The loudest number at iFlytek's 2025 developer festival was not a parameter count. It was a claim that the model understands you — your languages, your intent, even your voice cloned from a single sentence.
On 6 November 2025, at the 8th World Voice Expo in Hefei, iFlytek (科大讯飞) unveiled Spark (星火) X1.5, its latest deep-reasoning foundation model (基础模型). The headline engineering fact is quieter but more strategic: the model was trained and inferred entirely on domestic compute (全国产算力) — zero Nvidia GPUs in the stack.
What X1.5 is
Spark X1.5 uses a Mixture-of-Experts (混合专家) architecture with 293 billion total parameters and 30 billion active per token. That is roughly half the total size of the previous generation while keeping the active count small enough to deploy on a single domestic server.
iFlytek reports reasoning efficiency up more than 100% versus the earlier X1, and says the model's six core abilities — language understanding, text generation, knowledge Q&A, logical reasoning, math, and code — now benchmark against international leaders. The company itself claims overall performance at roughly 95% of GPT-5 across its evaluations, with 14 focus languages (Latin America, ASEAN) leading; that is iFlytek's own benchmark statement, not an independent ranking.
The "understand you" thesis
Where many labs compete on size, iFlytek leans on its two decades in speech and education:
- 130+ languages supported, with strong coverage of lower-resource ones.
- A world-first non-autoregressive speech-model architecture (per iFlytek), which it says cuts inference cost by a claimed 520% versus comparable autoregressive models.
- Personalised memory — the model builds a user profile across sessions to tailor replies.
- Voice cloning from one recording and a digital-human "Xiaofei (小飞)" that handles multi-person, multi-language interaction.
The pitch is that winning AI is not just smarter but more attuned: understanding emotion and intent, not only words.
Trained on domestic silicon
iFlytek says X1.5 was trained on a fully home-grown compute stack, with two hard problems solved along the way:
- Long chain-of-thought reinforcement learning, where training efficiency rose from about 30% to above 84%.
- Full-link training efficiency for the MoE model itself.
Running entirely on domestic accelerators is both a technical and a commercial stance. It insulates iFlytek from export controls and lets it market a "sovereign AI" story to Chinese enterprises and governments wary of foreign dependency.
Why a speech company is betting here
iFlytek's moat is not compute or parameters; it is two decades of speech data, acoustic models, and real products in classrooms, clinics, and cars. That gives it something pure model labs lack: distribution and feedback loops. A model that reads a child's handwriting, transcribes a noisy clinic, and clones a voice from one sample is monetising capabilities most labs only benchmark.
The risk is the opposite — that "understanding you" becomes a soft claim hard to defend against labs shipping larger, cheaper, open weights. iFlytek's answer is to stay close to verticals where accuracy and compliance matter more than leaderboard tops.
Where it actually ships
iFlytek's strength is applied, not abstract. The same underlying model powers:
- Education — a 4,000-plus-tag error-diagnosis system; an AI grading machine that takes a class from 60 minutes to 10 minutes of correction time.
- Healthcare — the medical large model reached what iFlytek calls "chief-physician-level" performance in graded-hospital pilots; its "智医助理" assistant reportedly lifted diagnostic reasonableness from 87% to 96%.
- Automotive — the iFLYSOUND cockpit audio system, in production at 19 car brands with over 1 million units shipped.
- Developer ecosystem — iFlytek open-platform developer count reached 9.68 million (≈ 2 million added in the prior year), and it open-sourced Astron, an agent platform with native RPA support.
Honest limitations
- The "95% of GPT-5" and "520% cost reduction" figures are iFlytek's own claims from its release; independent reproduction was not verified here.
- "Trained entirely on domestic compute" is a company statement; the specific accelerator vendors and cluster size were not detailed in the sources reviewed.
- The 293B/30B parameter counts and 130+ language coverage come from Chinese media relaying iFlytek's briefing; no external audit was available.
- Medical claims (chief-physician-level, 87%→96%) describe pilot or assistant settings, not replacement of licensed physicians, and were not independently confirmed.
- This article covers the X1.5 release (November 2025); later Spark revisions may have shipped since.
What readers can do now
- Test the multilingual edge — if your product serves ASEAN or Latin American users, trial Spark X1.5's language coverage against your current model; iFlytek's focus languages are where it claims to lead.
- Pilot the open-source Astron platform — for teams building RPA-style agents on domestic infrastructure, the open-sourced platform is a concrete starting point.
- Demand independent benchmarks — given that the headline numbers are vendor-reported, run your own eval before trusting "95% of GPT-5" for any production decision.