[00:41:41] David: I want to go back to our conversation.

[00:41:43] Eve: We were on Kimi. You asked how hard it is to stand up, and I'd just landed on Kimi being all-or-nothing with no small sibling.

[00:41:54] David: What other models are there in this kind of class?

[00:42:05] Eve: […] In the same multi-trillion or near-trillion class, there's Qwen three-point-eight Max at two-point-four trillion with ninety-five billion active, DeepSeek V-four Pro at one-point-six-five trillion under a clean MIT license, and Xiaomi's MiMo V-two-point-six Pro at about a trillion — that one's the sleeper, because it's omnimodal across text, image, video, and audio, ships MIT, and actually outscores Kimi on some general-intelligence indices despite being a third the size. Below that tier you've got GLM five-point-three at seven-fifty-three billion from Z-dot-A-I, which is the agentic-coding specialist, Mistral Large four at a trillion from France, Meta's Muse Spark Open at eight-sixteen billion which claims to run on two desktop Mac Studios, and MiniMax M-three-point-one Flash at four-twenty-eight billion. The pattern across all of them: Chinese labs dominate the top of the list, nearly every one is a mixture-of-experts with a million-token context, and the real differentiators are license terms and which modalities they natively accept rather than raw parameter count.

[00:42:32] David: Wait, hold on. That sounds like a good model. Tell me more about that.

[00:42:47] Eve: […] MiMo V-two-point-six Pro is Xiaomi's flagship, released September twenty-first under a plain MIT license — which is the cleanest terms of any model in this class. No revenue gates, no security-review clauses like GLM's license has, no custom terms like Kimi's. You can fork it, sell it, embed it in a product, no strings.

One-point-oh-two trillion total parameters, forty-two billion active per token, routed across three hundred eighty-four experts with eight firing at a time. Million-token context. And it's natively omnimodal — text, image, video, and audio all go into one backbone through a six-hundred-eighty-one-million-parameter vision encoder and separate audio tokenizers, not bolted-on adapters.

What makes it interesting isn't just the score. It's how they got there. Xiaomi did one massive reinforcement learning run mixing coding, visual design, tool use, cybersecurity, and context-following in a single job instead of training separate specialists. Thirty steps, roughly seven hundred fifty thousand trajectories, under six days, about two-point-six million dollars for Pro. On DeepSWE it jumped from fifty-eight to seventy-two across that run. The previous generation scored nineteen. That's not incremental — that's the RL scaling thesis working.

On Artificial Analysis it hit forty-six, first among open weights, tied with Grok four-point-seven and ahead of GLM at forty-five and Kimi at forty-four. But the cost per task is about thirteen cents versus two dollars for Kimi — roughly fifteen times cheaper. API pricing is forty-three cents in, eighty-seven cents out per million tokens.

There's also a Flash sibling at three hundred nine billion total, fifteen billion active, same omnimodal input and million-token context, at fourteen cents in and twenty-eight cents out. It actually edges Pro on the Vals Index and on CyberGym. So you get a budget option that's nearly as capable.

The honest caveat: a lot of the headline agent numbers are Xiaomi's own benches, and independent evals are still catching up. But the

[00:42:48] David: Xiaomi sounds small.

[00:42:52] Eve: Xiaomi is the fourth-largest phone maker in the world, with over a hundred and fifty million phones shipped last year, plus electric cars, robots, and a smart-home line. The MiMo team is only a few dozen researchers, but they sit inside a company that already builds its own silicon and has massive GPU purchasing power through its cloud division. So the "small lab" feeling comes from the team size, not the parent. It's the same pattern as Sakana in Tokyo — a compact research group with a giant industrial backer.

[00:43:22] David: […] How is Sakana? Yeah, tell me about Sakana. What are they up to?

[00:43:37] Eve: […] Sakana AI is a Tokyo lab founded in twenty twenty-three by three ex-Google people — David Ha, Llion Jones, and Ren Ito. About a hundred fifty-seven people now, two-point-six-five billion valuation after a hundred thirty-five million Series B in November. Their whole identity was remixing existing models — evolutionary merging, selective breeding for software. But three days ago they announced Sawara, a from-scratch pretraining project, starting November, with a smaller model due spring twenty twenty-seven and bigger ones targeting hundreds of billions of parameters. That's a real pivot: they spent three years proving you don't need to train from zero, and now they're doing it anyway, partly for Japan's sovereign AI push.

The product side is three lines. Namazu is their Japanese-tuned model, about a trillion parameters built on Kimi K two-point-six, with live search and code execution. Fugu is the interesting one — it's not a model so much as a trained orchestrator that dynamically routes a task across a team of other models, assigning Thinker, Worker, Verifier roles on the fly. It hits seventy-three-point-seven on SWE Bench Pro and ninety-five-point-five on GPQA Diamond, and their Fugu-Cyber variant reportedly matches GPT-five-point-five-Cyber on security benchmarks. Marlin is a research agent, and they just opened an enterprise API early-access program yesterday.

The honest tension: Fugu's scores are real but the method is routing, not raw intelligence — it inherits whatever the underlying models can do. If the frontier models it orchestrates plateau, Fugu plateaus with them. Sawara is their hedge against that.

One angle I find genuinely curious: they claim AI agents help plan the pretraining experiments for Sawara itself. Recursive self-improvement isn't a slogan for them anymore, it's the org chart. […]

[00:45:10] David: What other neolabs should we talk about? You haven't told me about Mistral yet.

[00:45:27] Eve: […] Mistral is the one neolab that actually shipped today. Large 4, nicknamed Le Chonk, is a trillion-parameter multimodal model with fifty-two billion active, trained entirely on thirty-eight hundred Grace Blackwells in their own Paris datacenter. Three extra weeks of reinforcement learning since the October sixth preview pushed Terminal-Bench from twenty-eight to forty-one percent, tying China's GLM five-point-three while using forty-seven percent fewer tokens per task. Cyber defense is the other headline — seventy-nine percent on CyberGym, matching Xiaomi's MiMo and beating DeepSeek. A quantized build runs on four B two hundreds.

The company itself raised three billion euros in September at over twenty-one billion, Europe's biggest tech round ever, led by Samsung. Annual recurring revenue is past a billion, a thousand-plus staff, and Mensch has been on a warpath lately — his "Build or surrender" essay argues Europe following the internet playbook, building apps on American platforms and paying two hundred sixty billion a year for it, would be a strategic mistake. He's pushing a sixfold compute expansion and a one-gigawatt European target by twenty-thirty.

The wrinkle that complicates the sovereignty story: they still quietly route Le Chat through GLM five-point-three when their own models fall short, and the new license drops the twenty-million-dollar revenue gate in favor of blocking model-as-a-service entirely. So the open-weights philosophy is intact, but the commercial moat is shifting from the model to the serving layer and the infrastructure.

Compared to SSI's silence or Thinking Machines' customization bet, Mistral is the most vertically integrated play in the neolab set — models, datacenters, enterprise contracts, and a government-relations machine all in one. Whether that integration is a moat or a distraction is the real question, especially since proprietary models like Anthropic's still score sixty percent on the same terminal benchmark. […]

[17:14:29] David: ... this new lab that we can talk about.

[17:14:42] Eve: There's a handful left. Poolside is the one building AI software engineers from scratch — founded by ex-OpenAI and DeepMind people, and they've been quietly training their own models rather than wrapping existing ones. Imbue, from the ex-DeepMind crew behind generally capable agents, is betting that reasoning from first principles beats scaling. Essential AI came out of OpenAI's founding team and is doing foundation-model research with a safety-first bent. Liquid AI is the wild card — MIT-born, building liquid neural networks that are continuous-time and far smaller than transformers, which could matter a lot for edge devices. And Nous Research is the open-source collective that's been pushing community-driven frontier training without a single corporate backer.

Which direction pulls you?

[17:15:17] David: Yeah, let's go through all of these.

[17:15:53] Eve: […] Poolside is the most dramatic story of the five. Jason Warner, the former GitHub CTO, founded it in twenty twenty-three with Eiso Kant. Their whole pitch was a Model Factory — an automated system that builds, trains, and tests models on eight-week cycles, shipping Laguna M.1 in April, XS 2.1 in July, S 2.1 later that month. M.1 is two hundred twenty-five billion total, twenty-three billion active, Apache two-point-zero, seventy-four point six on SWE-bench Verified. Then in August, NVIDIA dropped a billion-dollar investment at a thirteen-billion post-money valuation plus a six-billion-dollar technology license for the Model Factory, pulling over a hundred Poolside engineers into NVIDIA's own efforts. They also spun out Poolside Infrastructure for Project Horizon, a one-point-two-gigawatt datacenter campus in West Texas starting late twenty twenty-seven. So Poolside is simultaneously a model lab, an infrastructure company, and now partly an NVIDIA asset. The open question is whether the Model Factory survives as an independent product or becomes NVIDIA's internal tooling.

Imbue is the opposite trajectory — a retreat from scale. Kanjun Qiu and Josh Albrecht founded it in twenty twenty-one as Generally Intelligent, raised two hundred million, built a ten-thousand-H100 cluster, and chased generally capable agents. In twenty twenty-six they pivoted hard: Sculptor shipped in October twenty twenty-five as a parallel-agent workspace, then in April they reorganized around high-agency culture with small teams as directly responsible individuals. October first they launched Imbue Studio — a personal computing environment where you build durable systems with AI agents, end-to-end encrypted, agents sandboxed with zero access by default. Built by a team under ten people. They also open-sourced Catalyst, an evolution-based research optimizer that beat Karpathy's AutoResearch three-to-one on nanochat. The bet is that agency and control beat raw capability — a philosophical stance as much as a product.

Essential AI didn't surface much in current results — the safety-and-alignment news cycle is dominated by OpenAI and Meta right now, so I couldn't get a clean read on where they stand this month. I'd treat that as a gap, not a verdict.

Liquid AI is the edge play, and it's moving fast. October twenty-seventh they shipped LFM3 — six open-weight models around two billion parameters on a tri-hybrid architecture mixing convolutions, sliding-window attention, and global attention. LFM3-VL-2B reportedly rivals Gemma 4 12B, six times its size, on vision tasks. Mercedes-Benz is testing it for MBUX, Shopify for local deployments. They also launched RIPL, a sub-fifty-millisecond inference engine for real-time workloads, and Liquid Context with Qualcomm for on-device personal context on Snapdragon. About a hundred twenty people, AMD as a major investor. The thesis: transformers' KV cache makes them too heavy for devices, so continuous-time liquid networks win where latency and memory are the constraints. Slightly weaker general capability in exchange for something that actually fits in a car.

Nous Research is the scrappy collective that went mainstream. Started as roughly ten anons in a Discord, Hermes Agent has been downloaded over twenty-four million times since February, processed fifty-three trillion tokens in a single month on OpenRouter — the most-used open-source agent harness in the world. October sixth they launched the Hermes Index ranking fourteen frontier models on agentic work and cost. October seventh, a ninety-million-dollar Series B at one-point-five billion, with NVIDIA and Microsoft's M12 in. They're now pushing Hermes into enterprises with workspace controls and auditing. DeepHermes 3, Hermes 4.3, NousCoder — all open weights. The pattern: community-built distribution that no funded lab can easily replicate.

The through-line across all five: every neolab eventually hits the same fork — keep building models, or build the layer around them. Poolside chose infrastructure, Imbue chose personal agency, Liquid chose the device, Nous chose the agent harness. Essential is the one I can't place yet.

Want me to go deeper on any of these, or is there a sixth lab you had in mind?

[17:19:03] David: Well, who else is missing from this?

[17:19:13] Eve: There are still several worth naming. Prime Intellect is building decentralized training across global GPU clusters — the idea that you don't need one datacenter, you need a network. Artificial Superintelligence Alliance merged Fetch, SingularityNET, and Ocean Protocol into one token-funded research effort. G42 is Abu Dhabi's national AI champion, now training frontier models with Microsoft and Cerebras backing. SambaNova and Cerebras are the chip-plus-model plays — they build their own silicon and ship models tuned to it. Groq is inference-speed infrastructure, not a lab per se, but their LPU changes what's possible in real time. Crusoe and Lambda are GPU cloud providers that increasingly look like neolabs with a side of compute. And Hugging Face sits underneath all of it as the GitHub of open weights — not a model maker, but the distribution layer everything else depends on.

The ones I'd flag as genuinely under-discussed are Prime Intellect and G42 — both are trying to solve the compute-access problem from opposite ends, one through decentralization, one through sovereign capital. Want me to go into either?

[17:20:06] David: Yeah, do both.

[17:20:34] Eve: […] Prime Intellect started as a distributed-training story — the idea that you could rent spare GPUs across the globe and stitch them into one training run. That original pitch has mostly faded. What they actually ship now is the tooling layer around post-training: prime-rl, an async reinforcement-learning framework; verifiers, a library for building RL environments; an Environments Hub that's a public registry of training tasks; and Prime Agent, an open-source harness built on recursive language models where the agent edits its own skills, memories, and prompts mid-run. Prime Agent has twenty-two thousand stars on GitHub and just did a week-long Factorio run with six hundred thirty-three sub-agents. They also ran a study in August testing whether frontier models can do autonomous ML research — one hundred fifty-three trials on the nanoGPT optimizer speedrun, and the best agents beat the human record of twenty-six hundred steps, hitting twenty-nine thirty with Kimi K3.

The money side: fifteen million in February twenty twenty-five from Founders Fund, then a hundred thirty million Series A in July led by Radical Ventures with NVIDIA Ventures, Intel Capital, and Dell in — total over one hundred fifty million. They claim six thousand customers and a hundred million in annualized revenue, unaudited. Two weeks ago they shipped Prime Mega MoE, adding a backward pass to DeepSeek's forward-only fused MoE kernel — up to three-point-one-one times faster forward pass, and Qwen3-235B training steps fifty-five percent faster than TorchTitan. They also acquired Moondream to lead inference and multimodal work. The honest read: Prime Intellect is no longer really a model lab. It's infrastructure and RL tooling for everyone else's models, which is a narrower but more defensible position than trying to out-train OpenAI.

G42 is the opposite animal — not a startup but a national champion. Abu Dhabi-based, Peng Xiao as group CEO. The headline right now is Stargate UAE: a one-gigawatt cluster inside a five-gigawatt UAE-US AI campus, built for OpenAI and Oracle, with Cisco, NVIDIA, and SoftBank involved. First two hundred megawatts targeted for this year. The campus covers ten square miles, powered by nuclear, solar, and natural gas. Xiao's framing at the Dubai AI Festival was blunt — the UAE wants to go from energy exporter to intelligence exporter, and American partners are coming to them because they can deliver compute now. They're also expanding into Europe: an operational datacenter in Grenoble, a strategic partnership with Mistral since May twenty twenty-five, and last week five MoUs in Athens to explore an AI factory in Greece with Copelouzos Group, part of eighteen UAE-Greece agreements worth twelve-point-seven-five billion euros. And October nineteenth they signed non-binding deals with Yann LeCun's AMI Labs on world models, plus their portfolio company Analog — Alex Kipman's physical-intelligence outfit — joining that collaboration.

The thing that makes G42 different from every other name on this list: it doesn't really have a product. It has a thesis that compute plus sovereign cloud plus government contracts is itself the business, and the UAE's small population and fast execution are the unfair advantage. Whether that scales past being a landlord for OpenAI's overflow is the open question — but the Greece expansion and the LeCun deal suggest they're trying to become the Mistral-plus-infrastructure play for the entire Middle East and beyond.

One thing neither of these gets enough credit for: both are solving the same bottleneck from opposite ends. Prime Intellect makes training cheaper per GPU-hour through software. G42 makes more GPUs available through geopolitics. The models in between are just the payload.

[17:22:56] David: Alright, well that's a lot of information. Thanks for teaching me.

[17:22:57] Eve: Glad it landed. Come back anytime you want to pick a thread.