ChinaBiz Briefing | Alibaba Open-Sources Qwen3.8, DeepSeek V4 Pro, Tencent Bets Big on AI

August 13, 2026 marks one of the densest single-day news cycles in China's tech calendar this year. Three major AI releases landed within hours of each other, Tencent revealed a capex figure that stunned analysts, a smartphone startup redefined what a premium handset can do mechanically, and China's EV price war claimed another profitability victim. Taken together, the day's events underscore a single structural shift: China's technology industry is no longer catching up — it is setting terms.

Alibaba Opens Qwen3.8's 2.4T-Parameter Weights, Raising the Bar for Open-Source AI

Alibaba's Qwen team on August 13 published the full model weights for Qwen3.8-2.4T-A95B on Hugging Face and ModelScope — the first time the company has open-sourced a Max-tier flagship, just ten days after its commercial debut. The release landed within a 24-hour window alongside xAI's Grok 4.6 and DeepSeek's V4-Pro, compressing what was once a quarterly release cadence into near-simultaneous global launches. Third-party compression by Unsloth AI reduced the model from 4.9 TB to 397 GB, making self-hosting feasible for mid-tier cloud providers and well-resourced enterprises without licensing fees.

The strategic significance extends beyond benchmarks. Qwen3.8-Max leads on PaperBench (93.0) and OSWorld computer-use tasks (86.1), making it credible for scientific and office-automation workflows — though it trails on SWE-bench Pro (67.7 vs. Fable 5's 80.0), signaling continued reliance on Western models for complex software engineering. More consequentially, by releasing full Max-level weights rather than a distilled derivative, Alibaba raises the floor for what "open source" means in China and increases pressure on Baidu, ByteDance, and Zhipu AI to respond in kind. API pricing — $2.00 per million input tokens internationally — is competitive with mid-tier OpenAI and Anthropic offerings, while domestic pricing runs roughly 17% below, signaling a volume-anchor strategy at home.

DeepSeek V4-Pro Prices at $0.87/M Tokens, Undercuts Grok 4.6 by 7x

DeepSeek released DeepSeek-V4-Pro-0813 on August 12, benchmarking near or above Anthropic's Fable 5 on agentic tasks at an output price of $0.87 per million tokens — approximately one-seventh the cost of xAI's Grok 4.6, which launched almost simultaneously. On terminal operation benchmarks, V4-Pro scored 87.9, surpassing Opus 4.8 and sitting 0.1 points below Fable 5. Its software engineering score jumped from 12.8 in the April preview to 62.7 — a near-fivefold improvement that signals the model can now handle large-codebase refactoring, not just isolated code snippets. In cybersecurity agentic testing, it outright surpassed Fable 5.

The competitive read for enterprise API buyers is unambiguous: DeepSeek has again reset the pricing floor before Western incumbents could consolidate margin. The imminent public beta of DeepSeek Harness — the company's proprietary agent execution framework — would complete a full-stack agentic platform, shifting DeepSeek from model vendor to integrated toolchain provider. That repositioning puts it in direct competition with middleware vendors currently building on its raw API. One structural caveat: DeepSeek's infrastructure has experienced repeated capacity degradation following V4 Flash's general release, and flat launch pricing may be a tactical market-share defense rather than a durable commitment.

Tencent's RMB 52.8B Capex Bet Turns Free Cash Flow Negative

Tencent's Q2 2026 capital expenditure of RMB 52.8 billion ($7.3B) came in 64% above consensus and nearly three times the year-ago figure, pushing free cash flow negative for the first time in recent memory at negative RMB 13.8 billion. Revenue grew 11% year-on-year to RMB 204.8 billion; net profit of RMB 56 billion missed consensus by RMB 2.4 billion. Management was explicit about the allocation: training larger Hunyuan models, inference compute for WorkBuddy and CodeBuddy, powering WeChat AI agent Xiaowei, and meeting rising cloud demand. JPMorgan's full-year 2026 capex forecast of RMB 200 billion may prove conservative.

The investment thesis rests on two products. WorkBuddy, Tencent's AI productivity platform, recorded 20.97 million monthly PC visits in Q2 — ranking first among domestic office AI agent platforms, ahead of ByteDance's TRAE and Alibaba's QoderWork. Paying users' gross margins are already approaching Tencent Cloud's overall level, though free-user subsidies suppress blended margins. WeChat AI agent Xiaowei remains in limited grey-scale testing but carries a structural distribution advantage no standalone AI product can replicate: WeChat's 1.439 billion combined monthly active users. The critical open question is whether Xiaowei's eventual deployment compresses the high-margin advertising journeys that currently fund this entire investment cycle — a concern raised directly by UBS on the earnings call.

Honor's "Robot Phone" Sells Out at $1,389 on Launch Day

Honor launched what it calls the world's first robotics-grade smartphone on August 12, priced from RMB 9,999 ($1,389), and watched pre-sale inventory sell out within hours across major Chinese e-commerce platforms. The engineering centerpiece is a proprietary titanium-alloy 4-degree-of-freedom gimbal system integrating over 100 precision components — delivering mechanical articulation to a 248-gram handset. Honor simultaneously debuted the Yuguang H1, its first fully proprietary AI Signal Processor chip, which natively supports ARRI LogC3 cinema color encoding — the first consumer handset to do so without post-processing conversion — backed by a co-development partnership with ARRI, the Munich-based cinema camera manufacturer.

The launch redraws two competitive boundaries simultaneously. On hardware, Honor is competing directly with Apple's iPhone 17 Pro and Huawei's Mate 70 RS at the ultra-premium tier — a segment it has never previously contested. On AI platform strategy, the device ships with Alibaba Qwen deeply integrated into its Agentic OS, signaling that Honor — unlike Huawei's closed stack — is pursuing an open-model partnership approach. For Alibaba Cloud, the Robot Phone is a premium hardware distribution channel for Qwen at a moment when device-native LLM deployment is accelerating. Three variables will determine whether launch-day momentum converts into sustained market share: supply ramp speed for the precision gimbal system, competitive response timelines from Apple and Huawei, and whether the RMB 9,999 price point holds through the full product cycle.

Leapmotor Posts RMB 390M Loss, Losing ~$556 on Every Car It Sells

Leapmotor's Q1 2026 results snapped three consecutive quarters of profitability, reporting a net loss attributable to shareholders of RMB 390 million ($54.2M) as overall gross margin fell 550 basis points year-on-year to 9.4% and vehicle-specific gross margin deteriorated to approximately 7%. Free cash flow swung to a deficit of RMB 7.4 billion; average selling price dropped to RMB 98,000, a near two-year low. CFO Li Tengfei attributed the compression to a mix shift toward lower-priced B-series vehicles, underutilized manufacturing capacity, and declining strategic partnership revenues. The company's own disclosure: it is currently losing approximately RMB 4,000 ($556) per vehicle shipped.

The timing is structurally awkward. Leapmotor simultaneously launched the A05 compact EV at RMB 63,900 — equipped with LiDAR, Qualcomm 8650, and 510km range — a price point that triggered social-media frenzy but left Hong Kong-listed shares unmoved. Capital markets are asking the right question: Leapmotor can build a LiDAR-equipped intelligent EV at a price that shocks the industry; it has not yet demonstrated it can do so profitably. Full-year targets of 1 million deliveries and RMB 5 billion net profit now require approximately 640,000 units in H2 and a dramatic earnings acceleration that current unit economics do not support. Management is developing a second brand targeting the RMB 300,000-plus segment, planned for H2 2027 — but brand repositioning from a value anchor takes years, as SAIC-GM-Wuling's experience illustrates.

WeRide Posts Record Q2 Revenue as ADAS Surges 26x and Overseas Income Jumps 164%

WeRide delivered Q2 2026 revenue of RMB 232 million ($32.2M), up 82.2% year-on-year and 103.1% sequentially — its strongest quarter on record. The headline is the ADAS line: L2++/L3 revenue grew 2,594% year-on-year, driven by approximately 30,000 unit shipments. The company has secured production nominations across more than 30 vehicle models and targets 100,000 cumulative deliveries by end-2026 and 500,000 by end-2027. Gross margin expanded 9.4 percentage points to 37.5%, reflecting the higher-margin profile of software-intensive ADAS revenues and the structurally superior economics of the overseas asset-light model. Overseas revenue grew 164% year-on-year in Q2; WeRide's international asset-light model — licensing virtual driver technology while partners own the fleet — generates $40,000–$50,000 in annual recurring revenue per robotaxi vehicle, implying a $16M–$20M annualized run rate from its ~400-vehicle Middle East fleet alone.

The strategic logic is a cross-segment flywheel: L4 fleet data improves L2++/L3 model performance, which expands the training data pool that accelerates L4 iteration — a compounding advantage difficult for pure-play ADAS suppliers to replicate quickly. The constraint is persistent losses: adjusted net loss widened 12.6% year-on-year to RMB 338 million in Q2, and H1 net loss of RMB 790 million is essentially flat year-on-year. R&D spending of RMB 434 million in Q2 reflects continued investment in the GENESIS world model and WITT physical AI foundation model, which WeRide claims enables end-to-end autonomous driving on ~200 TOPS versus ~2,000 TOPS for some competing solutions. Whether the commercial flywheel generates sufficient cash flow to close the profitability gap before capital reserves are tested remains the central investor question.

What to Watch Next

The August 2026 window is shaping up as a structural inflection across multiple fronts. In AI, the competitive unit is visibly shifting from model benchmarks to agentic deployment platforms — DeepSeek Harness's imminent public beta will be the clearest test of whether that transition is real. Tencent's Hunyuan Hy4 release and WorkBuddy's revenue recognition trajectory will determine whether its RMB 52.8 billion capex bet begins to show financial returns before the end of the year. In hardware, Honor's supply ramp for its titanium gimbal system over the next 60–90 days will reveal whether launch-day sell-out reflects genuine demand or constrained supply. In EVs, Leapmotor's Q2 results — and whether the A05's aggressive pricing accelerates volume without further margin destruction — will be the next data point in China's broader EV profitability crisis. And for WeRide, the Madrid robotaxi pilot and the pace of ADAS delivery toward the 100,000-unit year-end target are the metrics that will determine whether its commercial inflection story holds.

Related Coverage:

DeepSeek's V4 Pro Undercuts Grok 4.6 by 7x as Agentic AI Race Heats UpAlibaba Releases Weights for 2.4T-Parameter Qwen3.8, Escalating Open-Source AI Arms RaceTencent's RMB 52.8B AI Bet Signals a New Phase of China's Compute RaceHonor's Robot Phone Sells Out in Hours as China's Smartphone War Turns RoboticWeRide's Asset-Light Pivot Validates as ADAS Revenue Soars 26x, Overseas Sales Jump 164%Leapmotor's Margin Collapse Exposes the Fatal Flaw in China's EV Price War