Search “DeepSeek”

Chaoyan Intelligence Releases and Open-Sources Its First Large Model, Apex-35B-V1.0

On October 28, 2026, Beijing-based large model startup Chaoyan Intelligence released and open-sourced its first large model, Apex-35B-V1.0. According to a report by Zhidongxi, the model has a total parameter count of 35 billion, activates approximately 3 billion parameters per token, supports a 262,000-token context window, and is optimized through mid-training and post-training based on Qwen3.6-35B-A3B. In testing, Apex-35B-V1.0 scored 60.9 and 40.0 on PaperBench and FrontierScience, respectively, higher than GLM-5.3's 42.5 and 30.7; it scored 84.0 on BFCL V4, higher than DeepSeek-V4.1-Flash-1017's 75.0. The model has been released with open weights under the Apache 2.0 license, with official provision of approximately 72GB BF16 weights and GGUF quantized versions at different precisions. No further developments have been disclosed at present.

CrowdStrike Report: VAPOR PANDA Uses AI Agents to Attack Over 200 Targets

On October 29, 2026, CrowdStrike Intelligence released a report stating that the China-linked group VAPOR PANDA used the AI-driven ARTEX intrusion framework to target more than 200 organizations in at least 14 countries over an 11-day period beginning October 17, 2026, and confirmed post-compromise activity at 9 of them. The report states that ARTEX relies primarily on the DeepSeek large language model, supplemented by another commercial model, with a main agent and planning and execution sub-agents sharing a fact and goal graph; in a single intrusion, the planning agent consumed approximately 1.4 billion input tokens, and operators issued instructions in Chinese and intervened at key junctures. CrowdStrike also stated that the group deployed the BLOODALCHEMY backdoor at some targets, and predicted that China-linked adversaries will increasingly adopt such agent frameworks over the next 12 months. The report was covered by Unite.AI on October 30, 2026, and there are no subsequent developments as of yet.

Replit releases ViBench 2.0, a long-horizon software development benchmark

Replit has released ViBench 2.0, designed to evaluate AI agents' ability to start from an empty project and build a complete product feature by feature. The benchmark includes 17 products, 241 feature requests, and 544 test plans. According to results published on the Replit Blog, GPT-6.1 Sol leads with 65% pass@1, GPT-6 Astra is at 60%, and Opus 5.5 is at 49%, with all three passing over 90% of test plans. Opus 5.5 completes a build in about 70 minutes on average, roughly 2.5 times faster than Sol; DeepSeek V4.1 Flash passes 77% of test plans at about $2.50 per build, but no open-weight model is able to pass a complete build.

Mistral calls Le Chonk the world's best open-weight model

On October 29, 2026, Mistral released the official version of Le Chonk, a trillion-parameter open-weight model, calling it the world's best open-weight model for automated coding and cybersecurity, available for free. Pierre Stock, Mistral's Vice President of Science, said the model was trained on increasingly difficult cybersecurity and coding problems, enabling it to match or surpass models from Chinese labs such as DeepSeek and Z.ai. Stock also emphasized that the West needs its own advanced open-weight models, because closed-source models can be shut down remotely, whereas open weights cannot have access cut off once released. Currently, this event has only been reported by Wired AI, with no further developments yet.

Doubao Work opens third-party large model services to enterprise customers

On October 29, 2026, Doubao Work opened third-party large model services to Team Subscription and Enterprise Subscription customers, initially integrating a series of models including Kimi 3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash. According to a report by Leiphone.com that day, the feature is enabled by default after launch, and enterprises do not need to separately procure from model vendors or submit additional applications; usage is deducted from the enterprise's existing AI credits based on consumption. Enterprises can define the available scope by member, department, group, or all employees, and specify the scenarios in which the models take effect, including the Doubao desktop app, Doubao Work desktop app, Doubao mobile app, Doubao web version, and Doubao Work within Feishu. On October 30, Wallstreetcn reported that after this opening by ByteDance's Doubao Work, enterprises can directly use their existing AI credits to make calls without separately purchasing APIs; the report also noted that Tencent WorkBuddy has already integrated models such as Hunyuan, GLM, MiniMax, Kimi, and DeepSeek, and pointed out that after multi-model integration, platforms still need to coordinate context length, tool calling, and cost at the Harness layer. Currently, the service is already available to the aforementioned enterprise customers.

Baseten becomes Postman's recommended inference partner for enterprise Agents

On October 30, 2026, Baseten became Postman's recommended inference partner for enterprise agents, responsible for providing production inference and post-training support for agents built through the Postman Agent Stack. In an engineering blog post, Baseten said that its seven model APIs released in 2026 all ranked in the top two in Artificial Analysis's throughput and latency rankings, with Beam 1.0, DeepSeek V4.1 Flash and others taking first place. The blog post also mentioned that Harvey used Baseten to post-train a 27B open-source model, matching closed-source frontier models on legal agent benchmarks, while the latter cost about $50 per task. There are currently no subsequent developments.

Prime Intellect releases the Prime Mega MoE training kernel

Prime Intellect releases the Prime Mega MoE training kernel. This kernel builds on the Mega MoE kernel from DeepSeek's open-source DeepGEMM library by adding a fused and overlapped backward pass, making it usable for training rather than inference only. In the prime-rl trainer, a full training step for the Qwen3 MoE model is 32% to 34% faster than the baseline and 48% to 55% faster than TorchTitan; single-layer forward pass speedup reaches up to 3.11x (from 12.0 ms to 3.9 ms), but the backward pass for Qwen3-235B-A22B is only 1.24x faster. The kernel supports BF16, MXFP8 on Blackwell, and blockwise FP8 on Hopper, is released under Apache 2.0 with prime-rl, and is enabled via the mega_moe dispatch backend.

The National Supercomputing Internet announces a price adjustment

On October 29, 2026, the National Supercomputing Internet announced its latest price adjustment. According to Leiphone.com, new users can enjoy a limited-time flash sale offer of 50 million tokens for 0.1 yuan on two models, DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash, and a number of new mainstream models have also been added to the TokenPlan package pool. The report also said that users who subscribe to TokenPlan can call the Auto model free of charge every night from 21:00 to 09:00 the next day between October 21 and 30, in order to reduce AI R&D and debugging costs. No further developments have been disclosed at present.

Doubao Work opens third-party large model services to enterprise customers

On October 29, 2026, Doubao Work opened third-party large model services to Team Subscription and Enterprise Subscription customers, initially integrating a series of models including Kimi 3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash. According to a report by Leiphone.com that day, the feature is enabled by default after launch, and enterprises do not need to separately procure from model vendors or submit additional applications; usage is deducted from the enterprise's existing AI credits based on consumption. Enterprises can define the available scope by member, department, group, or all employees, and specify the scenarios in which the models take effect, including the Doubao desktop app, Doubao Work desktop app, Doubao mobile app, Doubao web version, and Doubao Work within Feishu. On October 30, Wallstreetcn reported that after this opening by ByteDance's Doubao Work, enterprises can directly use their existing AI credits to make calls without separately purchasing APIs; the report also noted that Tencent WorkBuddy has already integrated models such as Hunyuan, GLM, MiniMax, Kimi, and DeepSeek, and pointed out that after multi-model integration, platforms still need to coordinate context length, tool calling, and cost at the Harness layer. Currently, the service is already available to the aforementioned enterprise customers.

View story:Doubao Work opens third-party large model services to enterprise customers

Liang Wenfeng is said to emphasize two underlying logics: openness and restraint

On October 27, 2026, reports said that DeepSeek founder Liang Wenfeng stressed two underlying logics at a closed-door meeting: first, proactively open-sourcing core capabilities so that the entire industry can grow on top of its foundation; second, amid the boom of wrapping applications around models to monetize, restrain short-term gains and relentlessly focus on basic architecture. The same report said that Liang Wenfeng, with a fortune of 280 billion yuan, appeared for the first time on the Hurun Rich List, ranking seventh nationwide; people familiar with the matter said he has recently been intensively catching up on embodied intelligence in Beijing and has begun to work out the next collision between models and the physical world. The above content comes from a single report and has not yet been cross-verified by other sources.

blakeandersonw open-source ai-app-playbook

On October 28, 2026, Blake Anderson open-sourced "ai-app-playbook," a set of operating procedures for coding agents to execute, covering product selection, building, publishing, and first-month profitability. The playbook has two layers: a human-facing playbook with chapters 00–13, and an Agent-facing AGENTS.md (Phase 0–14, with each step marked [AGENT] or [HUMAN]); before starting, one must pass the Gate A scoring. The tech stack is fixed as Cloudflare Workers, D1, Better Auth, OpenRouter routing to DeepSeek, Superwall plus Stripe, etc., and explicitly lists items not to use, such as Vercel, Supabase, and Next.js. Currently there is only this open-source release information, with no subsequent developments yet.

China's Ministry of Commerce and Ministry of Science and Technology plan to expand the export control technology catalog

On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48, soliciting public comments on revisions to the Catalog of Technologies Prohibited or Restricted from Export in China, proposing to add 40 technologies, 23 of which fall under artificial intelligence, covering advanced large models, AI agents, and large-scale computing clusters. The draft sets three thresholds for frontier model weights, with any one triggering restrictions: training compute of no less than 10^26 floating-point operations, parameter count of no less than 2 trillion, or ranking in the global top ten on indicators such as public benchmark rankings and global call volume at the time of export; high-performance agents intended for critical infrastructure, national defense, or core government uses are directly prohibited from export. The comment period for the newly added portions of the catalog runs until November 15, 2026, and the already-effective pre-export registration requirement applies immediately. On October 23, DeepSeek and MiniMax completed registration for open-source and overseas authorized use, and will continue to open-source their models.

Entertainment Capital had six major models each write 5 episodes of a food-and-beverage business-war short drama

Entertainment Capital had Doubao, Kimi, DeepSeek, GPT, Qwen, and Claude each write 5 episodes of a food-and-beverage business-war short drama, for a total of 30 scripts. According to a TMTPost·AGI report on October 24, 2026, Claude was deemed to have the most "Shaoxin flavor," with dialogue rhythm and functional characters such as apprentices and trainees closest to existing Shaoxin short dramas; Doubao came second, resembling more of a punchline generator, but three versions of its scripts did not specify what the shop sells. The report stated that among the 30 scripts, "conscience" appeared in all of them, "rules" appeared in 80% of the scripts, and the word "Shaoxin" did not appear even once. A Hushu report on October 25, 2026 further stated that the 30 scripts were highly homogenized: noodle shops appeared 6 times, and all six models chose a noodle shop in their first versions; Claude's dialogue was short and rhythmic, with an apprentice urging a price cut in every version, closest to the "Shaoxin style"; Doubao had dense punchlines, but 3 versions did not specify what the shop sells.

Entertainment Capital asked six major models to each write 5 episodes of a food and beverage business-war short drama

Entertainment Capital had Doubao, Kimi, DeepSeek, GPT, Qwen, and Claude each write 5 episodes of a food-and-beverage business-war short drama, for a total of 30 scripts. According to a TMTPost·AGI report on October 24, 2026, Claude was deemed to have the most "Shaoxin flavor," with dialogue rhythm and functional characters such as apprentices and trainees closest to existing Shaoxin short dramas; Doubao came second, resembling more of a punchline generator, but three versions of its scripts did not specify what the shop sells. The report stated that among the 30 scripts, "conscience" appeared in all of them, "rules" appeared in 80% of the scripts, and the word "Shaoxin" did not appear even once. A Hushu report on October 25, 2026 further stated that the 30 scripts were highly homogenized: noodle shops appeared 6 times, and all six models chose a noodle shop in their first versions; Claude's dialogue was short and rhythmic, with an apprentice urging a price cut in every version, closest to the "Shaoxin style"; Doubao had dense punchlines, but 3 versions did not specify what the shop sells.

View story:Entertainment Capital had six major models each write 5 episodes of a food-and-beverage business-war short drama

Artificial Analysis says Kimi K3.1 Pro scores 64% on the Cyber Index

Artificial Analysis reports that Kimi K3.1 Pro scored 64% on the Artificial Analysis Cyber Index, ranking first among openly available models. The index consists of three benchmarks: CWE-Bench-AA, DeepsecBench-AA, and CyberGym-E2E-AA; K3.1 Pro scored 95% on CyberGym-E2E-AA, higher than GPT-6 Sol (Daybreak Blue)'s 93%. Its cost per task is $2.81, higher than GPT-6 Sol (Daybreak Blue, max)'s $1.77 and DeepSeek V4.1 Flash 1017's $0.42. Earlier the same day, Artificial Analysis ranked DeepSeek V4.1 Flash 1017 first among openly available models with a score of 57%, second on the overall leaderboard, 12 points behind GPT-6 Sol (69%), which is offered through a trusted access program; this version improved by 16 points over its predecessor DeepSeek V4.1 Flash's 41%, mainly due to CyberGym-E2E-AA rising from 23% to 76%.

Huawei Secures $2.25 Billion Offshore Loan from Six Chinese Banks

Huawei has obtained a seven-year, $2.25 billion offshore loan provided by six Chinese banks. Related reports say the funds will be used for AI chip R&D, research, and inventory stockpiling. The reports also note that even though buyers such as DeepSeek plan to deploy its accelerators on a large scale, Huawei's production output remains strictly limited; Huawei's R&D investment in the first half of the year reached $17 billion, with negative cash flow of $5.6 billion. Currently, this event has only the above single-source report, with no subsequent factual developments.

Arena ranks Kimi K3.1 Pro (Max) No. 11 in Agent Arena

On October 24, 2026, Arena added Kimi K3.1 Pro (Max) to the Agent Arena rankings, placing it at No. 11. The ranking is based on 4K multi-real-agent sessions, achieving a +7.6% net improvement, up 6 percentage points from Kimi K3 (Max)'s No. 25. Its median cost per task is $1.03, 13% lower than the previous version; it ranks No. 8, No. 9, and No. 10 in the Work, Code, and Chat categories, respectively. Arena said the model's open weights will be released soon, at which point it will become the top-ranked open-source model in Agent Arena, surpassing DeepSeek-v4.1-Flash. There are currently no subsequent developments.

Shao Meng questions that several Pro models fall short of expectations

On October 24, 2026, Shao Meng raised questions, arguing that several recent models ending in Pro have fallen short of expectations. The related remarks named DeepSeek V4 Pro, Kimi K3.1 Pro, and the not-yet-released Gemini 3.5 Pro. Based on this, the poster suggested that building Flash models is the right path for most AI teams. Currently, this questioning comes only from Shao Meng's own post, with no other developments or official responses.

CAICT and others jointly release a large model token service performance monitoring report

The Artificial Intelligence Research Institute of the China Academy of Information and Communications Technology (CAICT), the Key Laboratory of Artificial Intelligence Large Models and Software/Hardware Evaluation of the Ministry of Industry and Information Technology, and the China Artificial Intelligence Industry Development Alliance jointly released the "Large Model Token Service Performance Monitoring Report (2026)." Based on monitoring data from January to August 2026, the report notes that in June 2026, domestic average daily token consumption was approximately 500 trillion, nearly four times that of December 2025. The report compares the performance of original manufacturers' official platforms and four open-source models—DeepSeek, Qwen, GLM, and Kimi—on third-party platforms in terms of call success rate, throughput speed, first-token latency, stability, and price. In August, the average TPS of domestic models was 78.23, double the 2025 average of 38.25, but still lower than the 112.82 of typical foreign models; the domestic normalized average price rose from 6.8 yuan in the second half of 2025 to 13.46 yuan. There are currently no subsequent developments.

LMSYS releases SGLang AWS Trainium plugin Beta

On October 22, 2026, LMSYS partnered with AWS to release the SGLang Trainium plugin Beta. Built on the newly released TorchNeuron, the plugin supports LLM inference serving and RL rollout generation on Trainium2 and Trainium3. Initial coverage includes Qwen3 (dense), Qwen3-MoE-30B-A3B, DeepSeek-V2-Lite, GLM-4, and GLM-5.2/5.3, and it supports data parallelism, tensor parallelism, expert parallelism, Radix Cache, BF16/FP8, and asynchronous scheduling. In the first end-to-end heterogeneous RL run, SGLang generated rollouts for Qwen3-MoE-30B-A3B on Trainium3, while Miles trained a GRPO policy with Megatron on a single H200 node. The Beta is currently limited to a small number of teams. The next day, Amazon announced the open-sourcing of TorchNeuron, a backend that converts PyTorch code into operations executable on AWS Trainium, allowing PyTorch code already running on other accelerators to migrate with almost no changes.