Mercury 3
Mercury 3 (inception/mercury-3) is Inception Labs' flagship diffusion language model (dLLM), released October 19, 2026. It decodes tokens in parallel via iterative denoising rather than strict left-to-right autoregression, pairing a 1M-token context window with a vendor-reported ~1,068 tokens/sec on NVIDIA GPUs and aggressive launch pricing of $0.10/$0.35 per 1M input/output tokens — positioning it as a latency-first alternative to autogregressive speed-tier models like Claude Haiku 5.5 and GPT-6 Luna.
Empirical Evaluation Results
Architectural Profile & Capabilities
Mercury 3 is a commercially served, text-only diffusion LLM from Inception Labs and the lab's most capable model to date, launched 2026-10-19. Unlike autoregressive frontier systems (GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro, DeepSeek V4 Pro, Grok 4.6, Qwen 3.8, Kimi K3), Mercury 3 generates and refines tokens in parallel through iterative denoising, which is the source of its throughput advantage and its unusual latency profile. OpenRouter characterizes it as more than 40x larger than the original Mercury model while holding comparable throughput. Published specifications: 1,000,000-token context window (up from 260K on Mercury 2.5), up to 131,072 output tokens, four reasoning-effort settings (instant, low, medium, high), temperature range 0.5–1.0, plus tool calling and structured output support. Pricing is $0.10/1M input and $0.35/1M output as a 50% launch discount off standard $0.20/$0.70; cached input is $0.01/1M on OpenRouter. Throughput claims diverge sharply by measurement context: Inception reports ~1,068 tokens/sec on NVIDIA GPUs and "1,000+ tokens per second" with 4–8x speed advantage over GPT-6 Luna and Claude Haiku 5.5 at matched reasoning effort, while OpenRouter's live provider telemetry shows ~241 tokens/sec and ~1.08s latency. Treat the 1,068 figure as controlled-hardware vendor telemetry, not a guarantee of served throughput. Benchmark posture: Inception's release table (as reported by BenchLM, MarkTechPost and eesel AI) shows Mercury 3 at 89.23 GPQA Diamond, 76.84 IFBench, 51.08 SWE-Bench Atlas, 46.31 DeepSWE v1, 60.35 MultiChallenge, 75.50 MRCR v2 and 71.33 DeepSearchQA. These are vendor-reported and largely self-run; no independent harness replication (Artificial Analysis, Vals AI, Scale Labs) was found as of this dossier. Critically, Inception has not disclosed parameter count or attention architecture, and no technical report accompanies this release. Comparisons should therefore be framed as "beats Haiku 5.5 / GPT-6 Luna on selected evals at 4–8x the speed," not as frontier-absolute capability — it does not contend with Opus 5, GPT-6 Astra, Gemini 3.8 Pro or DeepSeek V4 Pro on hard reasoning depth, and its SWE-Bench Atlas (51.08) and DeepSWE (46.31) results sit below top coding agents. Operational identity: a cheap, extremely fast, long-context text workhorse for high-volume orchestration, search, extraction and codebase-Q&A sub-agents — where latency compounds across many calls — rather than a deep-reasoning flagship.
Inception does not publish parameter count (a co-founder stated the company is 'not publicly announcing the model size right now'), so no active/total parameter figure or MoE routing claim can be verified; the model is architecturally a diffusion language model that generates and revises multiple token positions in parallel across denoising steps rather than committing sequentially per token. No verified public specification exists for its attention mechanism (MHA vs GQA vs MLA vs sliding-window/hybrid) — this dossier makes no claim on that point. OpenRouter notes Mercury 3 is >40x larger than the original Mercury model at comparable throughput. Served with four reasoning-effort settings (instant/low/medium/high) and up to 131,072 output tokens over a 1,000,000-token context; text-only modality (no vision/audio), with tool calling and structured output supported per Inception docs.
Recommended Workloads & Primary Use Cases
- •Throughput is provider- and setting-dependent: Inception's 1,068 tps is controlled-hardware telemetry, while OpenRouter serves ~241 tps at ~1.08s latency; reasoning-effort 'high' and long tool loops will erode the diffusion speed advantage dramatically.
- •All published Mercury 3 benchmark numbers (GPQA Diamond 89.23, IFBench 76.84, SWE-Bench Atlas 51.08, DeepSWE v1 46.31, MultiChallenge 60.35, MRCR v2 75.50, DeepSearchQA 71.33) are vendor-reported without a technical report, sample counts, or harness disclosure — no independent replication (Artificial Analysis, Vals, Scale Labs) was found. Treat as indicative, not audited.
- •MRCR v2 figure is internally inconsistent across secondary coverage: BenchLM and eesel list 75.50 for Mercury 3 while a MarkTechPost snippet cites '78.40 vs 75.50' in a comparative framing. Validate long-context recall on your own needle-in-haystack workload before relying on 1M-token retrieval.
- •Diffusion decoding changes failure modes versus autoregressive models: expect occasional token-level refinement artifacts, less predictable output determinism across denoising steps, and temperature sensitivity (supported range is narrow, 0.5–1.0) — pin seeds/params and regression-test prompts after provider updates.
- •No disclosed parameter count, attention design, or technical report means quantization/serving assumptions, true context degradation curves, and fine-tuning support cannot be planned from public documentation; also verify tool-calling schema fidelity and parallel-call behavior per provider, since agent reliability is not independently benchmarked for this variant.
- •Launch pricing ($0.10/$0.35) is explicitly a 50% promotional discount off $0.20/$0.70 standard — model costs at standard rates for multi-quarter budgeting and check whether cached-input pricing ($0.01/1M) is available on your chosen provider.
Calls to Mercury 3 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Inception.
More Models from Inception
Compare other engines available in this laboratory.
Inception's flagship 260K-context diffusion LLM (dLLM), generating and refining text in parallel at over 1,100 tps for real-time agents, code synthesis, and structured workflows accessible with Zero Data Retention on ARMES.
— | Speed-critical inference, real-time applications, rapid response scenarios