LONG AI
Home › Blog › Inception Mercury 3 Ships: Diffusion LLM

Inception Mercury 3 Ships: Diffusion LLM at 40x Scale, 1M Context, 254 tokens/s Measured

Category:行业资讯 · 9 min read · 2026-10-21

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Diffusion language model parallel generation illustration

Headline: Inception Mercury 3 (released Oct 20)

Inception, based in Redwood City, California, released Mercury 3. It's the company's fourth-generation model since Mercury 1 in June 2025, and currently the largest diffusion language model (dLLM).

It works differently from the familiar ChatGPT-style models: mainstream models write "one token at a time," while diffusion models "sketch a whole draft first, then refine many positions at once," like repeatedly de-noising an image in a paint program. The upside is parallelism; the downside, historically, is that it couldn't scale. Mercury 3 aims to prove that diffusion models can scale without losing speed.

Key figures (vendor claims, not yet independently reproduced)

Third-party test (partially replicated)

A third-party single-channel test measured a median of about 254 tokens/s - roughly a quarter of the vendor claim, so treat "1,068" as a best-case lab number rather than what you'll see in practice.

Other releases (6)

ModelDateTypeHighlights
Mistral Voxtral 310-20Voice model family + Agent SDK (commercial)Ears, mouth, and wiring in one shot: Small / Mini (4B) / Mini Realtime three transcription tiers, plus Voxtral Speech 3 synthesis (35+ languages, 50+ dialects), plus a LiveKit-based voice Agent SDK for calling systems and tools mid-conversation. Transcription covers 50+ languages with speaker diarization, hotword biasing, and endpoint detection. Vendor figures: macro-average WER 4.1% across 45 FLEURS languages, speaker diarization cpWER 18.4% (vs Gemini 3.5 Transcribe 38.4%); transcription from $0.0035/min, synthesis $13 per million characters, realtime under 500ms latency. EU-hosted or private deployment.
Ant Bailing Ming-3.0 voice series (open source)10-20Open weights, 5 modelsThree lines, five models: tts-0.5B / tts-tiny (7.9B total / 1.3B active), audio-flash (126B / 6.5B active) / audio-tiny (9.3B / 2.7B active), asr-tiny (8.6B / 2B active). TTS supports a "director mode" - natural-language control of emotion, stress, and rhythm, plus paralinguistic tags like [Laughter]. Third-party Artificial Analysis: tts-0.5B English robustness 93.2% (tied first), Controlled Voice 1069 Elo (first among open weights); vendor: audio-flash BigBenchAudio 99.40 first, asr-tiny 4.41% average error across four scenarios. Weights on Hugging Face / ModelScope.
DeepSeek DeepInfer (open source)10-20Open inference frameworkOpen-sources the entire inference framework that serves V4.1 Flash: distributed inference, speculative decoding, SWA Bounded Replay, structured output, compatible with NVIDIA and Huawei Ascend; ships deepseek-recipe aligned with the official API's token codec. Disclosed: single-cluster throughput 155T tokens/day, $38/hour per card at off-peak power pricing. Prefill and Decode split deployment, Decode on 64-128-card super-nodes.
Bark 3.0 (Conway Research)10-21Open weights (CC-BY 4.0)Only 163MB, built on NVIDIA Parakeet, vendor-claimed on par with Parakeet on average. A tiny voice model for edge, offline, and on-device use.
Inception Mercury Voice / Decide10-19/20Commercial APIAlongside Mercury 3: Mercury Voice under 170ms first-token latency for voice bots; Mercury Decide for high-concurrency classification. Fills the "realtime voice + batch decision" gaps.

Industry news

Realtime voice agent illustration

The trend in one line

Diffusion models have crossed the scale threshold - the architecture that was once stuck at small sizes is now competing head-on with autoregressive models on both speed and capability.

(Data from official announcements and public reports, 2026-10-21; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa