MaaS platform weekly tracking
Phase 2026-05-03 · Full public sample
Volcengine Ark（Doubao） | Coding Plan integrates multiple models, starting with CNY 8.91 | Continue to monitor new model integrations |
Alibaba Cloud Bailian | HappyHorse video generation, Qwen3.6 updates, optimization of free quota management | HappyHorse official version released. |
Baidu Qianfan | 7+ MCP Component Launch, Third-party Model Integration | ERNIE-5.0 Continuous Iteration |
Tencent Hunyuan | Hy3 Preview releases a model that tops OpenRouter and open-sources an offline translation model. | Hy3 official version released, larger size model |
SiliconFlow | Yang Pan resigned, financing in progress, FP8 endpoints 83.7 t/s | A new round of financing has been completed. |
Zhipu AI | GLM-5V-Turbo launch, GLM-5 paper public, three price increases | GLM-5.5 release expectations |
Kimi (Moonshot AI) | K2.6 released an open-source product, raised prices for the first time, and generated overseas revenue exceeding domestic revenue. | IPO Process of Going to Hong Kong |
MiniMax | M2.7 self-evolution model release, six-core implementation | Expected release of the new M3 model |
DeepSeek | V4 launched open-sourced, financing launch, Huawei Ascend adaptation | V4-Pro promotion ends May 31; old API July 24 will be discontinued |
01DeepSeekHeadlineDeepSeek V4 is released and open-sourced (April 24)
— V4-Pro has 1.6T total parameters, 49B activation parameters, V4-Flash 284B/13B, a context of 1M token, and MIT open-source licensing. The CSA/HCA hybrid attention mechanism is introduced, reducing KV Cache to 10% of V3.2 and reducing inference FLOPs to 27%. Fully compatible with Huawei Ascend 950PR chips, the training framework has been migrated from NVIDIA to Huawei. The first external financing has been launched, with the target valuation rising from $100B to $200-300B, and Tencent and Alibaba are in talks to participate. The V4-Pro limited-time 75% discount has been extended to May 31.
02KimiHeadlineKimi K2.6 released and open-sourced (April 20).
— 1T total parameters (activating 32B), the open-source model with the largest total parameter count in the industry, 256K context. The Agent cluster architecture supports 300 sub-Agents working in parallel and 4000 coordination steps. It surpasses GPT-5.4 in HLE, BrowseComp, and SWE-Bench Pro. Pre-investment valuation is $180B (about 120 billion RMB), and a new 1 round of $1B financing is underway. The API input price has risen from CNY 4/million tokens to CNY 6.5 (a rise of 58%), the first price increase. Overseas revenue has surpassed domestic revenue, and the month-on-month growth rate of paying users exceeds 170%.
03Tencent HunyuanHeadlineTencent Hunyuan Hy3 Preview launched and open-sourced (April 23)
— 295B total parameters (activate 21B), 256K context, fast-slow thinking fusion MoE architecture (192 expert paths, top-8 routing). First Token latency reduced by 54%, end-to-end duration reduced by 47%. SWE-Bench Verified 74.4%, Tsinghua mathematics PhD qualification exam highest score in China. The first report following Yao Shunyu's leading technical restructuring: Tencent's AI has shifted from "passive" to "competitive." April 29 tops the OpenRouter API call volume rankings; tool calling ranks No. 1 globally, programming ranks No. 2 globally.
04MiniMaxHeadlineMiniMax M2.7 Releases
— Self-evolving model that participates in its own evolution and can build complex Agent Harness. SWE-Pro 56.22% (close to the Opus level), GDPval-AA ELO 1495 (the highest among open-source models). Day 0 has been deployed on six chips (Pingtouge, Huawei, Hygon, Muxi, Tianshu, NVIDIA). Music 2.6, Speech 2.8 and Hailuo 2.3 were released simultaneously. The self-evolution workflow achieved a 30% performance improvement through more than 100 rounds of iterative loops.
05Zhipu AIHeadlineZhipu AIReleases GLM-5V-Turbo
The first multimodal Coding base model deeply integrates visual and textual capabilities from the pretraining stage. It supports generating executable code with image inputs such as sketches, style references, screenshots, and recorded screens. The AutoClaw platform, Z.ai platform, and BigModel open platform are opened for experience simultaneously.
06HeadlineAlibaba Cloud Releases HappyHorse (Happy Little Horse) AI Video Generation Model
The 15 billion parameter single-stream Transformer architecture supports native audio-video synchronization generation. Generating H100 5-second 1080 P videos takes only 38 seconds, with a speed reaching 2-3 times that of mainstream industry levels. It supports lip synchronization matching across 7 Chinese, English, Japanese, Korean, German, French, and Cantonese. It takes the top spot in the Artificial Analysis Video Arena anonymous blind test list. Pricing is 720 P CNY 0.9/ seconds and 1080 P CNY 1.6/ seconds.
07HeadlineGrok 4.3 API releases (April 30)
Input prices for 40% decrease to $1.25/million tokens, with 1M context, supporting native video input and PDF/XLSX/PPTX document output. The Heavy system is upgraded to 16-Agent parallel scheduling. It reaches Voice API STT/TTS $4.20/per million characters, and is OpenAI cheaper than 86-92%. AA Intelligence Index 53.
08SiliconFlowFollowSiliconFlow co-founder Yang Pan leaves
— Approaching completion of a new round of funding. Users exceed 9 million, corporate clients exceed 10, with 000 companies, daily average Token generation exceeds 100 billion tokens, with revenue at the hundreds-of-millions-yuan scale. Transitioning to the "Token Factory" positioning from AI Infra, launching a corporate-grade private deployment solution SiliconBrain.
09Baidu QianfanFollowBaidu Qianfan added the 7 and MCP/Skill components
Including Baidu Search, smart search generation, Baidu Baike, Baidu Netdisk AI video notes, Baidu Wenku AI picture books, document parsing, Baidu Maps, etc. New user registration includes a CNY 20 voucher. It supports the OpenAI/Anthropic protocol and connects to third-party models such as deepseek-v3.2, kimi-k2.5, glm-5, minimax-m2.5.
10Volcengine ArkFollowVolcano ArkCoding Plan Update
Connecting to GLM-5.1, MiniMax-M2.7, and Kimi-K2.6, supporting the 6 large model and multiple tools. The vehicle launch price starts from CNY 8.91, and TRAE tools are supported, with Claude Code specific adaptation.
11Alibaba Cloud BailianFollowAlibaba Cloud Model StudioQwen3.6 Series Update
Qwen3.6-Plus 27B is on par with Artificial Analysis Agentic Index and Claude Sonnet 4.6, while 64 can GB MacBook Pro 18 with TPS. Bailian is opening 200 call interfaces for multiple models, offering 70 million Token free for new users, and Batch 5 at a discount. A function to stop free quota usage immediately upon exhaustion has been added.
12Tencent HunyuanFollowTencent Hunyuan open-sources an offline translation model.
Hy-MT1.5-1.8B-1.25bit only supports 440M, B, and 33 for offline translation, and its translation quality outperforms Google Translate. Sherry sparse quantization technology has been included by ACL 2026. It is fully processed locally with zero privacy exposure.
13FollowToken pricing trend established
OpenAI GPT-5.5 token prices have doubled, and within three months, the cumulative increase in ZhipuGLM 5.1's pricing for token has reached 83%, while Kimi K2.6's price increase is 50%. The supply gap has widened: demand is growing non-linearly, while supply is expanding linearly.
Detailed Tracking Across Platforms9 platforms
The complete guide for Coding Plan was released, supporting the 6 large model (Doubao Code Model, Kimi programming, DeepSeek code, GLM-4, etc.), with multi-tool support
Integrate GLM-5.1, MiniMax-M2.7, Kimi-K2.6
The boarding price starts from CNY 8.91, the TRAE tool is supported, and Claude Code specialized adaptation is available.
The flagship model is officially released, and a new addition is provided.
Recent review 2 items
April 29 Coding Plan Personal Edition Package Update
ByteDance's ¥1 billion sends Doubao to the Spring Festival Gala
Publish the HappyHorse (Happy Little Horse) AI video generation model, with 15 billion parameters, generating 5-second 1080P videos from a single H100 image, requiring only 38 seconds, supporting 7 languages of lip sync matching.
Qwen3.6 Series Update: Qwen3.6-Plus 27B is level with Agentic Index and Claude Sonnet 4.6, and 64 GB MacBook Pro can run 18 TPS
qwen3-vl-flash-2026-01-22 integrates thinking and non-thinking modes, with its visual recognition capability significantly improved
The Qianwen OCR model is dedicated to text extraction and supports multi-language recognition
Bailian offers 200 call interfaces for multiple models to the open platform; new users get 70 million free calls for Token, with Batch discount rates at 5.
A feature for pausing after the free quota is used up has been added, to avoid unexpected charges.
Alibaba Cloud's Lightweight Application Servers Preinstall OpenClaw CLI Tools
Qianwen ¥3 billion threw a sum in (to drink milk tea for the user) to promote through ad placement
Recent review 2 items
Qwen cumulative downloads exceed 1 billion times, with derived models exceeding 200000 in number
Alibaba Establishes the "Foundation Model Support Team", with CEO Leading Coordination
Added 7+ MCP/Skill components: Baidu search, intelligent search generation, Baidu Baike, AI video notes for Baidu Netdisk, AI picture books for Baidu Wenku, document parsing, and Baidu Maps.
New users receive CNY 20 vouchers upon registration.
Supports OpenAI/Anthropic protocol compatibility, integrating third-party models such as deepseek-v3.2, kimi-k2.5, glm-5, minimax-m2.5, ernie-4.5-turbo, etc.
Platform reconstruction centered on Agent
Recent review 2 items
ERNIE-5.0 (Released monthly in 2026 year 1 month): Total parameters exceed 2.4T, the ratio of activated parameters is below 3%, natively full-modal modeling, and a super-sparse MoE architecture
The Wenxin image model has been added to ComfyUI.
Hy3 Preview released and open-sourced (April 23): 295B total parameters (activating 21B), 256K context, MoE architecture integrating fast and slow thinking, 192 expert route for top-8.
April 29 ranks first in the overall OpenRouter API call volume leaderboard, with tool calls ranked globally 1 and programming ranked globally 2.
The first Token latency has reduced 54%, and the end-to-end duration has reduced 47%
SWE-Bench Verified 74.4%, Highest Domestic Score in the Tsinghua Mathematical Doctor Examination
Connected to Yuanbao, ima, WorkBuddy, CodeBuddy, QQ, QQ Browser, Tencent Documents, etc.
Supports open-source agent platforms such as OpenClaw, OpenCode, KiloCode.
Open-source offline translation model Hy-MT1.5-1.8B-1.25bit, only 440M B, supports 33 languages, and its quality is better than Google Translate.
Sherry Sparse Quantization Technology Included in ACL 2026
Token Plan Individual Edition starts from CNY 28/ per month, with Tencent CloudAPI input of ¥1.2/million tokens
The first report led by Yao Shunyu after leading the technical refactor: a strategic shift to "focus on scenarios, not rankings"
Recent review 2 items
Tencent Yuanbao ¥1 billion marketing investment
Larger model is in training
Founder and VP of Growth, Yang Pan, departs
Nearly completing a new round of financing
Register to receive CNY 16 vouchers (including Pro model usage), and get free 20 million, token, and API quota.
Over 150 mainstream open-source large models have been launched, and the adaptation cycle for the new model is 1-3 days.
DeepSeek V4 Flash FP8 reaches 83.7 t/s throughput
Immersive translation integration, Cherry Studio integration
Collaborated with Huawei Cloud on the GLM-5.1 experience solution
Recent review 3 items
User usage exceeds 9 million, enterprise customers exceed 10 and 000, daily average Token generation exceeds 100 billion and tokens, revenue at the billions-of-yuan scale
Shift from AI Infra to the "Token Factory" positioning, and launch the enterprise-grade private deployment solution SiliconBrain
Collaborated with Zhipu, Huawei, Alibaba, and ByteDance
Publish GLM-5V-Turbo: the first multimodal Coding foundational model, deeply integrating visual and textual capabilities from the pretraining stage, supporting image inputs such as sketches, screenshots, and recorded videos to generate runnable code.
AutoClaw (Aolong), the Z.ai platform, and the BigModel open platform all simultaneously open their experience to Coding Plan users.
GLM-5.1 technical paper public: introduces DSA sparse attention, reduces KV Cache overhead by 75%, improves inference speed by 3 times, maximum context 202,752 tokens
GLM-5 (754B parameters, activated 40B) MIT license fully open-sourced, entirely trained on approximately 100000 Huawei Ascend 910B chips using MindSpore framework
Asynchronous reinforcement learning infrastructure: the training engine and inference engine are decoupled, and GPU utilization is significantly improved from 20-30%
SWE-bench Verified 77.8% (Open-source SOTA), HLE (Including Tools) 50.4 Score, AA Intelligence Index v4.0 50 Score (Open-source SOTA)
Full-stack adaptation for domestic chips: Huawei Ascend, Moore Threads, Hygon, Cambricon, Kunlun Xin, Muxi, Enflame
GLM-Image open-source tops Hugging Face Trending
Three consecutive price increases: the 2 monthly code subscription package is raised by 30%, the flagship 3 monthly package is raised by API to 20%, and April 8 is raised again to 10%, with a cumulative increase of 83%
Recent review 2 items
GLM Coding Plan Opens a refund channel
Launch CC-Bench-V2 Automated Software Development Evaluation Set
K2.6 was released and opened source (April 20): 1T's total parameters (activated with 32B), the open-source model with the largest total parameter count in the industry, 256K context.
Agent cluster architecture supports 300 sub-Agent parallel collaboration (improved from K2.5 100), with 4000 coordination steps.
Enhance long-range encoding capability, supporting uninterrupted encoding for 13 hours.
The only one among the top five global open-source projects that supports image and video understanding
Exceed GPT-5.4 in HLE, BrowseComp, and SWE-Bench Pro
BenchLM leads 84:83 in GLM-5.1.
kimi.com and Kimi applications, Kimi API, and Kimi Code programming assistant have been launched.
API's input price rose from CNY 4 to CNY 6.5/million tokens (a 58% increase), marking its first price hike; it is approximately 8.3 times cheaper than Claude Opus 4.7 input at the endpoint.
Pre-money valuation $180B (approximately 120 billion RMB); the new round of $1B funding is in progress
Overseas revenue has surpassed domestic revenue, with monthly year-on-year growth for paying users exceeding 170%
After K2.5 is released, 20's revenue exceeds 2025 for the entire year.
Cursor, Cloudflare, and Perplexity are deeply integrated, and Perplexity lists Kimi as the only introduced Chinese model
API Platform Added Support for Web Search, Rethink, Memory, Excel, Code-Runner, Quick Js, Date, Fetch and other tools
Recent review 2 items
In CES 2026, Jensen Huang used DeepSeek and Kimi K2-Thinking as GPU performance benchmarks
Progress of going to Hong Kong with IPO is underway
M2.7 release: a self-evolving model that participates in its own evolution, capable of building complex Agent Harness - SWE-Pro 56.22% (close to Opus levels), VIBE-Pro 55.6%, and Terminal Bench 2 57.0% - GDPval-AA ELO 1495 (the highest among open-source models) - supporting complex editing of Excel, PPT, and Word, with 97% skill compliance rate (40+ complex skills) - MLE Bench Lite 66.6% average medal rate, with the best run earning 9 gold, 5 silver, and 1 bronze, matching Gemini-3.1 - self-evolving workflow: analyze failure trajectories → plan changes → modify code → run evaluations → compare results → decide to keep or roll back, achieving 30% performance improvement over more than 100 iteration rounds - Day 0 enables six-chip deployment (Pingtouge, Huawei, Hygon, Muxi, Tianshu, NVIDIA) - offering M2.7 and M2.7-highspeed2 versions, with full support for automated Cache
M2.5 SWE-Bench Verified 80.2%（industry SOTA），Multi-SWE-Bench 51.3%，BrowseComp 76.3%
M2.5 Pricing: 100 TPS with input $0.3/million tokens and output $2.4/million tokens; the 50 TPS version costs half as much; working continuously for 1 hours requires only $1 or $0.3; 10000 US dollars allows 4 Agent units to work continuously for one year.
MiniMax Agent deeply integrates Office Skills, users have already built 10000+ expert configurations, internal 30% tasks are completed autonomously by M2.5, code generation accounts for 80% of new committed code
Sync releases of Music 2.6 (Cover's cover version, instrumental enhancement, Agent integration), Speech 2.8 (AI's voice 'human warmth'), Hailuo 2.3 (ultimate dynamic video generation), M2-her (multi-role immersive role-playing)
Globalized layout, serving 214,000, and enterprise clients
Full adaptation of MaxClaw by Jiejue Yuma, one-click deployment with OpenClaw
Recent review 3 items
DeepSeek V4's Stock Price Fell by Nearly Half After Its Release; JPMorgan Maintained "Overweight" Rating
2025 annual financial report full year total revenue $79.04 M
Expected Catalyst: New Models GLM-5.5, MiniMax M3
V4 launched and open-sourced (April 24): - V4-Pro: 1.6T total parameters, 49B activation parameters, the world's largest open-source weight model - V4-Flash: 284B total parameters, 13B activation parameters, a lightweight and efficient version - Context 1M token, maximum output 384K token, MIT open-source license
Technical Highlights: Introduce the CSA and HCA hybrid attention mechanism, reduce KV Cache to V3.2's 10%, reduce inference FLOPs to 27%
Fully adapted to Huawei Ascend 950PR chips, with the training framework migrated from NVIDIA to Huawei
DeepGEMM operator library adds FP8_FP4 mixed precision computation, optimizing the NVIDIA Blackwell architecture
Supports JSON output, Tool Calls, FIM Completion, and supports switching between thinking mode and non-thinking mode.
Evaluation performance: AA Intelligence Index 52 points (second in open source), GDPval-AA Agent evaluation 1554 points (first in open source), SWE-Bench Verified outstanding
Pricing strategy: - V4-Flash: input $0.14/million tokens, output $0.28 - V4-Pro: input $1.74, output $3.48 - V4-Pro limited-time 75% discount extended to May 31 (promotional price input $0.435, output $0.87) - input cache hit price reduced to the launch price 1/10 (April 26 takes effect)
API Migration Notice:
deepseek-chatanddeepseek-reasonerIt will be offline in 2026 year July 24, and needs to be migrated to.deepseek-v4-proordeepseek-v4-flashLaunched its first external financing, with the target valuation rising from $100B to $200-300B, as Tencent and Alibaba are in negotiations to join in
Rakuten Japan develops Rakuten AI 3.0 based on DeepSeek
DeepSeek MLA's attention mechanism is adopted by Kimi, and the Muon optimizer verified by Kimi is adopted by DeepSeek V4 (technical interlocking)
Recent review 1 items
Hallucination rate: V4-Pro 94%, V4-Flash 96% (still needs attention)
Follow time point9 items
2026-05-31 | DeepSeek V4-Pro 75% discount ends | DeepSeek |
2026-07-24 | The DeepSeek old endpoint (API) is offline under deepseek-chat/reasoner. | DeepSeek |
2026 H2 | Huawei Ascend 950 supernode mass production rollout, and the DeepSeek V4-Pro price may be lowered further | DeepSeek、Zhipu AI |
Recently | ZhipuGLM-5.5 release expectations | Zhipu AI |
Recently | MiniMax M3 release expectations | MiniMax |
Recently | Alibaba HappyHorse Official Release | Alibaba Cloud Bailian |
Recently | Tencent Hunyuan Hy3 official version and larger-sized models. | Tencent Hunyuan |
Mid-term | Kimi's journey to the Hong Kong market is IPO. | Kimi (Moonshot AI) |
Mid-term | DeepSeek External Financing Progress (Target Valuation: $200-300B) | DeepSeek |
Index of reported events9 period
- [2026-04-23]
Tencent Hunyuan Hy3 Preview launched and open-sourced (295B/A21B, 256K, fast-slow fusion MoE)
- [2026-04-24]
DeepSeek V4 is released and open-sourced (V4-Pro 1.6T/49B, V4-Flash 284B/13B, 1M context)
- [2026-04-20]
Kimi K2.6 released and open-sourced (1T/A32B, 256K, 300 sub-Agent).
- [2026-04-08]
Zhipu GLM-5.1 is released and open-sourced (754B/A40B, Huawei Ascend 910B training, MIT license)
- [2026-04-29]
Tencent Hunyuan Hy3 tops the total call volume ranking with OpenRouter API.
- [2026-04-29]
Tencent Hunyuan open-sources an offline translation model Hy-MT1.5-1.8B-1.25bit (440MB, 33 language).
- [2026-04-27]
SiliconFlow co-founder Yang Pan leaves, and a new round of financing is in progress
- [2026-04-30]
Grok 4.3 API is released (1M context, input reduced to 40%, native video input).
- [2026-04-26]
DeepSeek V4-Pro Input Cache Hit Price Reduced to Launch Price 1/10