Xiaomi's MiMo team published a dedicated case study walking through how MiMo-V2.6-Pro acted as a "co-scientist" for its Materials Core team, tasked with designing a new material to capture PFAS — the "forever chemicals" family linked to immune, liver, developmental, and cancer risks, now subject to EU restrictions under REACH. It's a far more detailed writeup than the brief mention in Xiaomi's main MiMo-V2.6 launch materials this week, and the extra detail is worth reading closely on its own, separate from the benchmark questions the launch post already raised.
The chemistry described is real, and specific enough to check
Credit where it's due: this isn't a vague "AI designs new material" claim with nothing underneath it. The target class is metal-organic frameworks — porous crystalline scaffolds built from metal nodes and organic linkers, work recognized with the actual 2025 Nobel Prize in Chemistry, awarded to Susumu Kitagawa, Richard Robson, and Omar Yaghi. The Nobel committee's own materials name "separating PFAS from water" as one of the field's known applications — this isn't an exotic use case MiMo stumbled into, it's an active, publicly funded research area with a growing literature: zirconium-based UiO-66 and UiO-67 frameworks specifically, functionalized with amino groups or fluorinated linkers, are already the subject of multiple 2025 papers on PFOA capture.
The three structures Xiaomi shows — C50 (an unmodified biphenyl-linker UiO-67-type Zr₆ framework, serving as reference), B50 (linker modified with a CF₃-biphenyl group), and A50 (linker swapped for the larger, flatter pyrene) — are internally consistent with the stated design logic: adding a positively charged site to attract PFOA's negatively charged carboxylate head, and enlarging the hydrophobic aromatic surface to accommodate its fluorinated tail. The interactive figure reports a computed adsorption energy of -2.78 eV and an N⁺···PFO⁻ distance of 3.90 Å for the selected candidate — numbers that look like genuine DFT-style output, not invented set-dressing, and land in a plausible range for an electrostatic ion-pair interaction. This reads as a model that navigated real, specific, checkable chemistry rather than producing plausible-sounding nonsense.
The "million to ten million times" figure has real physics behind it, and a real gap around it
The claim that A50 and B50 show PFAS adsorption "one million to ten million times" C50's is the kind of number that sounds inflated until you do the arithmetic. Binding affinity scales exponentially with binding energy: at room temperature, an improvement of roughly 0.35 to 0.41 eV in binding free energy alone would produce a 10⁶-to-10⁷-fold shift in an equilibrium binding constant. That's textbook physical chemistry, not a red flag by itself.
What isn't shown is the path from a single-point computed adsorption energy to that macroscopic ratio — the concentration regime, whether this is a thermodynamic equilibrium claim or something faster to reach kinetically, and how much the implicit or explicit solvent model used for a "-2.78 eV" gas-phase-style number carries over to real aqueous conditions. A range spanning a full order of magnitude (10⁶ to 10⁷) for what should be one computed number is also worth noting on its own: that's either honest uncertainty being disclosed, or a sign the underlying estimate is looser than the two decimal places on the adsorption-energy figure suggest.
Nothing here has been made, and nothing here has captured a PFAS molecule
This is the part worth being precise about, because the case study is precise about it too, to its credit: every result is a "dry-lab experiment" — literature review, hypothesis generation, a novelty check, and computational screening. Nothing was synthesized. No PFAS molecule has been placed in contact with A50 or B50 in a beaker. The stated purpose is explicitly to "identify more promising candidates for resource-intensive wet-lab testing," not to report a validated material.
That's a legitimate, useful role for an AI system to play, and Xiaomi doesn't oversell it in the technical description. But it's worth setting next to Anthropic's protein-binder design campaign, covered on this blog a month ago, because that's the closest recent comparison available and the contrast is stark. There, Claude's 1,320 designs were physically synthesized and tested by two independent outside contractors — Adaptyv Bio and Twist Bioscience — running genuinely different assay methodologies, with hit-rate and binding-affinity numbers coming from parties with no stake in the result. Anthropic also released the full dataset, including one target whose results were inconclusive and reported as such rather than dropped. That's what "verified" means when a lab wants the word to carry weight.
Nothing comparable happens here. A50 and B50 are proposals, not results. The headline framing — "co-scientist," a workflow diagram running from literature review to "candidate materials," binding numbers reported to two decimal places — reads like a finished discovery. The text underneath it is honest that it's a screening step. Those two things sit in tension throughout the case study, and a reader skimming the graphics would come away with a stronger claim than the actual evidence supports.
The "verification" on offer is two solicited testimonials, not blind review
Xiaomi includes quotes from two named, credentialed academics — Prof. Jinhu Dou of Peking University and Prof. Chen Wang of Xiamen University, both with genuine MOF-research backgrounds, which is more transparent than an anonymous endorsement. But it's worth being exact about what these quotes are and aren't. Both professors reviewed a case study Xiaomi had already completed and framed favorably; neither is described as having attempted to reproduce the binding-energy calculation, checked the novelty claim against the patent and literature search MiMo says it ran, or been given a blind comparison against an unlabeled human-designed alternative. Prof. Dou's assessment that the work was "on par with that of a well-trained doctoral researcher" and Prof. Wang's account of being "astonished" are genuine expert reactions to what they were shown — not the same category of evidence as a contract lab independently measuring whether PFAS actually binds to the material, and both professors say as much themselves: Dou writes that "these candidates deserve further experimental investigation," Wang says he's "eager to see these designs tested experimentally." The case study's own quoted experts agree the real test hasn't happened yet; it's the packaging around their quotes that reads as more conclusive than that.
The efficiency claim has no stated baseline
"MiMo-V2.6-Pro improved human experts' work efficiency tenfold, shortening the R&D cycle from one month to 2–3 days" is an estimate Xiaomi's own team made about its own work, with no description of what the one-month baseline is measured against — a comparable unassisted project, a typical industry timeline, or an internal rule of thumb. There's no control condition, no second team working the same problem without the model, and no methodology for how "tenfold" was calculated rather than asserted. It may well be roughly right; iterative literature review and computational screening are exactly the kind of work an LLM-plus-tools pipeline should compress. But it's a self-reported number about the model's own developer's productivity, presented with the same confident precision as the adsorption-energy figures, and it doesn't carry the same evidentiary weight.
What this actually shows
Strip away the framing and what's left is a real, technically literate demonstration: MiMo-V2.6-Pro searched literature, proposed chemically sound modifications to a known MOF platform within an active research area, ran automated computational tools to screen its own proposals, and produced two candidates with specific, plausible-looking binding metrics. That's a genuinely useful capability, and the chemistry underneath it holds up to scrutiny better than most single-example AI-for-science showcases do.
What it isn't yet is a validated material, an independently reproduced result, or evidence that goes beyond what two people Xiaomi selected to comment on it said about it. The gap between "designed a new material" in the framing and "proposed two computationally screened candidates for testing that hasn't happened" in the actual content is the whole story here — and it's a gap this blog now has a direct, recent point of comparison for, from a lab that closed it.