AI Industry Daily Radar - June 20, 2026
The most important AI industry developments from the last 24 hours.
Executive Summary
- Mistral releases Medium 3.5, a 128B dense open-weights model with 256k context window aimed at coding agents and long-context work, reporting 77.6% on SWE-bench Verified.
- Google DeepMind suffers two historic talent losses in one week: Transformer co-inventor Noam Shazeer joins OpenAI, and Nobel laureate John Jumper (AlphaFold co-creator) joins Anthropic.
- Cerebras unveils Kimi K2.6 running on its Wafer-Scale Engine at a world-record 1,000 tokens per second for trillion-parameter models.
- Perplexity launches Brain, a self-improving memory system for AI agents that learns from past work to boost accuracy and efficiency.
- AI agent infrastructure funding surges: NewCore raises $66M for agent identity security, while Genspark closes a $100M Series B extension at a $2.6B valuation.
Top Stories
1. Mistral Releases Medium 3.5: 128B Open-Weights Coding Model with 256K Context Window
Summary
Mistral AI has released Mistral Medium 3.5 in public preview, a 128-billion-parameter dense model released as open weights under a modified MIT license. The model is explicitly designed for coding agents, tool use, and long-context work, reporting 77.6% on SWE-bench Verified and 91.4 on the τ³-Telecom agent benchmark.
The release marks a strategic move in the open-weights space: a frontier-sized model that teams can self-host on as few as four GPUs, according to Mistral. The 256k context window is particularly significant for codebase-level agent tasks, where agents need to maintain context across large repositories, logs, and multi-file refactors. Mistral is also tying Medium 3.5 into Vibe, its agent product for long-running coding work, which now supports remote cloud sessions and CLI-to-cloud "teleport" functionality — enabling developers to start coding tasks locally and let them run in the cloud.
Alongside the model release, Mistral is previewing a Work mode in Le Chat for multi-step research, analysis, and cross-tool tasks, following the industry trend of chat interfaces evolving into agent orchestration platforms.
Source
2. Nobel Laureate John Jumper Leaves Google DeepMind for Anthropic
Summary
Nobel Prize-winning AI researcher John Jumper announced on Friday, June 19, that he is leaving Google DeepMind after nearly nine years to join Anthropic. Jumper co-created AlphaFold alongside DeepMind CEO Demis Hassabis, a breakthrough AI system that predicted over 200 million protein structures and earned both researchers the 2024 Nobel Prize in Chemistry.
Jumper's departure is the second major blow to Google DeepMind within a single week. Just days earlier, Noam Shazeer — co-inventor of the Transformer architecture and co-lead of Google's Gemini project — announced he would leave to join OpenAI. The back-to-back exits underscore the intensifying talent war between Big Tech and AI startups. As D.A. Davidson analyst Gil Luria noted, OpenAI and Anthropic hold an advantage over large companies like Google "because they can promise less bureaucracy and a more focused effort on pursuing Superintelligence."
Jumper joins Anthropic at a strategically significant moment. The company is locked in a high-stakes regulatory battle with the U.S. government over export controls on its Fable 5 and Mythos 5 models, and is hosting a science-focused event on June 30. His move signals that Anthropic is making a serious play in AI-for-science — a domain where DeepMind has historically led.
Source
3. Transformer Co-Inventor Noam Shazeer Joins OpenAI from Google
Summary
Noam Shazeer, co-inventor of the Transformer architecture and co-lead of Google's Gemini AI models, has left the company to join OpenAI, marking the highest-profile defection from Google DeepMind to OpenAI in the history of the frontier-model race. Shazeer confirmed the move on X on June 18.
The significance of this move is hard to overstate. Shazeer's name is literally embedded in the architecture behind every major modern AI system — the "T" in GPT. He returned to Google in late 2024 through the controversial $2.7 billion Character.AI acquisition, resuming leadership of the Gemini project. His departure now, to join the company that built the other three letters in GPT, represents what analysts describe as the clearest signal yet that the best frontier researchers believe the frontier is being set at OpenAI.
Shazeer's move follows a pattern: three other Gemini-team senior researchers have left for OpenAI, Anthropic, or xAI in the last six months. The talent drain from Google's AI labs is becoming a structural concern, with capital and compute increasingly abundant while top-tier research talent remains the binding constraint on progress.
Source
https://singularity.kiwi/noam-shazeer-joins-openai-2026/
4. Perplexity Launches Brain: A Self-Improving Memory System for AI Agents
Summary
Perplexity has announced Brain, a new memory system for its AI agent "Computer" that continuously learns from completed work to improve performance over time. Unlike traditional AI memory systems that store user preferences, Brain focuses exclusively on work context — building a knowledge graph of tasks, approaches, corrections, and outcomes from past sessions.
The system operates on a self-improving feedback loop. As Computer gains experience with projects, it learns which sources produce reliable results, which approaches are effective, and which mistakes to avoid. Early internal testing showed measurable gains: answer accuracy improved by 25% on previously encountered tasks, information recall increased by 16%, and task costs dropped by approximately 13%.
Brain is available in Research Preview for Perplexity Max and Enterprise Max subscribers. The launch signals a broader industry shift from stateless AI agents toward agents that accumulate knowledge and improve with use — a critical capability for production deployment at scale.
Source
https://mpost.io/perplexity-introduces-brain-signaling-a-shift-toward-self-improving-ai-agents/
5. Cerebras Unveils Kimi K2.6: 1,000 Tokens/Second on Trillion-Parameter Models
Summary
Cerebras has announced that it is serving Kimi K2.6, a trillion-parameter open-weight model, on its Wafer-Scale Engine at a world-record speed of 1,000 tokens per second. The model is widely recognized as a leading open-weight model for coding and agentic work, reportedly outperforming Claude Opus 4.6 and matching GPT-5.4 on benchmarks including SWE-bench Pro and DeepSearchQA.
The breakthrough is architectural: Cerebras stores Kimi K2.6 in 4-bit weights while computing at 16-bit floating point, distributing weights across multiple wafers with an on-wafer network fabric that provides over 200 times the bandwidth of NVLink on NVL72. Combined with custom kernels and speculative decoding, this enables trillion-parameter MoE model inference at speeds previously thought impossible.
Cerebras is making Kimi K2.6 available for enterprise trials, targeting customers running agentic coding, deep research, and production AI workloads where inference speed is the critical bottleneck. The 2.6 release extends Kimi's capabilities to full-stack workflows including authentication, database operations, and long-horizon agent execution.
Source
6. Genspark Raises $100M at $2.6B Valuation for Agentic AI Workspace
Summary
Palo Alto-based Genspark has closed a $100 million Series B extension at a $2.6 billion post-money valuation, just three months after its previous $1.6 billion valuation — a 63% step-up in a single quarter. The round was backed by Sozo Ventures, Korea Mirae Asset, and UpHonest Capital, bringing the company's total Series B funding to $485 million and cumulative funding to over $645 million.
Alongside the funding, Genspark launched AgentBase, an AI-native database and dashboard product, and announced a strategic partnership with Microsoft to embed its AI agents into the Microsoft 365 ecosystem. The company — which orchestrates over 70 AI models — has reached more than 2 million monthly active users and charges $30 per user per month for team plans.
Genspark's rapid valuation growth reflects the market's appetite for platforms that turn business goals into finished deliverables through agentic AI — a category that is increasingly attracting premium valuations as enterprises move from experimenting with AI to deploying autonomous agents in production workflows.
Source
7. NewCore Raises $66M to Build Identity Security for AI Agents
Summary
NewCore has emerged from stealth with $66 million in total funding to build an identity platform for humans, machines, and AI agents. The Tel Aviv and San Francisco-based startup, which includes a $16 million pre-seed and an expanded seed round, reached a reported $300 million valuation.
NewCore is addressing a problem that will only grow as enterprises deploy autonomous agents: AI agents currently lack proper identity management. They often use borrowed credentials with no permissions boundaries, audit logs, or revocation controls. NewCore's platform aims to give agents work identities with the same security properties as employee accounts — authentication, authorization, logging, and the ability to revoke access instantly.
CEO Zohar Alon previously founded Dome9 (acquired by Check Point), and his co-founders include a former Unit 8200 research leader and a former T-Mobile US CIO. The round was backed by Cyberstarts, Index Ventures, and Evolution Equity Partners. The investment signals that agent identity is becoming a critical enterprise battleground as organizations move from experimenting with agents to deploying them in production systems with real access to data and infrastructure.
Source
https://memeburn.com/newcore-raises-66m-to-give-ai-agents-work-identities-in-2026/
Industry Trends
Trend 1: The Great AI Talent Migration Accelerates
Two of the most consequential AI researchers in the world — John Jumper and Noam Shazeer — left Google DeepMind within the same week for Anthropic and OpenAI respectively. This is not an isolated pattern: multiple senior Gemini researchers have departed Google in the past six months. The structural dynamic is clear: startups like Anthropic and OpenAI can offer elite researchers more focus, less bureaucracy, and a direct path to superintelligence, while Big Tech labs struggle with slower decision-making and competing priorities. As capital and compute become increasingly commoditized, the quality of the 30-to-50 people deciding what to build next is becoming the binding constraint on frontier progress.
Trend 2: Open-Weights Models Are Closing the Frontier Gap for Coding and Agents
Mistral Medium 3.5 (128B dense, 77.6% SWE-bench), Kimi K2.6 (1T MoE, matching GPT-5.4), and GLM-5.2 (top frontend coding model) all launched or became widely available in the past week. These are not academic exercises — they are explicitly designed for production coding agents, tool use, and long-context work, with commercial-friendly licenses and self-hosting paths. The open-weights ecosystem is reaching a tipping point where teams with sensitive codebases and regulated environments have credible alternatives to closed API providers.
Trend 3: AI Agent Infrastructure Becomes a Standalone Investment Category
This week's funding announcements — NewCore ($66M for agent identity), Genspark ($100M for agent workspace), NeuralTrust ($20M for agent security), and Convey ($38M for AI teammate orchestration) — reveal a new investment thesis: the infrastructure layer that makes agents safe, governable, and productive in enterprise environments is itself a multi-billion-dollar opportunity. Identity, memory, observability, and security for AI agents are no longer afterthoughts. They are becoming independent product categories attracting significant venture capital.
Featured AI Products
Mistral Medium 3.5
- A 128B dense open-weights model with a 256k context window, designed for coding agents, tool use, and long-context work. Reports 77.6% on SWE-bench Verified and self-hostable on as few as four GPUs.
- It represents the strongest open-weights competitor yet aimed at the agentic coding workflow, offering teams a credible self-hosting path for sensitive production workloads.
- Official URL: https://mistral.ai/news/mistral-medium-3-5
Perplexity Brain
- A self-improving memory system for the Perplexity Computer agent that builds a context graph of completed tasks, learns from corrections, and improves accuracy by 25% on previously encountered work.
- Brain represents a shift from stateless agents to agents that accumulate knowledge — a critical capability for production-grade agent deployment.
- Official URL: https://www.perplexity.ai/
Cerebras Kimi K2.6 (via Cerebras Wafer-Scale Engine)
- A trillion-parameter open-weight model served at 1,000 tokens/second on Cerebras custom hardware, matching GPT-5.4 on coding and agentic benchmarks while being available for enterprise self-hosting.
- The combination of frontier model performance, open weights, and world-record inference speed makes this a compelling option for latency-sensitive production agent workloads.
- Official URL: https://cerebras.ai/
Key Takeaways
- The AI talent war has entered a new phase: Google DeepMind lost two of its most iconic researchers in a single week to Anthropic and OpenAI, signaling that the frontier is increasingly being set by startups, not Big Tech.
- Open-weights models are now credible alternatives to closed APIs for coding and agentic workloads, with Mistral Medium 3.5 and Kimi K2.6 both demonstrating frontier-competitive performance with self-hosting paths.
- AI agent infrastructure — identity, memory, security, orchestration — has emerged as a standalone venture category, with over $300 million in new funding announced this week across multiple startups.
- The Perplexity Brain launch points to the next frontier for agents: persistent memory and continuous self-improvement, moving beyond stateless question-answering toward systems that get better with use.
- Cerebras's 1,000 tokens/second milestone on trillion-parameter models demonstrates that hardware innovation is keeping pace with model scaling, potentially removing inference speed as a bottleneck for production agent deployment.