DeepSeek Math-V2 just hit IMO gold. UK schools now have a decision to make.

Five months ago, DeepSeek was a headline about a Chinese lab undercutting OpenAI on inference cost. Today it is a headline about the first open-weight model to score gold at the International Mathematical Olympiad — and the Department for Education has just put £3m on the table to put AI tutors in front of 450,000 disadvantaged pupils. The two stories are the same story, and UK education leaders have about one term to work out what they do with it.

This post was originally published in late 2025, before DeepSeekMath-V2 existed, before DeepSeek R2 launched, and before the DfE formally opened its tutoring tender. The update below reflects where the state of play actually is in April 2026.

DeepSeekMath-V2: the first open-weight model to win IMO gold

On 27 November 2025, DeepSeek released DeepSeekMath-V2 on Hugging Face and GitHub under a permissive licence. The headline numbers, confirmed by independent coverage:

• 5 of 6 problems solved at IMO 2025 — gold-medal territory.
• 118 out of 120 on Putnam 2024 with scaled test-time compute.
• Gold-level on the 2024 Chinese Mathematical Olympiad (CMO) as well.

Google DeepMind's Gemini Deep Think and OpenAI's reasoning system both hit IMO gold earlier in 2025 (both scored 35/42 within the official 4.5-hour window, per DeepMind's blog post). What makes DeepSeekMath-V2 different is not that it's the first to gold — it's that it's the first open-weight model to get there. You can download it, run it on your own hardware, and fine-tune it on your own data.

The model's core trick is self-verifiable reasoning: it writes proofs and then independently checks them using a verifier trained on its own output. For education use, that's the interesting bit — not the medal itself, but the fact that the model can flag when its own working is suspect.

DeepSeek R2 and the collapse of the reasoning-model price floor

Running alongside Math-V2 is DeepSeek R2, released in the first half of 2026 as a 32-billion-parameter dense transformer under MIT licence — down from R1's 671B Mixture-of-Experts. R2 scores 92.7% on AIME 2025 and runs on a single 24GB consumer GPU.

The practical implication for a UK MAT or sixth-form college: a reasoning model capable of elite-level maths now fits on a £2,000 workstation, with no per-token API bill attached. That's a very different commercial picture from the one most school IT budgets are priced against.

The DfE's £3m AI tutoring tender — and why DeepSeek isn't on the list

In February 2026, the DfE formally opened its procurement for AI tutoring tools covering English, maths, science and MFL for Years 9–10, targeting up to 450,000 disadvantaged pupils with rollout by end of 2027 and classroom testing beginning summer 2026. It is, by some margin, the largest single AI-in-education commitment any UK government has made.

What the tender explicitly is not is a green light for DeepSeek. Several UK edtech roundups have noted that DeepSeek models have been effectively excluded from approved tool lists on data-residency and confidentiality grounds — the model is open-weight, but the hosted API runs on Chinese infrastructure. This is the central tension for any UK educator in 2026: you can get gold-medal maths reasoning for free, but only if you self-host, and self-hosting requires engineering capacity most schools don't have.

There are three realistic paths:

1. Use the hosted DeepSeek API — cheapest, fastest, but almost certainly non-compliant with DfE and Ofsted guidance on pupil data.
2. Self-host DeepSeekMath-V2 or R2 — fully compliant, gold-medal capability, but requires a GPU server and an engineer who can keep it running.
3. Use a UK-hosted wrapper — a third party runs the open-weight model on UK infrastructure and sells it as a compliant service. This is where the £3m tender money will most likely land.

What this means for schools, MATs and edtech in the UK

Three shifts are already visible in spring 2026:

• Marking and feedback workloads are falling. Third Space Learning's 2025 teacher survey reported that teachers using AI-assisted marking were cutting between 30% and 40% off their weekly marking time. Post-R2, those numbers are going up, not down, because reasoning models give better rubric-aligned feedback than the GPT-3.5-class tools most schools first adopted.
• The private tutoring market is being priced against AI. A one-to-one GCSE maths tutor in London is £40–£70 per hour in 2026. A self-hosted Math-V2 instance costs fractions of a penny per query. The gap is no longer about capability — it's about trust, safeguarding and accountability.
• MATs with in-house engineering are pulling ahead. Academy trusts that already run central IT functions are deploying open-weight models on their own hardware. Trusts that outsource IT are waiting for the DfE-approved vendor list, which won't exist until autumn 2026 at the earliest.

Your 2026 deployment checklist (honest version)

Forget the five-step roadmap the original version of this post published. Here's the actually-useful version for a head of department or MAT CIO in April 2026:

1. Decide your data boundary first. Before you evaluate any model, answer: can pupil-identifiable data leave your network? If no, hosted APIs — DeepSeek, OpenAI or otherwise — are off the table.
2. Pilot against a narrow, marking-heavy workflow. Y10 mock paper marking is the canonical starting point. Run the model in shadow mode against human-marked scripts for a term and measure agreement rate.
3. Use the verifier, not just the answer. DeepSeekMath-V2's self-verification is the feature that makes it educationally defensible. An answer a teacher can't audit is an answer a teacher shouldn't use.
4. Wait for the DfE vendor list before signing a multi-year procurement. The tender closes for bids through 2026. Vendor selection will be announced before year end.
5. Train teachers on the failure modes, not the features. The 2026 Scientific American coverage on AI olympiad performance included explicit skepticism from the IMO president about reproducibility. Your staff need to understand where these models hallucinate confidently — especially on word problems and mis-stated questions.

How the Marketing Operating System model applies to education providers

If you run an edtech company, a tutoring business or an independent school that has to market itself, the DeepSeek story is not really a maths story. It's a distribution story. Reasoning capability is no longer your moat. Delivery, compliance, brand and parent trust are.

That is precisely the problem Anjin is built to solve. You don't need one more AI tool bolted onto your stack. You need an operating layer that runs your content, campaigns, landing pages, lead routing and reporting as a single workflow — one that can publish against a curriculum-aligned content plan, produce parent-facing ads in the brand voice, and close the loop back to which messages drove which open-day bookings.

That's the category shift. Schools and edtech will spend 2026 figuring out which AI models to deploy in the classroom. The organisations that also figure out how to run their growth function the same way will be the ones that still exist in 2030.

Anjin: the Marketing Operating System for education providers

Anjin is the Marketing Operating System built for founders, school marketers and edtech teams who can't run a ten-person growth department. It turns your brand, offer and audience into a coordinated stack of agents that write, design, publish, measure and iterate — on your infrastructure, under your brand, without the SaaS bill for every function.

For education providers specifically, that means open-day campaigns, prospectus landing pages, parent nurture sequences, results-day press coverage and MAT-level brand consistency all running from one system, rather than five tools and three freelancers. See Marketing Operating System pricing for how the £888 lifetime license applies.

Sources: GOV.UK — 450,000 disadvantaged pupils, GOV.UK — Edtech and AI companies invited, PublicTechnology, MarkTechPost, SCMP, Google DeepMind, arXiv 2511.22570, Decode the Future, Scientific American, Third Space Learning