GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
← Back to feed

DeepSeek R2 Scores 92.7% AIME on 32B Dense Model — Runs on a Single RTX 4090

DeepSeek R2, released in April 2026 under the MIT license, marks a deliberate architectural reversal. Where R1 was a 671-billion-parameter Mixture-of-Experts model requiring at minimum an 8× H100 cluster, R2 is a 32-billion-parameter dense transformer that fits within the 24 GB VRAM envelope of a single RTX 4090 or A6000.

The benchmark numbers are not incremental. R2 scores 92.7% on AIME 2025 — the competitive mathematics benchmark that has become a standard proxy for frontier-grade reasoning. R1, on independent re-evaluation, scored around 74% on the same benchmark. The gap is 18.7 percentage points on the same task family, from a model that is 21× smaller by parameter count.

Key Numbers

MetricDeepSeek R1 (Jan 2025)DeepSeek R2 (Apr 2026)
Architecture671B MoE (37B active)32B dense
AIME 2025~74% (independent)92.7% (announced)
Local hardware floor8× H100 cluster1× RTX 4090 (24 GB)
LicenseMITMIT
Token cost vs. Western frontier—~70% cheaper

Why the Architecture Shift Matters

MoE architectures deliver efficient inference at scale — only a fraction of parameters activate per token — but they impose severe minimum hardware requirements and are difficult to run locally. R2’s shift to a dense 32B design trades raw capacity for accessibility: a model in this weight class runs on hardware that thousands of developers already own, with no multi-node coordination required.

The reasoning capability improvement suggests DeepSeek has found significant gains through training methodology rather than parameter count. The Parcae paper (Prairie et al., 2026) has separately documented that Recursive Depth Transformers can match the output quality of models nearly twice their parameter count — pointing to architecture efficiency research converging with DeepSeek’s own trajectory.

Pricing and Competitive Position

At approximately 70% below comparable Western frontier reasoning API prices, R2 continues the pattern DeepSeek established with V3 and R1: price at a level that pressures the cost floor for the entire category. For developers running AIME-class reasoning workloads, R2 presents the first consumer-hostable option that clears 90% on the benchmark — previously only achievable via cloud APIs with significantly higher per-token costs.

The release lands as the company separately advances DeepSeek V4, a 1-trillion-parameter MoE model targeting multimodal reasoning at the frontier. R2 and V4 appear to serve different markets: R2 optimises for accessibility and cost, V4 for maximum capability.

MIT licensing means R2 weights can be fine-tuned and deployed commercially without restriction — a significant differentiator from most frontier-adjacent reasoning models.