Frontline Hotspot
Frontline Hotspot

iFlytek Spark X2.6: A 293B Flagship, but Only a 4B Is Open

iFlytek launched Spark X2.6 at its global 1024 Developer Festival in Beijing on October 22, 2026: a 293B-parameter MoE that activates 30B per token, a 1M-token context window, an end-cloud integrated design, trained entirely on domestic compute and served from a 10,000-card domestic inference cluster. The efficiency ledger, all vendor-reported: training 2000B tokens of 256K-length data on the Feixing-2 platform fell from 4.5 months to 1.2 months (a 4x gain), and the first fully domestic four-layer KV Cache architecture lifted hit rate from 63 to 95.7 percent with 4.5x inference efficiency. Self-reported scores put SWE Pro at 77.3 and SWE Multilingual at 95.0, both first, with SWE Verified at 95.4 behind only DeepSeek-V4.1-Flash 1017's 95.6 - yet there is no complete score table, no technical report, and no third-party replication. The key correction: the viral headline 'iFlytek open-sources X2.6' misreads the release. Only the on-device Spark-X2.6-4B is open under Apache 2.0 (Hugging Face under the XHToken namespace, ModelScope mirror, Ollama as SparkLLM), while the 293B flagship stays API-only. The 4B self-reports six gains including BFCL-V4 71.8 and SWE Verified 62.8. Launch pricing is 0.9 / 0.08 / 3.6 yuan per million tokens. The narrative layer predicts 60 percent local, 30 percent hybrid, 10 percent cloud workloads with routing accuracy up from 81.2 to 96.1 percent; the chairperson's 'only model trained on fully domestic compute' is a unilateral claim with no chip model disclosed. Unpublished: chip models, cluster size, total FLOPs, technical report, flagship weights.

Published October 23, 20268 min read
<!-- iflytek-spark-x2-6-hotspot | hotspot | iFlytek Spark X2.6: A 293B Flagship, but Only a 4B Is Open -->

On October 22, 2026, at the main forum of the iFlytek Global 1024 Developer Festival in Beijing, chairman Qingfeng Liu unveiled Spark X2.6, the company's new flagship model. The spec sheet is aggressive by any standard: a Mixture-of-Experts architecture with 293B total parameters and 30B active per inference, a 1M token context window, training carried out entirely on domestic Chinese compute, and deployment on a 10,000-card domestic inference cluster (all figures are launch-stage claims as reported by Global Times, Zhi Dongxi, and other outlets).

But the way this launch spread through social feeds introduced a distortion that shapes how most people understand it: many headlines, especially in Chinese-language coverage, read "iFlytek open-sources Spark X2.6." The reality is much narrower. The only model with published weights is Spark-X2.6-4B, an on-device small model released under Apache 2.0. The 293B cloud flagship's weights are not open; the Spark website offers API trials, nothing more. In other words, the thing that is actually open source is a 4B sidekick, not the 293B flagship headlining the event.

This distinction is not pedantry. The open-sourcing scope determines what a developer can actually take home: what you can deploy on your own hardware is the 4B model, and what you can wire into production is a paid flagship API. Those are two different products wearing one name. This article walks through the launch, the flagship specs and the domestic-compute math behind them, the open-sourcing scope, and the pricing narrative, and closes with a list of things iFlytek did not announce.

What iFlytek Actually Shipped

Spark X2.6 was only the centerpiece of a packed launch event. The same stage introduced SparkAudio 1.0, a speech foundation model; embodied multimodal interaction technology; the WinWork digital-employee platform; iFlytek Xingguang, a teaching assistant; and three hardware products: an AI recording ring at 1,299 yuan and AI glasses starting at 4,299 yuan, plus two assistant apps, Spark Loomy for everyday office work and Spark AStudio for professional coding (per Zhi Dongxi's coverage).

The bundle reflects a coherent industrial narrative. On stage, Liu said that "AI is becoming the most disruptive technological and industrial revolution in human history, bar none." He also offered a line worth remembering: "A fully self-controllable large model stack is the inevitable choice for AI security and ultimate cost efficiency, and that requires end-cloud unified large models with a fully self-controllable stack." Together, these two sentences capture the core of iFlytek's route: model, chips, and on-device hardware need to move as one.

There is a policy backdrop here. iFlytek was placed on the US Entity List in October 2019 and remains on it (Federal Register 84 FR 54002). For this company, "domestic compute" is not marketing rhetoric; it is a survival route it was forced onto seven years ago, now repackaged as an industry-wide selling point.

The Flagship: 293B-A30B and the Domestic Compute Math

Start with the architecture. A 293B total / 30B active MoE design follows the current flagship playbook: large total parameters for knowledge capacity, sparse activation to keep per-inference cost down. The 1M token context window means an entire code repository or several hundred pages of documents fit into a single prompt. The cloud version is available for trial on the Spark website — trial, note, not open weights.

The informative part is the efficiency accounting on both the training and inference sides. On training, iFlytek says a "long-horizon task dynamic parallel training technique" solved the low mixed-training efficiency that domestic chips' memory bandwidth imposes on long tasks. The worked example: training on 2,000B tokens of 256K-length data on the "Feixing-2" domestic supercomputing platform's 10,000-card cluster took 4.5 months before optimization and 1.2 months after, a fourfold gain in ultra-long-context training efficiency (official figures relayed by Zhi Dongxi and Tencent News). Liu also said Spark X2.6 was trained "over the past two to three months."

The inference-side numbers are equally specific: distributed inference on heterogeneous domestic compute, a four-layer KV Cache architecture that iFlytek describes as the first implemented on fully domestic hardware, inference efficiency on Feixing-2 reaching 4.5 times the pre-optimization level, and the KV Cache hit rate rising from 63 percent to 95.7 percent.

One trap worth flagging: coverage mentions "10,000-card" in two places, and they mean different things. Global Times' "10,000-card inference cluster" refers to X2.6's inference deployment; Zhi Dongxi's "Feixing-2 10,000-card" refers to the training platform. Both are four-digit card counts, but they describe two separate systems. This article keeps them distinct, and so should any citation of these figures.

As for flagship scores, the launch-stage figures are: 77.3 on SWE Pro, ranked first; 95.0 on SWE Multilingual, ranked first; and 95.4 on SWE Verified, second only to DeepSeek-V4.1-Flash 1017's 95.6 (Tencent News relaying the on-screen data; official claims). On the agent side, enterprise-operations agents and the general tool-calling benchmark MCP-Atlas are described as first-tier; on reasoning, AA-LCR v1.1 is on par with MiniMax M3. All of these scores come from the launch event itself. iFlytek published no complete score table and no technical report, and no third-party retest exists at this time.

What Is Actually Open: Only the 4B

Now the central clarification. What shipped under the open-source banner is the on-device Spark-X2.6-4B, under Apache 2.0. The 293B flagship's weights are not open — the English-language outlet TSN put it bluntly: the flagship's weights are not.

The 4B model lives in three places. On Hugging Face, the weights sit under the XHToken namespace (see the official release post), repo address: huggingface.co/XHToken/Spark-X2.6-4B. For downloads inside China, the ModelScope page is confirmed live with an INT4 quantized build, repo address: modelscope.cn/models/XHToken/Spark-X2.6-4B-INT4. For one-command local runs, use Ollama:

bash
ollama run SparkLLM/Spark-X2.6-4B

Note that two namespaces coexist: HF and ModelScope use XHToken, while Ollama uses SparkLLM — do not mix them up when citing links. According to the model card, the release includes BF16 main weights plus GGUF and INT4 variants, with support for Transformers, SGLang, vLLM, MLX, and llama.cpp, and deployment guides for both NVIDIA GPUs and Huawei Ascend NPUs (the SGLang Ascend image, cann9.0.0).

On specs, the 4B model has roughly 4.1B parameters and is architecturally identical to its predecessor X2.5-4B: 36 layers, with a 3-to-1 ratio of sliding-window attention layers to full attention layers, and a 1,048,576-token context. The capability gains come from post-training, not structural change. Three thinking tiers (none, low, high) are available, and loading requires trust_remote_code. One citation detail: the ModelScope page lists the INT4 build at 1.97B parameters and 3.43GB — that is the post-quantization figure, while BF16 full weights are about 4.1B. The two numbers coexist because of quantization, not model shrinkage.

There is also an interesting timeline wrinkle: the weights went live before the event. SparkLLM announced the release on X on October 20 (Tuesday evening), two days before the developer festival opened on October 22. Strictly speaking, the 4B model was listed first and showcased second; the keynote gave it a formal debut ceremony after the fact.

The 4B benchmarks are likewise all self-reported (thinking=high tier, no third-party retest): BFCL-V4 71.8 (predecessor 65.1), MCP-Atlas 62.4 (54.6), BrowseComp 53.1 (40.9), SWE Pro 54.3 (44.4), SWE Verified 62.8 (41.6), and AA-LCR v1.1 67.3 (64.7). Official claims place it best-in-class among open models of similar size, with some benchmarks matching or exceeding roughly 9B models, and inference efficiency on par with or better than Qwen3.5-4B and Gemma-E4B. For context, the predecessor X2.5-4B was open-sourced on September 1 and has been downloaded 2.56 million times, placing it among the top four on-device models globally (official figures). This 4B update is, by that standard, an incremental iteration.

Pricing and the 60/30/10 Story

The flagship API ships with limited-time pricing: 0.9 yuan per million input tokens, 0.08 yuan per million for cache hits, and 3.6 yuan per million output tokens. Zhi Dongxi reports that both the input cache-miss price and the output price undercut DeepSeek-V4.1-Flash-1017. Low pricing plus a 1M context window aims squarely at long-document, long-session enterprise workloads.

On the industry narrative, Liu offered a forecast for how workloads will split: 60 percent of future tasks running locally, 30 percent in end-cloud collaboration, and 10 percent purely in the cloud. Underpinning that split is a routing mechanism — after introducing a commentary-model-based end-cloud routing co-training mechanism, routing accuracy rose from 81.2 percent to 96.1 percent (official figures), and the on-device model's agent task coverage stands at 41.9 percent. In his words: "For large models to go deep into business, they must handle complex tasks while keeping data boundaries intact and costs down."

One widely quoted claim deserves its own calibration. As relayed by the English outlet TSN, Liu said Spark is the only model trained entirely on domestic compute, that all other models are trained on NVIDIA chips, and that iFlytek is the smallest of the bunch. Two caveats must accompany that claim: first, it is a one-sided statement — iFlytek named no comparison targets, and TSN flagged it as unverified; second, iFlytek has not disclosed which domestic chips were used for training. The concurrent industry backdrop complicates the "only" framing too: DeepSeek-V4-Pro has been reported to complete post-training on Huawei Ascend 910C, while Reuters reported ByteDance, Alibaba, and Tencent ordering over 3 million NVIDIA RTX PRO 5500 units. Faced with these public data points, "only" deserves a question mark.

On the financial side, iFlytek's first three quarters of 2026 brought 18.187 billion yuan in revenue, up 7.05 percent year over year, with net profit attributable to shareholders of 21.89 million yuan — back to profitability — though non-GAAP net profit was negative 763 million yuan. Q3 R&D spending exceeded 1.7 billion yuan, over 25 percent of quarterly revenue. A company still losing money on an operating basis is betting a quarter of its revenue on R&D, which explains why this "domestic compute plus end-cloud unification" story has to land.

What Was Not Announced

Setting the launch-event noise aside, the list of things iFlytek did not disclose is not short:

The specific domestic chip models used for training: not announced. The total scale of the training cluster: not announced. Total training compute in FLOPs: not announced. A technical report: not published. The flagship 293B weights: not released. Third-party leaderboard results (LMArena and similar): none exist yet.

That means every efficiency number above — 4.5 months cut to 1.2, a 95.7 percent KV Cache hit rate, 4.5 times inference efficiency — has to be filed under "official claims," unverifiable by outsiders. The flagship's first-place rankings rest on a benchmark subset iFlytek itself selected; the missing complete score table is itself a signal. Liu's response was to change the venue: real-world industrial deployments (state-owned enterprises, education, healthcare, judicial systems) outperform public benchmark scores, he said. That claim can be neither falsified nor verified — a classic industrial-narrative hedge.

Third-party analysts frame it more neutrally: the industry's focus is shifting from parameters and benchmark scores toward domestic-compute deployment, data security, and real-world usability (analyst Liu Dingding, via Global Times). By that standard, Spark X2.6's real value lies not in leaderboard rank but in offering a positive engineering answer to the question of whether large models can be trained efficiently and served cheaply on domestic chips — provided those numbers survive later verification.

The actionable part for developers, meanwhile, is settled: the 4B weights are downloadable now, Apache 2.0 carries no extra conditions, and the Ascend NPU guide gives domestic hardware an official deployment path; the flagship can be tested against your own long-context scenarios at the limited-time price. Judge capability on your own datasets — that beats waiting for any leaderboard.

Closing

Spark X2.6 was a launch event flattened by its headlines. The real structure has three layers: a 293B flagship with aggressive specs and aggressive pricing but closed weights and incomplete evidence; an open 4B on-device model with a clean license and a clear target, but entirely self-reported scores; and a "10,000 domestic cards" infrastructure narrative with vivid numbers and missing key parameters. Understanding those three layers matters more than remembering any single benchmark score. When the chip models, cluster scale, and technical report surface — or when a third-party leaderboard appears — we will come back to cover the verification half.

This article is AI-assisted and human-edited. Last updated: 2026-10-23

Related

Frontline Hotspot

DeepSeek Ports the NVIDIA Stack to Ascend: Four Repos, One Day

On 2026-09-30 DeepSeek open-sourced its Huawei Ascend infrastructure components (official basis relayed via AI-tool aggregators): a full kernel stack mapped one-to-one to its NVIDIA-side repos - TileKernels (TileLang-based operator library) gaining an Ascend backend selected automatically at runtime so one Python API runs on both GPU and NPU, DeepGEMM-Ascend (fully API-compatible, BF16/FP8/FP4 GEMM and MegaMoE), DeepEP-Ascend (EPBuffer V2.5-compatible), and FlashMLA Ascend sparse-attention kernels (410 TFlops prefill at 95% of hardware peak, 360 TFlops decoding at 83%, with a deep-dive technical report, README official basis). "Multiple tests near hardware limits" is the official claim. Cold thought: DeepEP-Ascend ships no LICENSE file as of the snapshot, some fused kernels are CUDA-only, and the validated stack centers on Ascend 950 series.

Oct 1, 20269 min read
Frontline Hotspot

280B Total, 14B Active, Three Regions Shut Out

On November 5, 2026 Baidu released ERNIE 5.5 Flash Base and published its weights: 280B total parameters with 14B activated per token, pretrained from scratch on 33.5T tokens, 128K training context and 1M for fine-tuning and inference, 43 layers with 37 heavily compressed attention layers plus 6 MLA-DSA layers, 512 routed experts with top-10 selection and one shared expert, and FP8 plus BF16 checkpoints on Hugging Face and ModelScope. This is an end-of-pretraining Base checkpoint with no mid- or post-training, not a chat model; a post-trained version is promised without a date. The custom ERNIE 5.5 Flash Base Model License Agreement v1.0 excludes the European Union, the United Kingdom and South Korea from its territory, allows free non-commercial use, requires a prominent Powered by ERNIE notice and a one-time notice for commercial use up to 50 million dollars in related annual revenue, requires separate prior written authorization above that threshold, and bans using outputs to improve unrelated AI systems.

Nov 6, 20269 min read
Frontline Hotspot

Kimi K3.1 Pro Weights Are Out: 2.8T Parameters, 1.56TB

At midnight Beijing time on November 5, 2026, Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face, honoring the promise made when the model launched on October 23. The model has 2.8 trillion total parameters with about 104B activated per token, a 1048576-token context, 69 KDA plus 24 Gated DSA attention layers, a 401M vision encoder, and MXFP4 weights with MXFP8 activations under quantization-aware training, totaling roughly 1.56TB. The vendor-reported Terminal-Bench 4.0 score is 65.7; on the third-party Agent Arena of real user sessions it ranks first among open models with a net improvement rate of 5.29 percent, ahead of MiMo-V2.6-Pro at 2.68 percent (preliminary). The Kimi K3.1 Pro License requires a separate agreement for Model-as-a-Service businesses exceeding 20 million dollars in trailing 12-month revenue and prominent naming for products above 100 million monthly users or 20 million dollars monthly revenue.

Nov 5, 20269 min read