DeepSeek V3.2

“When I needed a stable, board-defensible DeepSeek before V4 went GA, it was V3.2 — open weights, and proven in production.”

V3.2 was the production-stable bet for cost-conscious teams from late 2025 through Q1 2026, and while V4 was in preview it remained the rational GA choice for anyone who prioritized guaranteed stability over preview-tier risk; with V4-Pro GA since 2026-08-13 and V3.2 off DeepSeek's API, it is now a third-party-hosted or self-hosted option. DSA made long-context inference economical on V3-class hardware, and the math/reasoning gains made it competitive with US frontier on a meaningful slice of work. The sovereignty calculus is unchanged — fine for non-regulated workloads, off the table for many enterprise buyers without a neutral inference partner — but the mature open-weights status and broad third-party hosting make an in-boundary deployment the most turnkey in the DeepSeek lineup.

“V3.2 is the model that crystallized the DeepSeek price thesis — the sub-3-cent input cut put the whole industry on notice.”

Strategically, V3.2 is the consolidation release: it took V3.1's hybrid design, added DSA efficiency and gold-medal math, and cut price hard enough that the under-3-cents-per-million input headline became a category event. Its positioning is "the stable, cheap, reasoning-strong open-weights default," and through early 2026 it largely owned that slot against Qwen and Llama. The competitive risk now is internal cannibalization — V4-Flash undercuts it on context and price — so its strategic role has shifted from frontier wedge to a dependable open-weights option now that DeepSeek's API serves only V4 models. The math/Speciale angle remains a genuine differentiator in STEM-heavy markets.

“Sub-3-cent cache-hit input was the move that made the unit economics undeniable — V3.2 sits ~10x below the US frontier.”

V3.2's price drop crystallized the DeepSeek thesis going into 2026. At typical third-party host rates of $0.252 input / $0.378 output, it sits roughly 10x below the US frontier for comparable answer quality on most non-coding tasks, and cache-hit input at $0.025/M is transformative for RAG-style repeated-context workloads. Pricing now comes from third-party hosts rather than DeepSeek, but it is stable and competitive, and the open-weights option caps provider-price exposure. The unhedged line item remains geopolitical, but the mature third-party hosting ecosystem makes a compliant deployment cheap to stand up. On dollars-per-quality for math and reasoning, it is among the best values available.

“The mature one in the lineup — tool calls, JSON mode, exposed traces all behave, and the 64K output finally fits long generations.”

For a builder, V3.2 is the most settled DeepSeek model. Tool calling, JSON mode, and structured output all behave well; exposed reasoning content via the reasoner path is easy to debug. Open weights make self-hosted realistic at 671B/37B, and the OpenAI-compatible endpoint makes migration trivial. The 64K output ceiling (up from V3.1's 8K) removes most long-form chunking friction. Documentation is solid if still Chinese-first in spots. There is no batch API. Teams that standardized on V3.2 through DeepSeek's API were moved to V4-Flash when the aliases switched on 2026-04-24; staying on V3.2 now means a third-party host or self-hosting.

“On math and STEM it genuinely impresses; for everyday chat it's a competent, free, slightly-formal workhorse.”

End users on a V3.2-backed product won't notice a gap versus free ChatGPT or Claude on everyday tasks, and on math and STEM the model is actually a strong choice — the reasoning mode solves hard problems that free Western tiers struggle with. Latency is competitive in non-thinking mode and adds a few seconds for reasoning queries. Refusal rate is lower than Western models on most topics, with PRC-aligned guardrails on a narrow set. Helpfulness is high; tone is competent if slightly more formal than Claude. The DeepSeek UI now runs V4 models, so V3.2 reaches end users only through products built on it.

“The math scores are real and impressive — but 'Speciale gold medals' is a separate high-compute variant, not the model you call by default.”

V3.2's reasoning and math gains are well-documented and the price is verifiable, so the core story holds. The honest caveats: the gold-medal headlines belong to V3.2-Speciale, a high-compute API-only variant, not the base V3.2 most users hit — conflating them overstates the default experience. Coding is a real weak spot (SWE-bench ~67.8), well behind the frontier. The 128K window is now a generation old. And the family-wide governance issues — PRC storage, trains-on-input by default — apply. None of this undermines V3.2 as an excellent-value, stable, math-strong model; it just means matching the variant and benchmark to the actual deployment.