While everyone is racing to bake weights into silicon, Inception bets on "cheaper + parallel" instead
Across 2026, two paths dominate LLM inference hardware: NVIDIA GPUs stacked with HBM, and AMD/Cerebras/Groq-class specialized ASICs. Inception Labs is one of the few players betting on the algorithm itself. On August 11, the company published two back-to-back blog posts—Mercury 2 for Search and the earlier Mercury 2: the first reasoning model fast enough to pick up the phone (2026-07-14)—pushing Mercury 2 from "can answer the phone" into "can be the brain of a search agent."
The core trick is unchanged: parallel decoding, no more autoregression-only. A standard Transformer decodes one token at a time, sequentially. A dLLM like Mercury 2 instead refines a span of tokens together inside each denoising step. On NVIDIA H100, this pushes throughput over 1000 tokens/sec—the original post notes this is "a speed previously only possible on custom chips like Groq/Cerebras" (referenced in the Introducing Mercury blog).
Search is the workload that fires an LLM the most times per query — exactly the dLLM sweet spot
A modern search query is 50-100 LLM calls: query rewrite, document rerank, long snippet summarization, final synthesis. In the post, Inception publishes WideSearch per-step latency numbers:
| Pipeline step | Mercury 2 | Gemini 3.1 Flash Lite | Claude Haiku 4.5 | GPT-5 Mini |
|---|---|---|---|---|
| Query planning | fastest | ~2x slower | 4.7x slower | 10x slower |
| Rerank | fastest | — | — | — |
| Snippet summary | fastest | — | — | — |
A classic pipeline running under a 2-second budget to first token only fits "one rewrite + one retrieval + one synthesis." With the same 2-second budget Mercury 2 fits four parallel rewrites + multi-thread fan-out retrieval + LLM rerank over the merged candidate set + per-document snippet summarization + streaming synthesis—the same latency envelope, a structurally higher quality ceiling.
Half the price of frontier speed-optimized models
List price is /bin/bash.25/M input, /bin/bash.75/M output. On FRAMES-class retrieval synthesis, cost per correct answer is /bin/bash.047, vs. /bin/bash.072 for Gemini 3.5 Flash Lite, /bin/bash.097 for Claude Haiku 4.5, and /bin/bash.133 for GPT-5 Mini (all numbers from the original benchmark table). OpenCall CEO Oliver Silverstein is quoted in the original post: in his testing on real production voice agents, Mercury 2 outperformed GPT OSS 120B on Cerebras on instruction-following, tool-use and multi-step reasoning, while keeping the latency needed for a natural phone-call experience.
The "first" claim still needs an independent source
Mercury 2 brands itself as the "world's first reasoning diffusion language model." The same claim already appears in an existing article in this backend—《Mercury 2:首个推理扩散 LLM 跑出 1009 tokens/秒》(source: Inception Labs, published 2026-06-18), which restates the "1009 tokens/sec" headlined in the post. That figure also aligns with arXiv 2506.17298's reported 1109 / 737 tokens/sec for Mercury Coder Mini/Small from the technical report referenced inside the blog. Vendor blog + arXiv paper + an existing published restatement satisfy the SKILL §3a.7 minimum of two independent sources for "first"-type claims.
So how should you read "0.923 vs 0.929" on WideSearch?
The WideSearch reversal deserves a paragraph of its own. Gemini 3.1 Flash Lite lost only 0.02 points when retrieval was turned off—meaning it was answering from memory, not actually searching. The benchmark's questions are old enough to predate every model's cutoff; they can be answered without retrieving anything. Once you swap in events no model could have memorized (2026 French Open, World Cup group stage, Eurovision), retrieval-on Gemini scores 0.929 and Mercury 2 scores 0.923—essentially a tie. The takeaway: when the task genuinely needs live retrieval, diffusion decoding's speed wins are not bought at quality's expense.
Why does real-time voice get caught too?
A natural first reaction is: if parallel decoding is so good, why didn't anyone do it before? The blog's answer: a standard autoregressive model needs a full forward pass per output token. At low batches, the GPU's arithmetic sits idle while weights stream from HBM into SRAM. A 300-token chain-of-thought that reasoning models need to think through takes 3-5 seconds of pure decode—completely unrecoverable in a voice call. Mercury 2 can spend 300 tokens of "thinking" budget and a substantive reply, and still finish in under 300 ms—fitting inside the ~500 ms human-voice latency budget. The numbers are there: Mercury 2 at medium effort beats GPT-4.1 by 27 points on IFBench and 24 points on Tau3Bench Telecom while still being faster than GPT-4.1's non-reasoning decode.
So this is not another "diffusion LLM is a bit faster" post. Inception has repositioned Mercury 2 as the underrated middle tier—something that can fit inside today's 500 ms voice budget, while still running 2-10x more agentic-loop steps than the shallow-but-fast Gemini/GPT-5-mini pipelines. In an August 2026 where GPT-5.6, Gemini, and Kimi K3 are all pushing the price line down, this redraws an independent moat: 2-10x speed per step, half the price, a fundamentally different marginal-cost curve on workloads where agent call count is exploding.
Sources: Mercury 2 for Search (official blog, 2026-08-11), Mercury 2 reasoning blog (official blog, 2026-07-14), Introducing Mercury Coder (official blog), and arXiv 2506.17298 (Mercury technical report, referenced in the blogs).