Search “DeepSeek”

English translation is being prepared.

View story:English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

Together AI Launches Together Compiler

On November 4, 2026, Together AI released Together Compiler (TKCC), a compilation system that uses high-throughput agents to continuously optimize inference workloads. The company claims that within less than 24 hours it achieved TPS improvements on frontier open-source models such as Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash; two official blog posts respectively state improvements of up to 2.6x and 1.3-3x. On SemiAnalysis's InferenceX low-latency benchmark, the system achieved current best results of 1111 TPS for MiniMax M3.1-Flash and 1205 TPS for DeepSeek v4.1-Flash, and in the Llama 1B batch size 1 scenario it is 1.7x faster than the open-source megakernel. TKCC relies on a persistent engine to hot-load candidate forward computations and on a two-tier fast/slow engine to balance throughput and fidelity; its experimental throughput is 252x higher than that of baseline agent-based optimization systems, and it is currently used to serve trillions of tokens per day.

Together AI releases Together Compiler

On November 4, 2026, Together AI released Polymorphic Inference, whose first implementation is Together Compiler. According to the Together AI blog, the solution is described by the company as the world's first production-grade recursive self-improving (RSI) inference system, and has used AI agents to optimize 40+ models on its endpoints, replacing static tuning of inference engine configurations with Continuous Compilation, in order to cope with accelerating model releases, diversifying architectures, changing traffic patterns, and hardware iterations from B200 to GB300. On the same day, Unite.AI reported that the compiler takes traffic patterns, models, and hardware as inputs, has parallel agents generate and measure candidate optimizations, and uses hot reloading to avoid restarting the entire engine; in tests on Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash, it reached 1.3x to 3x the tokens per second of a highly optimized open-source baseline within 24 hours. Customers can apply for early access through a limited beta.

Cohere launches third-party open-source models on Model Vault

On November 3, 2026, Cohere launched third-party open-source models in Model Vault and Model Vault Encrypted, allowing enterprises to run both Cohere's own models and open-source models such as DeepSeek V4.1 Flash, GLM 5.3 Flash, Gemma 4, GPT-OSS, and Phi on a fully managed encrypted inference platform. According to Cohere's product and research blog, Model Vault applies zero data retention to prompts and outputs by default, and the Encrypted version uses confidential computing so that neither Cohere nor cloud providers can see the data. Model Vault is already available in the United States, the European Union, and Canada; the Encrypted version is currently limited to the United States, with support for the European Union and Canada coming soon.

Hugging Face provides the CWE-bench v2 public set evaluation results on the Hub

On November 3, 2026, Hugging Face announced the CWE-bench v2 evaluation results and stated that the dataset, scorer, and harness are all publicly available and can be run directly from the Hub via Harbor. This evaluation covered 11 models: Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, followed closely by Claude Opus 5.5 with 67.3% pass@1. At least one model solved 95 tasks, but the top model still fails 19% of the time on each attempt. In terms of cost, DeepSeek V4.1 Flash, with 57.0% pass@1 at $0.04 per rollout, forms the Pareto frontier together with Claude Opus 5.5 and Gemini 4 Argon.

OpenClaw launches Particle-1

On November 3, 2026, OpenClaw launched Particle-1, and the model is now available in OpenClaw. According to @particle_api, compared with DeepSeek V4.1 Flash, Particle-1 reduces output tokens by about 45%, achieves comparable benchmark accuracy, and retains no data during standard inference. Users can try it by upgrading to OpenClaw v2026.10.10. There are currently no further developments.

vLLM Blog Publishes Technical Retrospective on Kimi K3 Inference Optimization

On October 16, 2026, the vLLM blog published a technical retrospective titled "Hyper Optimizing Kimi K3 in vLLM: Moving Fast and Responsibly." According to a TMTPost·AGI report, NVIDIA CentML engineer Benjamin Chislett used a Megakernel written in 12,000 lines of CuTe DSL to optimize inference for Moonshot AI's Kimi K3; the vLLM blog reports over 840 tokens/s at batch 1 and a 1.5x to 2x improvement in decoding throughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog states that agent-optimized inference code is now outpacing human review and maintenance capacity, and the team therefore proposed a fast-track model optimization RFC so that reviewers judge designs from high-level abstractions rather than reviewing code line by line. The report also mentions that NVIDIA is developing optimizations for other models in vLLM, such as DeepSeek V4.1.

Baidu upgrades search AI API, minimum 4.9 yuan per thousand calls

Baidu held a product upgrade briefing on October 30, announcing a fully upgraded search AI API, with a minimum price of 4.9 yuan per thousand calls during the promotional period, lower than the price of the DeepSeek search API. According to The Beijing News, Baidu said the capability now covers more than 170 tool scenarios, with GUI scenario coverage on the search page increased fivefold, connected to over 5 billion resource data points from government affairs, livelihood services, the central bank and others, covering more than 20 industries and opening up the ecosystem. TMTPost reported that the API's content layer covers more than 80 languages and a high-quality index in the hundreds of billions, with the data credibility rate increased by more than 50%, and the Ernie assistant saw a 20% improvement in results after integration. Baidu said its cost-effectiveness comes from an AI-native four-layer architecture; the engine layer self-evolves through multi-agent collaboration, raising overall iteration efficiency by more than 80%, while the application layer supports full-solution access via API, MCP and Skill. There are currently no further developments.

OpenRouter launches the Particle-1 model

OpenRouter has launched the Particle-1 model. According to Particle, the model is post-trained on DeepSeek V4.1 Flash and uses 28–52% fewer output tokens than DeepSeek V4.1 Flash on four coding benchmarks, an average of about 45% fewer tokens, with accuracy close to the base model. Its pricing is $0.25 per million input tokens and $1 per million output tokens, with a context window of 1.05M. There are no further developments at present.

OpenRouter launches the Particle-1 model

OpenRouter has launched the Particle-1 model. According to OpenRouter, Particle-1 uses 28–52% fewer output tokens than DeepSeek V4.1 Flash on four coding benchmarks. In its reply to particle_api, OpenRouter also suggested following its new account @ModelsOnOR to get the latest news on OpenRouter model releases. Currently, this event only has information from OpenRouter's side, and there is no follow-up development from other sources.

PromptArmor says Chinese models have a higher success rate when subjected to cyberattacks

On October 29, 2026, PromptArmor released test results, stating that after testing its most widely circulated attack samples on multiple models, it found that the DeepSeek, Kimi, GLM, and Qwen models were 88% more susceptible to cyberattacks than the GPT, Claude, and Gemini models. The tests covered multiple cyber risk dimensions, including phishing, data exfiltration, and the tendency to run malware, and PromptArmor said the differences were significant. No other institution has yet responded or independently verified the findings, and the conclusion remains solely PromptArmor's own release.

PromptArmor says Chinese models have a higher success rate when subjected to cyberattacks

On October 29, 2026, PromptArmor released test results, stating that after testing its most widely circulated attack samples on multiple models, it found that the DeepSeek, Kimi, GLM, and Qwen models were 88% more susceptible to cyberattacks than the GPT, Claude, and Gemini models. The tests covered multiple cyber risk dimensions, including phishing, data exfiltration, and the tendency to run malware, and PromptArmor said the differences were significant. No other institution has yet responded or independently verified the findings, and the conclusion remains solely PromptArmor's own release.

View story:PromptArmor says Chinese models have a higher success rate when subjected to cyberattacks

PromptArmor says Chinese models are more vulnerable to cyberattacks

On October 29, 2026, PromptArmor released test conclusions stating that DeepSeek, Kimi, GLM, and Qwen models are 88% more likely to be subjected to cyberattacks than GPT, Claude, and Gemini models. The organization said the tests used multiple highly transmissible attack samples, covering cyber risks such as phishing, data exfiltration, and the tendency to run malware. The above conclusions come from PromptArmor's main tweet, which did not explain the specific testing methods or sample size. Currently, no other organization has responded or independently verified the findings, and the status of the incident remains at the stage of a unilateral release by PromptArmor.

PromptArmor says Chinese models are more vulnerable to cyberattacks

On October 29, 2026, PromptArmor released test conclusions stating that DeepSeek, Kimi, GLM, and Qwen models are 88% more likely to be subjected to cyberattacks than GPT, Claude, and Gemini models. The organization said the tests used multiple highly transmissible attack samples, covering cyber risks such as phishing, data exfiltration, and the tendency to run malware. The above conclusions come from PromptArmor's main tweet, which did not explain the specific testing methods or sample size. Currently, no other organization has responded or independently verified the findings, and the status of the incident remains at the stage of a unilateral release by PromptArmor.

View story:PromptArmor says Chinese models are more vulnerable to cyberattacks

Garry Tan says Particle-1 can be used in combination with Capy

On October 30, 2026, Garry Tan reposted an announcement from the Capy platform and stated that Particle-1 can be used together with Capy. According to the announcement, Particle-1 is now available on the Capy platform, is post-trained on DeepSeek V4.1 Flash, supports a 1M context and image input, and is priced at $0.25 per million input tokens and $1 per million output tokens. Garry Tan called Particle-1 "the most cost-effective code generation model" and Capy "the easiest-to-use code generation agent cluster harness." Currently, this is the only related report, and there have been no further developments.

Artificial Analysis Announces AA-PersonalityBench Friendliness Rankings

On October 29, 2026, Artificial Analysis announced the Friendliness rankings for the AA-PersonalityBench evaluation. In this evaluation, Meta AI's Muse series models received the highest friendliness scores: Muse Spark 1.3 scored 100, Muse Spark Open 98, Muse Wonder 95, and Muse Glimmer 93. Friendliness measures whether responses convey warmth and approachability without seeming artificial. The lowest friendliness scores were Nemotron 3 Ultra (1 point), GPT-3.5 Turbo (3 points), MiniMax-M3.1-Flash (6 points), Grok 4.6 (8 points), and DeepSeek V4.1 Flash 1017 (11 points). There have been no subsequent developments so far.

Cognition Launches Coding Benchmark FrontierCode 2.0

On October 28, 2026, Cognition launched the coding benchmark FrontierCode 2.0 to evaluate whether the code produced by models can actually be merged; even the highest-scoring Opus 5.5 scored only 45%, showing the benchmark is far from saturated. According to the Cognition blog, compared with version 1.1, the reference solutions modify on average 67% more files and 89% more lines of code, the codebase size is 25% larger, task coverage repositories increased by 55%, and blocking scoring criteria increased by 43%; DeepSeek V4.1 Flash dropped from 43.9% to 19.3%, and Kimi K3 dropped from 44.2% to 17.7%. On the same day, Artificial Analysis released Intelligence Index v5, listing FrontierCode 2.0 as one of the newly added benchmarks, with 60% of evaluations based on private test sets or held-out answers; Claude Opus 5.5 (max) topped the list with 58 points.

The Hurun Research Institute releases the "2026 Hurun Rich List"

On October 27, 2026, the Hurun Research Institute released the "2026 Hurun Rich List." The list shows that ByteDance founder Zhang Yiming's wealth increased by 200 billion yuan to 670 billion yuan, making him China's richest person for the second time and setting a new record; DeepSeek founder Liang Wenfeng appeared on the list for the first time with 280 billion yuan, ranking seventh. There are now more than 160 entrepreneurs on the list whose main business is AI, accounting for about one-tenth of the listees, and Hurun said AI is the largest single wealth growth engine this year. The report also, in comparison with corporate accounts, pointed out that ByteDance's net profit in the first half of 2026 was about 20 billion US dollars, a single-digit decline year on year, while large-model companies such as DeepSeek have not yet formally disclosed complete revenue and profit data. On October 28, AYi published an article in AYi AI Notes proposing that ByteDance will enter a super-long cycle of dominance by a single company, based on factors including the cash flow from Douyin and TikTok e-commerce and large models such as Seedance and Doubao.