We scraped the latest 178 official blog/news articles from 11 vendors—OpenAI, Anthropic, Google DeepMind, Meta AI, Microsoft Research, xAI, Mistral, as well as Qwen, DeepSeek, Zhipu GLM, and Moonshot AI (Kimi)—and ran readability, jargon density, and lexical richness analyses. The conclusion is somewhat counter-intuitive: DeepSeek, which reads the "easiest," actually packs the highest jargon density; the vendor whose writing style most resembles an academic paper is none other than DeepMind itself.
We didn't settle for subjective impressions like "who feels more obscure to read." Instead, we wrote a unified Python analysis script and ran three sets of metrics on each vendor's latest 10–20 official articles (after stripping navigation bars, footers, pure code blocks, and pure tables, keeping only the body text).
Readability: Flesch-Kincaid Grade (corresponding to the US school grade level—higher is harder) and Flesch Reading Ease (0–100—lower is harder to understand), applicable only to English texts—there is no mature equivalent formula for Chinese. Jargon density: we built a custom dictionary of ~60 English + ~40 Chinese high-frequency AI industry terms (agent, RAG, mixture-of-experts, multimodal, reinforcement learning…), counting the number of terms per 1,000 words. Lexical richness (TTR): unique word count / total word count; the closer to 1, the less word repetition.
Sample sizes vary by vendor: Mistral, xAI, Microsoft, and DeepSeek each have 20 posts; Anthropic, DeepMind, and Meta each have 15; OpenAI has 16; Zhipu GLM has 15; Kimi has 12; Qwen, due to rate-limiting during scraping, only yielded 10—these differences and finer limitations are detailed in the "Limitations" section at the end.
Among the 9 English vendors where Flesch-Kincaid scores could be calculated, Google DeepMind is the hardest to read (grade 16.3, equivalent to graduate-level reading), followed closely by Qwen (16.0) and Meta AI (15.5)—the common thread among these three is that their official blogs are packed with long compound sentences and passive voice, reading more like paper abstracts than press releases. Conversely, DeepSeek (11.1), xAI (12.3), and Mistral (12.5) read the "easiest", but this comes with a caveat—DeepSeek doesn't have a real blog; what we scraped were API documentation site update announcements and GitHub READMEs. This type of text habitually uses short sentences and lists to stack information, naturally pulling down the grade score.
* DeepSeek has no real blog; the sample consists of API update announcements + GitHub READMEs, a unique text genre—see main text for details.
A single-sentence comparison makes this more intuitive. Google DeepMind's sentence looks like this: "For example, if a robot was asked, 'Based on my location, can you sort these objects into the correct compost, recycling and trash bins?' it would need to search for relevant local recycling guidelines on the internet, look at the objects in front of it and figure out how to sort them based on those rules — and then do all the steps needed to completely put them away."—69 words, requiring the reader to mentally hold five or six subordinate clauses to finish it. In the same sample, DeepSeek's sentences look like this: "Rivals Gemini-3.0-Pro." "128K context for both." "Enhanced capabilities." "Reduced hallucinations."—four sentences totaling fewer than 20 words in telegraphic style. But note that "context" and "hallucinations" in these short sentences are themselves jargon—which is precisely the issue the next section will address.
This is the most counter-intuitive chart in the entire piece. DeepSeek's jargon density is the highest across the board (40.3‰, roughly 40 industry terms per 1,000 words), nearly 7× that of the lowest-density Kimi (6.1‰). Close behind are Qwen (33.3‰) and Microsoft Research (32.7‰)—the common thread among these three is that their "official text" is essentially closer to documentation written by a technical team for their peers, rather than press releases written by a PR team for the general public.
Cross-referencing with the readability ranking in the previous section reveals an interesting misalignment: DeepSeek reads the "easiest," but packs jargon the "densest"—short sentences + high-density professional terminology. Grammatically effortless, yet for those unfamiliar with the field, the information density is actually higher. This is a classic "telegraphic" writing style. Imagine: "V4-Pro adopts DSA sparse attention, 1M context enabled by default, Agentic Coding first surpasses closed-source SOTA"—you recognize every word, but string them together and it's all jargon.
"Memora sets new state-of-the-art on LoCoMo and LongMemEval, outperforming Mem0, RAG, and full-context inference while using up to 98% fewer context tokens."
"The reliability and safety we've prioritized in building Claude, paired with UST's experience in governance and regulated delivery, enables this work to move out of a pilot and into the systems that run a business."
Both are saying "our technology is strong/reliable," but Microsoft and Qwen choose to prove it directly with industry jargon, while Anthropic chooses terms like "reliability," "governance experience," and "moving from pilot to production systems"—words that an ordinary business reader can instantly grasp. The same message, two completely different allocations of "translation cost."
On the TTR (Type-Token Ratio, unique word count / total word count) leaderboard, DeepSeek (0.642) and xAI (0.560) sit at the top, seemingly the "most deliberate, least repetitive" in word choice. But this metric has a well-known statistical trap: the shorter the article, the naturally higher the proportion of unique words—a 200-word announcement rarely repeats the same word three times, whereas a 1,600-word deep technical explainer almost inevitably repeats core terms like "model," "agent," and "benchmark" over and over.
Viewing TTR alongside article length makes this clearer: DeepSeek (avg. 233 words/post), xAI (399 words/post), and Kimi (447 words/post) happen to be the three vendors with the shortest articles in the sample, and their TTR does indeed rank at the top—this is more likely a statistical illusion caused by length, rather than genuinely "richer diction." Conversely, Microsoft Research (1,579 words/post, TTR 0.392) and OpenAI (1,466 words/post, TTR 0.393) sit at the bottom of the TTR rankings, precisely because their articles are long enough that core terminology must appear repeatedly. If we must compare while controlling for length, the three vendors with similar word counts—Anthropic (943 words), Mistral (948 words), and Meta (967 words)—also have very similar TTRs (around 0.45), suggesting that at equal length, lexical diversity is actually quite comparable.
Merging the jargon word frequencies across 9 English vendors, the result is unsurprising—"agent" appeared 962 times, 2.7× the runner-up "eval" (359 times). Whether it's OpenAI's Agentic workflows, DeepMind's multi-agent safety research, or Meta's brand case studies, almost every vendor is desperately trying to prove they "understand agents best."
After merging the statistics for the two Chinese vendors, the top buzzwords are "reasoning" (118 times), "multimodal" (52 times), and "tool calling" (30 times)—clearly a different discourse system from the English list's "agent / eval / token": Chinese emphasizes capability-descriptive terms like "reasoning" and "multimodal," while English emphasizes product/evaluation paradigm terms like "agent" and "eval," indirectly confirming the divergence in terminology choices between the Chinese and English AI circles.
This assessment is a cross-referential comparison based on real scraped data, not a publishable corpus linguistics study. Several limitations should be noted:
Genres are not fully equivalent: Microsoft uses a Research Blog (academic-leaning), Qwen uses a GitHub Pages tech blog, DeepSeek has no blog at all, and Kimi's Chinese content mainly comes from developer platform product notes—these are not the same genre as the relatively uniform "company news pages" of OpenAI/Anthropic/DeepMind. The jargon dictionary is a custom-built list of ~100 Chinese and English terms, with limited coverage, using substring matching; cross-language absolute values should not be directly compared. Flesch-Kincaid uses heuristic syllable estimation, not an academically standard method; absolute values will have deviations, but the relative ranking between vendors remains valid as a reference. TTR is highly sensitive to article length; this article does not apply length normalization, so please do not interpret TTR in isolation from word count. Sample sizes are inconsistent (10–20 posts each), and all are from a single snapshot in early July 2026; vendor blog styles will shift with product cycle adjustments.
Returning to that counter-intuitive finding at the beginning—DeepSeek reads the easiest, yet packs the most jargon. Behind this lies a bigger question: when a vendor's "official blog" is actually a technical announcement written for peers, rather than a brand narrative for the public, jargon density measures not "how much this company loves industry slang," but "whether this company has a PR team to translate jargon into plain language." Chinese large model vendors generally lack a Western-style corporate blog tradition—whether this is a gap in PR maturity or a normal divergence under different technology dissemination cultures is a question left for readers to judge for themselves.
First published 2026-07-24