Atlas

Models

← All models

Mercury 2

Inception Labs · 2026-02-24

Mercury 2 is Inception Labs' diffusion-based reasoning language model, announced in February 2026 with tunable reasoning, native tool use, and structured JSON output. Unlike autoregressive models, it generates text through parallel refinement, producing multiple tokens simultaneously. Mercury 2 supports a 128K-token context window and is designed for high-throughput production reasoning workloads.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo280
AA-Briefcase v1.1 Elo266
AA-LCR36.3
AA-LCR v1.143.7
AA-Omniscience Index-50.7
ALE-Bench1190
Arena Code WebDev — Brand & Marketing1154
Arena Code WebDev — Consumer Product1169
Arena Code WebDev — Data & Analytics1143
Arena Code WebDev — Frontend1169
Arena Code WebDev — Gaming1199
Arena Code WebDev — HTML1199
Arena Code WebDev — React1160
Arena Code WebDev — Reference-Based Design1185
Arena Code WebDev — Simulations1211
Arena Text — Business, Management & Financial Ops (No Style Control)1349
Arena Text — Chinese (No Style Control)1417
Arena Text — Creative Writing (No Style Control)1289
Arena Text — English (No Style Control)1378
Arena Text — Entertainment, Sports & Media (No Style Control)1315
Arena Text — Exclude Ties (No Style Control)1312
Arena Text — Exclude Ties (Style Controlled)1295
Arena Text — Expert (No Style Control)1358
Arena Text — Hard Prompts (No Style Control)1361
Arena Text — Hard Prompts English (No Style Control)1373
Arena Text — Instruction Following (No Style Control)1327
Arena Text — Life, Physical & Social Science (No Style Control)1344
Arena Text — Longer Query (No Style Control)1330
Arena Text — Medicine & Healthcare (No Style Control)1363
Arena Text — Multi-turn (No Style Control)1357
Loading Atlas data…