over agent traces, and AWS pitched task-aware compression as a fix for RAG's document-scale ceiling.
Netflix details its in-house LLM serving platform
Netflix built its own serving layer on Triton and vLLM rather than relying on a third-party inference API.
LangChain's SmithDB adds full-text search over agent traces
SmithDB indexes large, deeply nested agent-trace JSON in object storage with a median 400ms search latency.
AWS pitches task-aware knowledge compression as RAG's next step
Traditional RAG breaks down on analytical tasks spanning hundreds of documents; AWS's TAKC pre-compresses knowledge bases into task-specific representations instead.
Why agentic AI needs its own API gateway pattern
InfoQ argues traditional API gateways assume deterministic services, an assumption agentic AI breaks, pushing engineering leaders toward AI gateways as a dedicated architecture seam.
China's Model Cost Advantage Keeps Widening, Even as DeepSeek Wobbles 3 items
A widely syndicated report put Chinese models at nearly a third of enterprise LLM tokens at one-tenth the cost of US rivals, even as DeepSeek paused a new funding round and US lawmakers weigh restrictions that could slow the catch-up.
Chinese models now handle nearly a third of enterprise tokens, at one-tenth the cost
A widely syndicated report puts Chinese open models at roughly 30% of enterprise LLM token volume, priced at about one-tenth of comparable US models.
DeepSeek pauses a new fundraising round
The pause follows leaked comments attributed to founder Liang Wenfeng that went viral days earlier.
Congress debates restrictions that could slow the US AI catch-up
Washington Examiner frames pending congressional measures as a risk to US competitiveness against China's AI pace.
Enterprise Adoption Signals 2 items
Anthropic named Cognizant a Global Premier Partner after training over 30,000 of its associates on Claude, and new OpenAI research tracks how ChatGPT is reshaping what workers actually do inside their roles.
Anthropic names Cognizant a Global Premier Partner
Cognizant has trained 30,000+ associates on Claude and is embedding it across its platforms under the expanded partnership.
OpenAI research tracks how ChatGPT is reshaping job tasks
The study finds ChatGPT users are taking on tasks that cross traditional role boundaries, not just automating existing ones.
You are caught up for this edition