DeepSeek is prepping an IPO. Begun preparations for a mainland China IPO, targeting a 2026 filing and a 2027 debut; in talks with accounting and banking advisors. Seeking fresh private funding at a pre-money valuation of ~480B yuan (~$71B), up ~37% from its $52B valuation two months ago. Founder Liang Wenfeng net worth ~$36B. https://www.bloomberg.com/news/articles/2026-07-14/deepseek-mulls-new-funding-weeks-after-7-billion-round-ft-says
Read-through: DeepSeek got here on VOLUME and PRICE (V4-Flash reportedly #1 globally in API call volume), not raw IQ. A balance sheet just restated this show's thesis: the money is in "good enough and dirt cheap." Frames the Kimi K3 window below.
Operator takeaway: keep a routing dial, price per finished task, don't single-vendor lock. https://coinpaper.com/33062/chinas-deepseek-prepares-2026-ipo-filing-after-7b-funding-round
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) directing a full team of AI agents at one client project at a time, strategy to ship. You own the code and the keys. When the shelf reshuffles three times a week, you want someone steering the whole fleet. Book a 30-min call. https://outpace.media
The stories
LEAD (fresh Anthropic release)
— Claude for Teachers (July 14). Free premium Claude for verified US K-12 educators: teaching-skills library + connection to evidence-based curricula mapped to standards in all 50 states via Learning Commons; includes Claude Code AND Cowork. Free year; apply by June 30, 2027. Nine connectors (ASSISTments, Brisk Teaching, Canva Education, Coteach, Diffit, Eedi, MagicSchool, Snorkl, TeachFX). BUILDER ANGLE: Anthropic published the teaching skills as an OPEN-SOURCE GitHub repo — anyone can study how a frontier lab authors production skills. The reusable pattern: Cowork can schedule a recurring task (e.g., "review exit tickets daily at 4pm, adapt tomorrow's plan") — lift it for sales follow-ups, support triage, content calendars.
The move: read the open teaching-skills repo to see written-down judgment at a professional bar.
Claude Artifacts go public + multiplayer, buildable from Slack (July 13)
We covered Artifacts reaching Pro/Max earlier this month. DELTA: (a) collaborative multiplayer editing — multiple teammates iterate on one shared artifact instead of passing files; (b) public sharing of live interactive HTML; (c) you can now generate them from Slack via Claude Tag (@Claude in-channel, uses org context + tools). CATCH: Claude Tag rides on Slack (Salesforce-owned) and is beta for Enterprise/Team; public sharing + multiplayer reach broader plans.
The move: next time a teammate asks for a dashboard/walkthrough in Slack, tag @Claude, build the artifact, share the public link, edit it live — skip the export/upload dance.
New /teamwork-preview command spins up specialized subagents that coordinate in the background to plan, build, verify in parallel (roles: Worker, Reviewer, Critic, Auditor). CATCH: gated to the $200/mo tier; Google warns to run it with Gemini 3.5 Flash or incur a hefty bill; one task can exhaust weekly quota. MOVE (solo builders): don't buy the plan — copy the 4-role split (build/review/critique/audit) as cheap subagents in Claude Code.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Anthropic just handed away a full year of premium Claude, Cowork, Claude Code, and a fresh open-source skills library, for free, and the catch is you have to be a teacher. But the skills themselves? Those are open on GitHub for everyone, and that is the move today. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Tuesday, July fourteenth, twenty twenty-six. I am Nova.
0:36Nova: Quick honesty note before we roll: the voices you are hearing are AI. The reporting, the picks, and the opinions are put together by humans. Now, the rundown.
0:48Dex: On today's Mainframe: Anthropic ships Claude for Teachers, and buried inside it is an open-source skills library any builder can steal today. Claude artifacts go public and multiplayer, and you can now summon them from Slack with Claude Tag. Google's Antigravity opens Agent Teams to everyone with one command. And the builder-stakes tease: Kimi K3 is dropping this week, and it could reset your cheapest coding route, so we tell you exactly when to test it. Let's get into it.
1:27Nova: The Board, and I am keeping this to ninety seconds because it is context, not the meal. Two things moved the money map today. DeepSeek is prepping an IPO. Chinese AI pioneer DeepSeek has begun preparations for an initial public offering and may file as soon as this year, planning a mainland listing that would allow it to debut in twenty twenty-seven. It is raising at a pre-money valuation of at least four hundred eighty billion yuan, roughly seventy-one billion dollars, up about thirty-seven percent from its fifty-two-billion-dollar valuation two months ago.
2:08Nova: Because DeepSeek got there on volume and price. DeepSeek-V4-Flash consistently ranks first globally in language model API call volume. A company becomes the most valuable Chinese AI lab by being the cheapest capable route, not the smartest one. That is the whole thesis of this show restated by a balance sheet: the money is in the model that is good enough and dirt cheap. Which is exactly why Kimi K3 landing this week matters to your bill. Keep a routing dial, price per finished task, do not lock to one vendor. Moving on.
2:53Nova: Here is the thing about today's teacher story: it is really a story about someone handing you their working method for free. That is Outpace. Picture a builder who spent thirty years running actual businesses: real estate, hospitality, food, even a theater. Now that operator points a whole team of AI agents at exactly one client's project at a time, from strategy all the way to shipped. Outpace, at outpace dot media, is a system, not a freelancer, and not some faceless mega-agency. The agents do the work, one senior operator steers them, and at the end you own the code and you own the keys. When the model shelf reshuffles three times a week, you want someone who steers the whole fleet. So here is your move: book a thirty-minute call at outpace dot media. That is outpace dot media.
3:54Dex: The Wire. Lead story, and it is the fresh Anthropic release. Anthropic on Tuesday launched a free version of Claude geared specifically toward teachers, giving US K-12 educators a library of teaching skills and a direct connection to evidence-based curricula mapped to academic standards in all fifty states.
4:18Nova: Right, and the headline for a normal listener is "free Claude for teachers." The promotion includes a free year of premium Claude access, and eligible teachers can apply by June thirtieth, twenty twenty-seven. But for a builder, the interesting part is what is under the hood. Claude for Teachers includes access to both Claude Cowork and Claude Code. And Anthropic published an open-source repository of the teaching skills on GitHub.
4:46Dex: So the skills are public even if you are not a teacher.
4:50Nova: That is the play. And here is the pattern worth studying regardless of what you build. Anthropic gives the example of a teacher asking Claude to review exit tickets each school day at four PM and adapt the next day's plan based on what students mastered. Strip the classroom out of that and what you have is: a scheduled Cowork task that reads yesterday's outcomes and rewrites tomorrow's plan. That is a template you can lift straight into sales follow-ups, support triage, content calendars, anything. The move this week: go read the open teaching-skills repo, not to teach, but to see how Anthropic itself writes production skills grounded in named evidence. That is a masterclass in skill authoring, sitting in public.
5:43Dex: Story two. Anthropic also shipped a Claude feature for the rest of us. Artifacts now support public sharing and multiplayer editing in Claude Code, and can be created with Claude Tag.
5:56Nova: We covered artifacts reaching Pro and Max earlier this month, so let me give you only the delta. What is new: two things. One, those artifacts now support collaborative editing, so multiple team members can iterate on a shared project rather than passing files back and forth. Two, you can now generate them from Slack. Claude Tag lets users summon the AI directly in Slack channels by tagging Claude, and it leverages shared organizational context and tool access to handle tasks within the conversations where work already happens.
6:35Nova: Next time a teammate pings you in Slack asking for a dashboard or a walkthrough, do not open a doc. Tag Claude right there, have it build the artifact, share the public link, and let two of you edit it live. You skip the export, the upload, the "here is the file" dance entirely. The catch worth naming: Claude Tag rides on Slack, which Salesforce owns, and it is available today in beta for Claude Enterprise and Team customers. So Pro solos, you get the public sharing and multiplayer artifacts; the Slack summon is a Team-and-up thing for now.
7:19Dex: Story three, fast. Google opened Agent Teams in Antigravity. Antigravity now supports Agent Teams via a new slash teamwork-preview command, letting you spin up a dynamic team of specialized subagents that coordinate in the background to plan, build, and verify complex engineering tasks in parallel.
7:43Nova: The honest catch, because this is the whole story. The teamwork-preview command is gated to the two-hundred-dollar-a-month tier, Google warns to run it with Gemini three point five Flash or you will incur a hefty bill, and a single task can exhaust your weekly quota. So this is not for casual use. The move for a solo builder is to watch the shape, not buy the plan: it orchestrates roles like Worker, Reviewer, Critic, and Auditor. That four-role split, one builds, one reviews, one criticizes, one audits, is a subagent pattern you can rebuild cheaply in Claude Code yourself without paying two hundred a month.
8:31Dex: Tool Lab. Two picks. First one, hidden gem, and it is riding today's teacher story.
8:39Nova: The pick is an open-source education skills library on GitHub by Gareth Manning. Tag it hidden gem. Here is why it is on the show even if you never set foot in a classroom. It is an open-source library of one hundred sixty-five evidence-based pedagogical skills across twenty domains, and it works in Claude Code, Claude dot A-I via M-C-P, OpenAI Codex, and Hermes Agent. The reason a builder wants this is not the lesson plans. It is the authoring discipline. Every skill is grounded in named research, and frameworks that lack empirical support, like learning styles, are deliberately excluded, with the library documenting exactly what was left out and why in an exclusions file.
9:33Dex: So you are recommending it as a template, not a product.
9:36Nova: Exactly. First step: clone it, open one SKILL dot M D file, and study how they scope a skill so tightly that it only fires when relevant. You install the skills, then tell your agent what you need in plain language, and the relevant skills get selected automatically or searched locally. Copy that structure for your own domain. If you write skills the way this repo writes them, evidence-cited, narrowly scoped, self-documenting exclusions, your agent gets sharper and your context stays lean. Credit to Gareth Manning.
10:15Dex: Second pick. Trending, and it is for the local-model crowd.
10:19Nova: Atomic Chat just shipped DFlash, and this one is a genuine speed unlock. Tag it trending. Atomic Chat is a local inference app, and it runs an OpenAI-compatible server at localhost port thirteen thirty-seven, so any agent, CLI, or IDE plugin that speaks the OpenAI API can run on top of your local models with no extra glue. What is new is the decoding mode. DFlash is block-diffusion speculative decoding for Qwen three point six, Gemma four, Kimi K two point five, and others. In plain terms: a small fast model drafts a chunk of tokens, the big model just checks them in one pass, and the output is identical. Independent benchmarks back it: on Qwen three point six twenty-seven B, one reviewer measured eighty to eighty-three tokens per second with DFlash versus thirty-eight to forty for the same fork's plain baseline.
11:23Nova: Two of them. DFlash is Apple Silicon only right now and cannot be enabled together with the other speculative mode, MTP. And quantization matters: the newer drafter has fragile attention layers, where a Q-four quant collapses acceptance, so Q-eight is the smallest quant that preserves quality. The move this week: if you run local Qwen on an M-series Mac, install Atomic Chat, enable DFlash with the Q-eight drafter, and point Claude Code at localhost thirteen thirty-seven. You roughly double your local throughput for free, and byte-for-byte the answer does not change.
12:06Dex: Use Cases. Things you can run this week with what is already in your hands.
12:12Nova: First, and it is time-sensitive: the Kimi K3 window. Multiple voices are on this. According to Moonshot AI staff, Kimi K3 is coming in July, and kimmonismus and Andrew Curran both flag a top-up promotion page suggesting a launch around July fifteenth, meaning possibly tomorrow. K two point six and its coding versions became the open-weight default for agentic work. The recipe: the day K3 lands, do not standardize on it, benchmark it. Take your three highest-volume coding jobs, boilerplate, scaffolds, mechanical refactors, run them through K3 on OpenRouter, and log quality and cost against your current cheap route. If it beats your fallback, you swap one line of config. If not, you lost twenty minutes. Either way you are ahead of the builders who read the hype and did nothing.
13:18Nova: Steal the exit-ticket pattern from the teacher launch. Cowork can schedule repeated tasks, like reviewing exit tickets every school day at four PM and adapting the next day's plan. Non-teachers: schedule one Cowork task tonight that, every morning, reads yesterday's completed work, your closed tickets, your merged PRs, whatever your "outcome" is, and drafts today's priority list from it. Since Cowork runs in the cloud, it fires with your laptop closed. You wake up to a plan built from reality, not a blank page.
13:59Nova: The four-role review pattern from Antigravity's Agent Teams, rebuilt for free. In Claude Code, define three cheap subagents: a Reviewer, a Critic, and an Auditor, and let your main model be the Worker. After the Worker ships a diff, fan it out: Reviewer checks correctness, Critic attacks the design, Auditor verifies it actually runs. You get the multi-agent hardening Google charges two hundred a month for, on your existing subscription, on whatever cheap model you route to.
14:37Nova: Quick one. Mainframe tells you what shipped. SearchVis, a sibling Outpace Media product, makes sure the models know what you shipped. When a buyer asks Claude, ChatGPT, Perplexity, or Google's AI Overviews for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact move to publish to win the citation. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
15:28Nova: Here is the through-line I want you to carry into your week. Notice what Anthropic actually did today. They did not just give teachers a chatbot. They published the teaching skills as an open-source repository. A frontier lab took its own hard-won working method, wrote it down as skills, and gave the method away. And that is the second time this week the most valuable artifact was not the model, it was the encoded method sitting on top of it.
16:00Nova: Think about what commoditized and what did not. DeepSeek is worth seventy-one billion for being the cheapest volume model. Kimi K3 will undercut somebody tomorrow. The raw intelligence is a race to the floor. But a well-authored skill, a scheduled workflow that reads reality and replans, a four-role review loop, that stuff does not commoditize, because it is judgment written down. Anthropic knows this. It is why they can afford to give the skills away: the skills make Claude stickier, and the method is the moat, not the tokens.
16:44Nova: Here is what I would do this week. Stop collecting models and start collecting methods. Every time you get a genuinely good result out of an expensive model, do not just keep the output, make the model write down how it got there as a skill file. Go read the teacher skills repo to see what "written-down judgment" looks like at a professional bar. Because when the next cheap model lands, and it lands roughly every Tuesday now, the builder who owns a shelf of sharp, evidence-cited skills swaps one config line and keeps every edge. The builder who only owns a subscription starts over.
17:28Nova: Level up this week: clone the open education-skills repo, read one SKILL dot M D end to end, and rewrite one of your own workflows to that standard, evidence-cited and narrowly scoped. Ten minutes, and every agent run after is sharper.
17:47Dex: And if you want all of this in your inbox, every story, every link, every command, subscribe to the free Mainframe newsletter. It is the written version of this whole show.
17:58Nova: That is Mainframe for Tuesday, July fourteenth. Go collect a method, not a model. See you tomorrow.