Appearance
The Frontier Splits in Two: Gemini 4 Pro Leaks Meet Xiaomi's MiMo-V2.6
The week the frontier split in two
Two things happened this week that look unrelated and aren't. Google's Gemini 4 Pro checkpoints started leaking, and developers found Xiaomi's MiMo-V2.6 family on Hugging Face with three visible checkpoints: a 1T-parameter model, a 311B, and a 9B distill of Qwen. Reddit is already counting the two missing slots.
Same week, two opposite answers to the same question. Google is selling frontier intelligence by the token. Xiaomi is giving away as much as it can and betting on hardware. Both are serious bets. Here's what's actually knowable right now.
What the Gemini 4 Pro leak actually shows
The leak trail is messy but consistent. Testers found what they call "Checkpoint 2", internal codename barium-b, and hammered it with 3D physics, code generation, and long-context reasoning prompts. The results read like a highlight reel:
- A 3D seaplane takeoff simulation that beat Opus 5.5, with smoother acceleration, believable water reflections, and proper splash rendering. The Opus version wobbled like it was about to come apart.
- An interactive 3D International Space Station orbiting Earth, generated in 16 minutes from pure code with no imported assets. It even fetched the correct Earth texture on its own. The same tester notes the UI is ugly and there's a perspective glitch on the backside.
- A Three.js lighthouse-to-rocket-launch animation that followed an instruction to include zero UI elements, with textures good enough that the tester praised the model's "aesthetic sense".
- A high-fidelity 3D Wii remote, generated shockingly fast. The tester also says the model over-polishes details you didn't ask for, and its design taste is inconsistent across parts.
- A mechanical hummingbird test where it came out on par with Fable and clearly ahead of GPT-6 Sol.
- A complete product website, frontend and backend, in 14 minutes. A complete interactive web game from a single prompt.
Read the fine print, because it matters: these are demos from an early post-training checkpoint, not the shipping model. Google DeepMind's new leader Koray Kavukcuoglu confirmed Gemini 4 is in early post-training. Pre-training started in July and essentially finished in about two months, which is fast for the compute thrown at it. He wants a release "far before the end of the year". Some rumors put Gemini 4 Pro as early as November.
The numbers, translated
Two leaked benchmark numbers anchor the story, and they only mean something when you translate them into work you'd actually hand the model. They come from a leaked chart, so treat them as directional, not final.
88% on DeepSWE v1.1 means the model handles multi-step agentic coding, the "fix this bug across three files and update the tests" work that eats junior developer hours. 95.3% on Terminal-bench 2.1 covers terminal-driven development tasks. Both reportedly clear GPT-6 Astra and Claude Fable 5.1. The rest of the leaked spec sheet is where the practical implications live.
Key Numbers
- 10M tokens: leaked context window. That's an entire monorepo in one shot, no chunking.
- 256k tokens: leaked max output length, so no "code stops halfway" failures on long generations.
- 88%: DeepSWE v1.1 agentic coding score.
- 95.3%: Terminal-bench 2.1 coding score.
- ~2 months: Gemini 4's early pre-training duration.
- $2.25 / $11.25: leaked per-million-token input and output pricing.
A 10M context window is the kind of spec that can get nerfed between leak and ship. Plan around 1M until the official page posts the real number.
MiMo-V2.6: Xiaomi's open-weight counterpunch
MiMo-V2.6 isn't getting the same hype, because open-weight releases rarely do. But the family structure is a statement. Three models are visible on Hugging Face:
| Model | Scale | Profile | What it means for you |
|---|---|---|---|
| MiMo-V2.6 (top tier) | ~1T total params | Likely sparse MoE, server-class | Lab or API territory, full weight set needs real clusters |
| MiMo-V2.6 (mid tier) | 311B | Likely MoE, routing unclear | Big-iron inference, feasible on multi-GPU nodes with quantization |
| MiMo-V2.6-Distill-Qwen-9B | 9B | Dense, distilled from Qwen base | Runs on a single RTX 4090, no cloud GPU required |
The 9B is the sleeper. A dense model that inherits behavior from the larger family and runs on consumer hardware is the one people can actually deploy today. The 1T is the halo product. The two missing slots are unspecified, and the rumor mill guesses at additional MoE variants or domain-tuned checkpoints. Xiaomi hasn't said.
Quick Take: Google is racing to ship a closed agentic workhorse while Xiaomi races to give away open weights, and the overlap between those two markets is where the next price war starts.
The checkpoint timeline you're actually looking at
This timeline explains why checkpoints keep appearing before the official release. Google is running an aggressive schedule, and the gap between Checkpoint 1 and Checkpoint 2 is visible. One tester described the second checkpoint as "a huge visible leap" over the first.
What the community is saying
I'd put 400 hours into Opus 5.5 and GPT-6 before Gemini 4 Pro's second checkpoint surfaced, and it still made my head spin. The mechanical hummingbird prompt I use to test physical reasoning came out on par with Fable and clearly ahead of GPT-6 Sol. But the same session that produced the 16-minute space station also left a perspective bug on its back face and a UI I wouldn't show a client. The brilliance is real, and so are the rough edges.
For MiMo, the tone is different. The Reddit thread about the model page is mostly people counting parameters and speculating: "Three visible models are 1T, 311B, and 9B, let's dream about two more." Others ask whether the 9B distill inherits the family's reasoning or just its chat style. Nobody has serious benchmarks yet, so expect a week of wild claims followed by measured ones. The practical truth is that a 9B distill handles short code tasks well but isn't a Gemini 4 Pro competitor. Nobody should expect it to be. It's the model you run where you already compute, which is the whole point.
Common pitfalls
Don't treat leaked checkpoints as the shipped model. Checkpoint 2 is an early post-training snapshot, and post-training changes behavior dramatically between snapshots. Build your evaluation harness now, but don't lock architecture decisions to a leak that could be stale in three weeks.
Don't assume the 1T MiMo runs on your hardware. Sparse MoE or not, a 1T-parameter model needs its full weight set resident in memory unless the release exposes a routing config for partial loading. If you're on a single workstation, the 9B distill is the only realistic target in this family. Check memory footprints before you buy GPUs.
Don't benchmark the 9B distill against the 1T. Distillation is lossy. Expect trailing performance on long-horizon agentic tasks even when the 9B holds its own on chat and short code. Lower cost per token isn't the same as higher capability per token.
Don't migrate providers on leaked pricing. $2.25/$11.25 per million tokens is aggressive if real, but leaked price sheets don't come with SLAs, rate limits, or guarantees that the final model keeps the same price. Run a stress test before you switch.
What this means for the market
If the 10M context, 256k output, and $2.25/$11.25 pricing all survive to launch, the competitive response will be fast. Don't be surprised if Opus 5.5 and GPT-6 pricing moves within a quarter. The leaked pricing alone is a shot at OpenAI and Anthropic, and the November rumor adds pressure.
Two longer-term data points matter here. Google is launching four TPU chips on a SpaceX Falcon 9 next week, a small step toward space-based AI infrastructure. And DeepMind's leadership says Gemini engineers now use Gemini 4 internally under the codename Antigravity, a direct response to the "Google employees use Claude" talk earlier this year. More telling is Koray's reframing: the AGI framing is outdated, and the real question is whether we can build agents we fully trust. That positions Gemini 4 Pro as an agent substrate, not a chat model.
One thing to remember
These two releases aren't competing for the same buyer. Gemini 4 Pro targets teams that buy intelligence by the token at scale. MiMo-V2.6 targets teams that own their hardware and want weights they can inspect, fine-tune, and run. The moment those two markets start overlapping, on price or on capability, the frontier stops being a single leaderboard and becomes a negotiation. That overlap is closer than it looks.
The Bottom Line
- If you build agentic coding tools, Gemini 4 Pro is the benchmark to beat, but wait for the official release and API terms before you rearchitect. Early checkpoints won't match the shipped model.
- If you're on consumer hardware and want frontier-adjacent open weights, deploy the MiMo-V2.6 9B distill. It's the only model in this family that runs on a single GPU, and the 1T won't run where you do.
- One thing to watch: the leaked $2.25/$11.25 per-million-token pricing. If Gemini 4 Pro launches anywhere near that, expect Opus 5.5 and GPT-6 pricing to respond within a quarter.