Mistral's Leanstral 1.5, UBTech's U1 Humanoid, and Baseten's $1.5B Infrastructure Surge

Three announcements define early July 2026's convergence of AI capability, physical robotics, and infrastructure scale. Mistral's Leanstral 1.5 (119B total / 6.5B active MoE, Apache 2.0) solves 587/672 PutnamBench problems via test-time scaling — the first open-weights model to challenge proprietary formal mathematics systems; UBTech's U1 humanoid secures 13,000 launch-day pre-orders for its VLAM-powered, mood-sensing consumer robot; and Baseten's $1.5B Series F validates the "picks and shovels" investment thesis as enterprise inference infrastructure becomes the AI stack's critical bottleneck.

🤖 Mistral Leanstral 1.5 — Open-Source Formal Mathematics

What Lean 4 Theorem Proving Is and Why It Matters

Lean 4 is a dependently typed programming language developed by Microsoft Research that serves as both a general-purpose programming language and a formal proof assistant. When code is "proved" in Lean 4, the Lean kernel mechanically verifies every logical step against the axioms of mathematics. A Lean 4 proof is mathematically certain — no human reviewer needed.

Applications of automated Lean 4 theorem proving:

Leanstral 1.5 — Architecture and Performance

Model architecture:

PutnamBench performance:

The Putnam Competition is the most prestigious undergraduate mathematics competition in North America, featuring 12 problems per year over 2 sessions. PutnamBench is a formalised version of all Putnam problems from 1962–2024 (672 total problems) encoded as Lean 4 theorem-proving tasks.

Leanstral 1.5 at 87.4% is the highest-scoring open-weights model ever on PutnamBench — and scores higher than all closed proprietary models except OpenAI o3's specialised math configuration.

What "test-time scaling" means technically: Standard inference: one forward pass → one output. Test-time scaling: the model runs multiple candidate proof attempts (beam search over Lean 4 syntax), verifies each attempt with the Lean kernel in real-time, uses the verification result to guide further generation. It is an online search problem, not just text generation:

1. Generate candidate proof tactic (e.g., ring_nf, omega, norm_num)
2. Apply to current proof state → Lean kernel returns: success / failure + error message
3. On failure: use error message as conditioning context for next tactic attempt
4. On success: advance proof state and repeat for next subgoal
5. Budget: 512 tactic attempts per theorem (controllable compute budget)

🦾 UBTech U1 — Consumer Humanoid with VLAM Architecture

The Vision-Language-Action Model (VLAM) Distinction

Traditional industrial robots execute pre-programmed scripts. Humanoid robots with VLAMs can:

• See: RGB + depth cameras feed a vision transformer
• Understand language: Natural language instructions parsed by an LLM
• Act: An action model maps understanding to motor commands

U1 VLAM Architecture:

U1 hardware specifications:

13,000 first-day pre-orders = $1.157B in committed revenue — the largest single-day commercial commitment in consumer robotics history. Primary purchaser demographics from UBTech's announcement:

• 38% commercial (small business: retail, hospitality, elderly care)
• 29% institutional (hospitals, elder care facilities)
• 21% research institutions
• 12% individual consumers (early adopter demographic)

Market context: Global humanoid robotics VC funding H1 2026: $18.3B (up from $3.2B in H1 2024 — a 472% increase).

💾 Baseten $1.5B Series F — The Inference Infrastructure Market

Why Inference Infrastructure Is the New AI Moat

The AI capital allocation cycle has phases:

1. Training infrastructure (2020–2023): GPU clusters, data pipelines, foundation model training
2. Model development (2022–2025): Pre-training, RLHF, fine-tuning, model evaluation
3. Inference infrastructure (2025–present): Cost-effective, low-latency serving of models to billions of users

The inference cost problem for enterprise:

Cold start time is the interval between a user's first request (when the model is not loaded) and the first response. For enterprise use cases with variable traffic, long cold starts = poor user experience or permanent model residency = expensive idle GPU cost.

Baseten's technical differentiators:

Series F $1.5B round — terms:

• Lead: Thrive Capital (led), with a16z, Sequoia, and Databricks Ventures
• Valuation: $8.2B post-money (up from $1.1B at Series D in 2024)
• Use of funds: 55% GPU cluster expansion (targeting 80,000 H200s by Q2 2027), 25% multi-region expansion (EU and APAC), 20% R&D (custom inference kernels, speculative decoding library)
• Revenue context (disclosed): $420M ARR growing at 180% YoY at time of raise

📌 The Bottom Line

• mistral-leanstral-1-5-lean4-theorem: 119B total / 6.5B active MoE (inference cost = 6.5B dense); Apache 2.0 open-weights; PutnamBench: 587/672 (87.4%) — highest open-weights score ever, beats all closed models except o3 math-config; test-time scaling mechanism: candidate tactic → Lean kernel real-time verify → error conditioning → next tactic (512 attempts per theorem); applications: DO-178C aerospace, DeFi protocol security (eliminates re-entrancy/overflow class), Lean-verified OS primitives.
• ubtech-u1-humanoid-vlam: VLAM stack: ViT-L/14 (8 cameras/30fps) + Mistral 7B + diffusion transformer action model + multi-modal affect model; Orin NX 275 TOPS on-board (≤100ms latency); specs: 168cm/55kg, 48 DOF, 12 DOF hands, 5kg payload, 4hr battery, $89K retail; 13,000 first-day pre-orders = $1.157B committed revenue; split: 38% commercial/29% institutional/21% research/12% consumer; global humanoid VC funding H1 2026: $18.3B (+472% vs H1 2024).
• baseten-1-5b-inference-infrastructure: Cold start: 7B <1s, 405B <15s (vs 18–300s standard cloud); GPU memory sharing (multi-model per GPU at token level vs full model residency); autoscaling sub-second (pre-provisioned pools vs 3-8min instance launch); pricing: per-token+compute (vs per-GPU-hour); 400+ pre-optimised models; $8.2B valuation (from $1.1B 2024); $420M ARR +180% YoY; $1.5B: 55% GPU (80K H200s by Q2 2027) + 25% EU/APAC + 20% custom kernels.

📬 Stay Updated

Get the best of AI & technology delivered to your inbox every week. Subscribe to our free newsletter →

Disclosure: This post contains affiliate links. If you purchase through our links, we earn a small commission at no extra cost to you. We only recommend products we believe in.