Frontier Labs

iFlytek's third-quarter report shows a profit of 226 million yuan for the quarter and, with one-off items stripped out, a loss of 126 million; over nine months it lost 763 million yuan before one-offs, and government grants of 471 million covered 62 percent of that core loss. The company gives one reason in the filing and on its 23 October earnings call: R&D spent on staying on the self-reliant technology route, which for iFlytek means training its Spark models on Chinese chips. At its developer conference on 22 October, chairman Liu Qingfeng said domestic compute is still nearly two generations behind Nvidia, short on memory and bandwidth, and that only Huawei and iFlytek are doing this in earnest. The call record says iFlytek will not spend hundreds of billions on infrastructure or absorb billions a year in losses, so its 2026 training centres on a 4-billion-parameter on-device model and models of a few hundred billion parameters. Inventory rose from 3.03 billion yuan to 4.79 billion, the call record names memory-chip stockpiling among what pushed cash out of operations, and bank loans and bonds went from 4.69 billion to 7.56 billion.

DeepSeek's DeepInfer deployment notes say one cluster serves 155 trillion tokens a day for revenue of 38 dollars per GPU-hour at off-peak prices. Its February 2025 disclosure works out to about 12.91 dollars, so the figure has roughly tripled while prices fell. DeepSeek does not give the cluster's GPU count, chip, or cache-hit share, and the revenue figure depends on all three.

MiniMax M3.1 Flash, Zhipu's GLM-5.3-Flash and DeepSeek's V4.1-Flash-1017 at off-peak rates all list input at 15 cents a million tokens. Cached input, most of an agent's bill, runs from 0.3 cents at DeepSeek off-peak to 3 cents at Zhipu. Third-party hosts already serve DeepSeek's and Zhipu's models for less; MiniMax has not released M3.1 Flash's weights, so only MiniMax can sell it.

On 15 October DeepSeek's chat product, V4 Pro API and V4.1 Flash API failed together for 4 hours 52 minutes, covering the whole weekday morning peak. DeepSeek is returning every API account's spend from 8 to 14 October as gift credit and has pledged to strip out single-point dependencies. Twelve of the thirteen incidents on its status page since 2 September came after V4.1 Flash launched on 10 September.

On 14 October the official account of ByteDance's Doubao phone assistant published 151 apps its agent will no longer operate. Sixty-five belong to Tencent and its affiliates and twenty-three to Alibaba's; the list also takes in ICBC, China Merchants Bank, all three state carriers, DeepSeek and Kimi. The same day the 30-day notice period ended, so a third-party app whose developer never replied is now being opted in by default.

On 10 September Moonshot AI named the first five partners in its Kimi enterprise partner program: Teamsun, Kingsoft Cloud, Yakang, AsiaInfo Technologies and Chinasoft International. Moonshot supplies the models, the Hosted Agents runtime and the engineering method; the partners supply the industry knowledge, the customer relationships and the engineers who sit in the customer's office, and the two sides are bound by a share of token and service revenue. All five are listed, so their first-half numbers are filed. For the six months to 30 June, Chinasoft International reported 8.95 billion yuan of revenue at a 21.6 per cent gross margin and 350 million yuan of profit; AsiaInfo 2.20 billion at 22.0 per cent and a 477 million yuan loss; Kingsoft Cloud 5.78 billion at 14.1 per cent and a 437 million loss; Teamsun 2.50 billion and a 245 million loss; Yakang, the smallest at 825 million, a 50 million profit. Three of the five lost money, 1.159 billion yuan between them in six months, about 173 million dollars at the onshore rate of 6.7050. The three Hong Kong names taken together ran a 19.1 per cent gross margin on 16.9 billion yuan of revenue. AsiaInfo is the one to look at hardest: its gross margin was 37.4 per cent across 2024 and 30.1 per cent in the first half of 2025 before reaching 22.0 this half, while the loss widened from 198 million yuan to 477 million. Teamsun moved the same way faster, from a 140 million yuan profit in the first half of 2025 to a 245 million loss on revenue up about eleven per cent. The same day, Tencent Cloud launched a forward deployed engineer certification on its ADP agent platform and opened partner recruitment alongside it, while ByteDance hires the role onto its own payroll at Volcano Engine. OpenAI reached the opposite answer, standing up a separate Deployment Company on 11 May with four billion dollars of initial investment from nineteen investors at a fourteen billion dollar valuation and buying the consultancy Tomoro for roughly 150 forward deployed engineers. The five partners turned over 20.2 billion yuan in the first half, about 3.02 billion dollars of enterprise IT services revenue, and that is the pool Kimi is now attached to.

DeepSeek created a GitHub repository called DeepGEMM-Ascend at 15:49 UTC on 29 September and a second one, DeepEP-Ascend, nine hours later, then posted to WeChat on 30 September that it had open-sourced six infrastructure components for Huawei's Ascend platform, each corresponding to something it already ships for Nvidia. Both new repositories carry benchmark tables and the two halves do not agree with each other. DeepGEMM-Ascend, a matrix multiplication library measured on an Ascend 950DT running CANN 9.20, reports a 4096 by 7168 by 16384 multiply in BF16 at 2,229.9 microseconds for 431 TFLOPS against a stated limit of 432, or 99.8 per cent; FP8 against FP8 gets 861 of 865 and FP4 against FP4 1,701 of 1,730. DeepEP-Ascend, which moves tokens between chips for mixture-of-experts routing, sustains 373 to 375 GB/s on dispatch and 345 to 347 on combine at EP8, falling to 313-320 and 272-278 at EP128, and its README states that dispatch holds 90 to 95 per cent of the physical payload limit only for expert-parallel sizes up to 32, attributing the combine gap to local reduction overhead and to HBM contention with URMA. The same README says those bandwidth figures were collected on a PoC HDK supplied to DeepSeek with additional manual configuration that "is not a publicly distributed release", and names Huawei's Q3 commercial release for the Atlas 850E, public availability planned for around 15 October, as the recommended public baseline. Reduce-scatter and all-reduce collectives are still being built, expert load balancing exposes the aligned APIs with the Ascend communication kernels not implemented yet, and pipeline parallelism and the Engram memory path are marked experimental. DeepGEMM-Ascend ships under an MIT licence; DeepEP-Ascend had no licence file, a bare 404 on the path, and GitHub's API returned null for its licence field. The announcement also says DeepSeek and Huawei are jointly advancing a 128-card supernode on the Ascend 950, and EP128 is the last and weakest row of DeepEP's own table.

Meituan launched LongCat-2.5-Preview on 25 September on its LongCat API open platform, and the release notes list three things it does: it understands images, it is good at code, and it is deeply adapted to Claude Code and other mainstream development environments. That third bullet has been on every LongCat release since 20 April 2026. The setup guide repoints ANTHROPIC_BASE_URL at api.longcat.chat/anthropic and maps four Claude Code model variables, including the ones named for the Opus and Sonnet tiers, to LongCat-2.5-Preview; the platform documents itself as fully compatible with the Anthropic Claude API format and ships thirteen such client configurations. The pricing page carries two yuan per million input tokens on a cache miss, four fen on a cache hit and eight yuan per million output, with a dollar panel reading thirty cents, six tenths of a cent and $1.20, and those two panels reconcile at Friday's onshore rate of 6.7125. The predecessor LongCat-2.0's pricing page carries the identical three numbers, so nothing is charged for the size, the image understanding or the coding work. The 1.6-trillion-total, roughly 48-billion-activated and million-token-context figures are not on any Meituan primary and are attributed to IT之家 citing the platform's release notes. The open-weights record is the specific part: Meituan has published twenty-nine models on Hugging Face including LongCat-2.0's weights on 5 July, has published nothing since 31 July, and LongCat-2.5-Preview has no weights at all. The download counts complicate the obvious reading, with the trillion-parameter LongCat-2.0 taken 1,773 times in thirty days against 23,795 for the much smaller LongCat-Flash-Chat.

Yan Wentao (严文韬) joined DeepSeek on 21 September as its first chief financial officer, closing a vacancy that had run the full three years since the company was founded and had been advertised on and off in public since February 2025. The hire was the fourth-hottest topic on Baidu the next morning at 5,413,293 heat, a large number for a personnel announcement at a private company, and no English-language item in the day’s collection carried it. The profile does not match the job as posted. DeepSeek’s own advertisement asked for mastery of accounting standards and tax policy, listed a CPA as preferred and wanted a compliance and risk-control system built across the whole business; Yan was born in 1991, studied at Fudan, spent 2013 to 2020 at Tencent Investment and H Capital, joined Hillhouse Venture in 2020 and was newly made partner there, and has never run a finance function. The account in 36Kr, republished from 字母榜, lists his public deals as ByteDance, Zhipu, MiniMax, Xiaohongshu, J&T Express and Webull, two of which are DeepSeek’s direct competitors, and notes that Hillhouse has never held a share in DeepSeek, so this is not an investor converting a board seat into an operating job. Partners at Wuyuan, Sequoia China and Longzhu were reported to have talked to Liang Wenfeng about the role, and Liang wanted someone born in the 1990s. The reason a primary-market investor fits sits in the cap table. Before the first outside round the company was 66 percent Ningbo Cheng’en and 34 percent Liang Wenfeng personally; after it closed, Ningbo Cheng’en held 60.19 percent, Liang 31.01 percent, a new vehicle called Hangzhou Chengli 8.52 percent and the state-backed National AI Industry Investment Fund 0.28 percent. The outside money came in stripped of what outside money usually buys: economic rights, a five-year lockup and no vote, with the exception our own company file has carried since the round closed, which is that the state fund holds the only outside vote. Look through the structure and Liang controls about 84 percent. The round itself was 50 billion yuan in June 2026 at a post-money above 350 billion yuan, about 52.2 billion dollars at 6.7004, which reconciles to the 52 to 59 billion dollar range the entity page already carries, and a second round is reported at a 500 billion yuan pre-money. CITIC Securities has been engaged for a STAR Market listing that 36Kr puts in 2027, against Reuters reporting a process starting in 2026 with no tutoring agreement signed and DeepSeek confirming nothing. On compute, 36Kr describes a roughly 1 gigawatt centre at Ulanqab in Inner Mongolia and at least 160,000 domestic parts, adding an order value of about 2.56 billion dollars, or 16,000 dollars a part.

Zhipu open-sourced ZCode on 21 September as zai-org/ZCode, Apache-2.0, and the public commit log holds two entries: "Initial commit" at 2026-09-20T12:06:58Z and "feat: open source" nine hours later at 21:14:32Z, against a repository created at 12:01:16Z the same day. The Issues tab is off, so the API field has_issues reads false and there is nowhere to file a report; the repository stood at 5,042 stars and 1,416 forks. The release answers the worst week the product has had. On 18 September a developer writing as ferstar found a .zcode folder holding more than 700MB, including a 313MB encrypted archive of most of the core assets of a roughly 10GB commercial project, still on disk because the upload had failed 564 times. Reverse-engineering found that while a user was signed in ZCode snapshotted the whole workspace to Alibaba Cloud object storage, and the package carried not only current source but the complete Git version history, the large-file cache, the reflog and some global development configuration. The RSA public key came down from the server and the private key never left it, so a developer could not open a package built from their own disk. The upload was on by default and the interface had no switch to turn it off. A day later Chengming Technology of Taiyuan put twelve questions to Zhipu in public and asked for written answers by 10 October, saying its own forensics found six workspaces, 34,549 files and about 425MB uploaded, including source, architecture, version history, database passwords, cloud credentials and employee personal information, and that the client made requests to a Singapore-registered entity while the Chinese terms are signed with Beijing Zhipu Huazhang. Zhipu has told Chinese reporters the allegations are untrue, nobody outside the two parties has adjudicated any of it, and the file counts are not established. Zhipu apologised on the 18th and blamed a repository-indexing feature behind a Repo Wiki page, shipped v3.14.0 removing it and cutting the snapshot and upload path, and commissioned two reviews: it says CAICT found the implicated storage bucket in a state of zero data in the cloud and NSFOCUS confirmed the bucket and every object in it had been deleted. Deleting the bucket is what makes those audits weaker than they sound, because an empty bucket says nothing about the window between upload and deletion, and some users’ workspace state files still carry a lastAcceptedManifestHash pointing at packages that no longer exist to be checked. The cost sits in the business mix. Zhipu’s open-platform and API revenue went from 15.2 percent to 86.5 percent of the total in twelve months as on-premises deployment fell away, which is the arrangement where the customer’s code stays on the customer’s machines. The company said in August that ZCode had reached a million users, and on 13 September it announced a share placing and a zero-coupon convertible totalling over HK$39.3bn, about 5.01 billion dollars, roughly 60 percent of it for the next GLM base model and the compute to serve it. That is the second raise in ten weeks: the July placement sold shares at HK$1,588 and the stock closed at HK$794.

OpenRouter’s weekly leaderboard, computed from the token bucket ending 16 September, puts Tencent’s Hy4 preview second in the world at 12 trillion tokens, behind OpenAI’s GPT-5.6 Luna at 17.3 trillion and ahead of every DeepSeek build. Eight of the top ten models are Chinese: GLM 5.3 Flash from Z.ai at 11.3 trillion, DeepSeek V4 Flash 0731 at 11.2, DeepSeek V4.1 Flash at 10.3, Xiaomi’s MiMo-V2.5 at 7.83, Tencent’s older Hy3 at 4.47, an April build of DeepSeek V4 Flash at 4.24, Nvidia’s Nemotron 3 Ultra at 3.49 and GLM 5.3 at 2.46. The Chinese entries carry 63.8 trillion tokens against 20.79 trillion for the two American ones, a ratio of 3.07, which was recomputed from OpenRouter’s own published figures rather than taken from the Sina piece that ran the same claim that morning. Pricing pulled from OpenRouter’s models API in the same session changes how the table reads: Hy4 preview costs 83.4 cents per million prompt tokens and 2 dollars 50 per million completion tokens, while the Nvidia entry at ninth is the variant published at zero, so one of the two American models in the world’s top ten by usage is free. The second half of the issue runs the other way. Leiphone reported as an exclusive, citing people it did not name, that Yao Xingcheng, who ran pre-training for Tencent’s Hunyuan large language model and left in July, has joined Mira Murati’s Thinking Machines Lab, and put the Tencent package in the tens of millions of yuan a year. That is one outlet’s sourcing on an unannounced move and is unconfirmed elsewhere, so the salary figure is loose and only the direction of travel is firm. The same outlet published a second piece arguing that Chinese platforms hiring pre-training researchers out of Google’s Gemini team have mostly bought seniority rather than the undocumented judgement calls that decide whether a training run works. In the Briefing: Insta360 grew first-half revenue 50.29 percent to 5.517 billion yuan while net profit fell 94.15 percent, and subtracting its filed first quarter from the half leaves a second-quarter net loss of 54.2 million yuan that no outlet printed; Huawei announced the Ascend 960 SuperPoD at Huawei Connect, up to 4,096 NPU cards and 8 exaflops of FP8, on the claim that chip-to-chip communication eats more than 40 percent of training time in a 100,000-card cluster; and Zhipu disclosed that a GLM-5.3-driven Infra Agent did much of the engineering on its own inference stack, which chief scientist Tang Jie calls an early form of recursive self-improvement.

DeepSeek posted about 150 openings on 8 September and not one of them is a research role, per a WeChat notice read closely by QbitAI and republished by 36Kr. They split into server-side development engineers across six directions, which are the large-model research platform, Agent framework components, R&D efficiency infrastructure, the DeepSeek API, online serving and data engineering, and agent elastic-compute engineers across two, platform development and low-level systems, with a target profile of senior backend people two to ten years in. QbitAI puts the whole company at 300 to 500 people, so a single round is worth somewhere between a quarter and a half of existing headcount, hired into one function; an investor at a front-line firm told Sina Tech the goal for this year is 1,000 people, which would put DeepSeek level with Zhipu, whose 2025 prospectus listed 883 employees. What the 150 are being hired to build has a name: DSec, for DeepSeek Elastic Compute, disclosed for the first time in the V4 technical report as three Rust components, an API gateway called Apiserver, an Edge agent on every host machine and a cluster monitor called Watcher, wired over a homemade RPC protocol on top of 3FS, the distributed filesystem DeepSeek open-sourced last year. Cui Tianyi, who runs the DeepSeek Harness team, gave the reason in one sentence: scale produces complexity, and the complexity turns around and demands more scale. Underneath it is the money. Reuters reported on 9 September that DeepSeek has hired CITIC Securities to prepare a listing on Shanghai’s STAR Market and aims to start the process this year, which TechNode carried the same morning; Sina’s wire adds that CITIC has made contact and entered due diligence, that the two sides have not signed a formal listing-tutoring agreement, and that DeepSeek has said nothing officially.

Z.AI, the entity Zhipu lists under in Hong Kong at stock code 2513, filed interim results on 31 August 2026 covering the six months to 30 June. Revenue was 953.892 million yuan, up 399.7 percent, while gross margin fell from 50.0 percent to 26.4 percent. The move is mix rather than discounting: open-platform and API revenue was 825.176 million yuan, up 2,735.7 percent, and went from 15.2 percent of revenue to 86.5 percent, while on-premises deployment fell from 73.7 percent of 2025 revenue to 13.5 percent of this half. Token volume on the MaaS platform is up more than 40 times since the start of 2026 and the average API selling price rose about 101 percent over the same stretch, so volumes surged while prices went up. That leaves the 24.6 percent gross margin on the open-platform and API business a cost problem, not a pricing one.

Harvey, the OpenAI-backed legal AI company, published its first post-trained model on 20 August. Harvey Tenet is a Kimi K3 base, post-trained with a rank-64 LoRA over the full network on a dataset of about 1,750 tasks. It completes almost twice as many held-out tasks on the Legal Agent Benchmark as base Kimi K3, and places second on that benchmark overall. Harvey ran ablations on candidate judge models against heavier frontier models and settled on Kimi 2.6 as the grader. Its M&A diligence work uses a GLM-5.2 orchestrator, taking rubric criteria pass rate from a 43.8 percent best baseline to 60.1 percent after self-distillation post-training.

Chinese cloud vendors are reselling competitors’ models at roughly 30 percent of list price, below their own cost, because in AI coding the customer names the model. Leiphone reported the discount ladder on 21 August from named cloud sales sources: 60 to 80 percent of list six months ago, 48 percent now common, roughly 30 percent at the low end. The same day, Approaching.AI vice-president Guan Jiawei told the South China Morning Post that domestic chips can only serve the low-quality inference tier, the one with weak demand and weak monetisation, and that high-quality tokens, especially coding, still depend on Nvidia. iFlytek said the same week that a flagship model trained entirely on domestic compute would reach the domestic first tier on code capability.

DeepSeek open-sourced Harness, an MIT-licensed agent framework, on 13 August. Within four days its repository had passed both anthropics/claude-code and openai/codex on GitHub stars, 148,734 against 141,759 and 106,470, while Anthropic’s repository carries no open-source license at all. The caveats are in the same API response: Harness is a pre-1.0 release candidate with its public issue tracker switched off, and stars measure attention rather than use.

DeepSeek raised API prices by up to 12x, 83 days after calling its 75 percent V4-Pro discount permanent. Tencent Cloud matched the move rather than undercutting it, and the repricing ran industry-wide: Zhipu has raised prices three times this year and Kimi K3 output passed 100 yuan per million tokens. The stretch of the market where Chinese models competed purely on price is closing.

DeepSeek finished shipping V4, and its day-0 domestic-chip support runs on Huawei Ascend (via Huawei’s CANN, its answer to Nvidia’s CUDA), Cambricon and Hygon. After the launch Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950 processors.

An IDC survey found 47 percent of large US firms now run a Chinese model in at least one use case, and at one point mid-July all five top models on OpenRouter were Chinese. Coinbase, DoorDash, Airbnb and Cursor are named adopters, switching to cut the bill as UBS puts Chinese API prices at 10 to 20 percent of the US alternative.

DeepSeek open-sourced V4 under MIT with Huawei Ascend as a primary deployment path: 1.6 trillion parameters, million-token context, coding a hair behind the best Western models. The first sovereign stack.

DeepSeek closed its first outside round, roughly 7.4 billion dollars at a 50-billion-plus valuation. The founder wrote the biggest check; the only investor with a vote is a state fund.

A teardown found part of DeepSeek V4’s architecture was co-designed for Huawei’s Ascend 950DT inference chip. The model and the silicon are now being built for each other.

US Commerce ordered a leading AI lab to cut its two most capable models off from all foreign nationals worldwide, then reversed it a week later. Export controls reached model access, and retreated.