# Mercury 3 Mercury 3 is a diffusion LLM (dLLM) from Inception (inceptionlabs.ai), released in October 2026. Model ID: `mercury-3`. Inception calls it "the most capable diffusion-based LLM released to date", and the claim it really cares about is in the title of the launch post: diffusion scales. Why does that matter? Almost every [[Large Language Models (LLMs)|LLM]] you use is autoregressive. It writes one token, then the next, then the next, and each step waits on the previous one. A diffusion LLM starts from a noisy draft of a whole block and refines many tokens in parallel over a few denoising passes (the same idea as image [[Diffusion Models]], applied to text). Fewer sequential steps means much faster output. The open question since Mercury 1 (February 2025) was whether that approach could keep up on hard tasks once you scale it. Mercury 3 is Inception's answer: more than 40x the total parameters of Mercury 1 (mini), roughly 3x the "intelligence" on their aggregate benchmark score, and the same speed. On HN, Volodymyr (user volodia, Inception co-founder and chief scientist) made the argument I'd keep in mind: reasoning models think longer and longer per answer, and since diffusion produces those thinking tokens in parallel, the speed gap with an autoregressive model of the same size grows as reasoning gets longer. ## Specs and pricing - **Speed**: 1,068 output tokens/s on NVIDIA Blackwell GPUs, per Inception. Their table puts [[GPT-6 Luna]] at ~130 and [[Claude Haiku 5.5]] at ~240, hence the "4-8x faster" claim - **Context**: 1M tokens (Mercury 2.5 had 260K); up to 131,072 output tokens - **Price** (per million tokens): $0.20 input, $0.70 output, cached input $0.02. At launch it's 50% off: $0.10 / $0.35 / $0.01 - **Reasoning effort**: `instant`, `low`, `medium`, `high`. The default is `low` (Mercury 2.5 defaulted to `medium`) - **Features**: tool calling, structured outputs, text only. Temperature goes from 0.5 to 1.0 - **API**: OpenAI-compatible, on the Inception platform (100M free tokens), plus [[OpenRouter]], Baseten and [[Vercel]] for hosting. Streaming arrives block by block rather than token by token, and an optional "diffusing" mode shows the text being denoised - **Weights**: closed. Inception isn't releasing them and won't disclose the model size Luna and Haiku 5.5 both cost $0.10 / $0.50, so even at the regular price Mercury 3 charges twice as much for input. At the launch price it's the same input cost and 30% cheaper on output. One HN commenter (2001zhaozhao) pointed at the real difference: Haiku 5.5 charges five times more once a prompt goes over 100K tokens, while Mercury 3 keeps the same price up to 1M. ## Benchmarks All three models at `medium` effort, published by Inception: | | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 | |---|---|---|---| | GDPval (Elo) | 1,237.2 | 1,204.0 | 1,225.0 | | τ³-bench Banking | 30.24 | 23.00 | 13.06 | | MultiChallenge | 60.35 | 51.49 | 52.48 | | AA-Omniscience-Public | 46.44 | 46.94 | 37.22 | | IFBench | 76.84 | 61.82 | 54.65 | | GPQA Diamond | 89.23 | 86.36 | 84.51 | | AA-LCR v1.1 | 77.67 | 78.30 | 74.70 | | MRCR v2 | 75.50 | 62.72 | 78.40 | | SWE-Bench Atlas | 51.08 | 33.33 | 32.26 | | DeepSWE v1 | 46.31 | 37.76 | 51.62 | | BrowseComp+ | 55.18 | 50.50 | 45.78 | | DeepSearchQA | 71.33 | 60.00 | 65.22 | Mercury 3 leads on 8 of 12. Luna wins knowledge accuracy and long-context reasoning by a hair; Haiku 5.5 wins MRCR v2 and DeepSWE. Don't compare these with Anthropic's own Haiku 5.5 table: Anthropic reports GDPval-AA v2.1 at 1620 Elo, so the versions or setups clearly differ. ## Who uses it, according to Inception Early-access partners quoted in the launch post: - **TinyFish** (web agents): 50% of 50 live Mind2Web navigation tasks, "significantly more than GPT-6.1-Sol", with the same tokens and steps at a fraction of the per-token cost - **RevisionDojo** (tutoring): hints, marked answers and definitions in about a second, and "a drop-in swap for our existing stack" - **Octen** (search on smart glasses): median generation time of 0.9 s and median end-to-end response of 2.2 s - **Lorikeet** (customer support in regulated businesses): better latency and higher eval scores than other models in early testing All four are products where something waits on the model at every step (an agent, a student, a caller), so latency compounds. That's the niche Inception goes after. Two sibling products target the same thing: Mercury Voice (optimized for time to first answer token, enterprise only) and Mercury Decide, a [[Decision Models (DMs)|decision model]] that Inception says beats [[Jev]] on JevBench and LegalBench. ## What HN testers found The HN thread (255 points, 91 comments) had quite a few people trying it on the spot: - **It's good for its class.** UncleOxidant tried it on a repo and was "pleasantly surprised". GodelNumbering got a working HTML renderer for UML diagrams and correct follow-up edits. Escapado built a small SwiftUI OCR-and-translate app end to end in 8-9 steps (about 45 minutes, 26M tokens, 1.3M uncached) using Oh My Pi as the harness; it one-shotted everything except a coordinate calculation it fixed when told. ilitirit noted it pushed back on his opinions instead of agreeing - **The 1,000 tokens/s isn't what people saw.** Escapado measured ~160 tokens/s, ilitirit 300-400 after a slow start. minraws and snake_doc both flagged a slow time to first token, and snake_doc found DeepSeek V4.1 Flash beating it on latency and intelligence on OpenRouter - **Pelicans.** [[Simon Willison]] ran his pelican-on-a-bicycle SVG test via OpenRouter and the Inception API: "not bad, a little quirky". One reply noted that the `medium` pelican looked better than the `high` one, which may be why the benchmarks were run at medium - **The fast loop is the feature.** ghrl used it in [[OpenCode]] for "grilling" sessions where the next batch of questions shows up within seconds, and it can write a 600-line HTML prototype per option in about 2 seconds to help you decide Volodymyr's own framing was modest: Haiku- or Luna-level in daily use, not a replacement for Opus as the main model driving an agent, but good for [[AI Subagents]] (they used it to implement kernel optimization ideas proposed by a bigger model). ## Critiques - **The scaling chart** plots quality on a linear axis against parameters on a log axis. arendtio's comment: that combo has "a high probability of looking fantastic". The quality score is also Inception's own average of 11 benchmarks - **Batching economics.** Several commenters explained why [[Gemini|Google]] went quiet after Gemini Diffusion: diffusion doesn't get the same memory-bandwidth savings from batching many users together, so it shines at low concurrency and gets expensive per token at scale (clhodapp, citing a Google talk; LarsDu88; willy_k). famouswaffles added that [[DiffusionGemma]] fell further behind its autoregressive sibling as tasks got harder. Volodymyr's reply: they've been "happy with how performance has scaled with batch size", and the launch prices are what that allows - **No size, no paper, no weights.** Questions on model size and training got "not publicly announcing", with technical reports promised later. SillyUsername had been waiting for open weights; for local diffusion, DiffusionGemma is the main option, and nolist_policy pointed out that llama.cpp doesn't support it ## My take I care more about the scaling curve than about any single benchmark row. If dLLM quality keeps rising with scale while the speed stays put, the cost of an agent step drops on two axes at once: price per token, and wall-clock time per step. For agent loops, I'd argue the second one matters most. A model that answers in one second instead of eight lets an agent try several approaches in the time it used to take for one. I'd use it where Luna and Haiku 5.5 already fit: subagents, extraction, routing, tool-calling loops, voice ([[Model routing]]). The 1M context at a flat price is a real argument against Haiku 5.5's over-100K rates. But I'd measure throughput and time to first token on my own traffic before believing 1,068 tokens/s, and I'd run it side by side with the current model before swapping anything ([[Shadow Evaluation]]). Every number above, except the HN anecdotes, comes from Inception. ## Caveats - Benchmarks, speed figures and partner quotes come from Inception's launch post. Independent testing so far is anecdotal (HN users, Simon Willison's pelicans) - Measured speeds on launch day were 160-400 tokens/s, well below the advertised 1,068 - The launch discount ends after an unspecified period; regular input price is twice Luna's and Haiku 5.5's - Model size and training details are undisclosed ## References - Inception, "Diffusion at Scale: Introducing Mercury 3": https://www.inceptionlabs.ai/blog/introducing-mercury-3 - Inception docs, Models (pricing, context, limits): https://docs.inceptionlabs.ai/get-started/models - Inception docs, Reasoning Efforts: https://docs.inceptionlabs.ai/capabilities/reasoning-efforts - Inception docs, Streaming & Diffusion: https://docs.inceptionlabs.ai/capabilities/streaming - Inception docs, integrations (linked by the co-founder on HN): https://docs.inceptionlabs.ai/integrations/overview - Hacker News discussion (item 50155245): https://news.ycombinator.com/item?id=50155245 - Volodymyr (Inception co-founder), launch comment on HN: https://news.ycombinator.com/item?id=50155371 ## Related - [[Diffusion Models]] - [[DiffusionGemma]] - [[Large Language Models (LLMs)]] - [[GPT-6 Luna]] - [[Claude Haiku 5.5]] - [[Decision Models (DMs)]] - [[Jev]] - [[AI Subagents]] - [[Model routing]] - [[AI Inference]] - [[AI Reasoning Models]] - [[Context Window]] - [[OpenRouter]]