Skip to content
Top 9 stories discussed in the past 48 hours · Updated Oct 20, 11:50 AM ET

China AI · Trending

Startup BuzzNew launches, every week
No.01Surging↑ 116%

DeepSeek open-sources DeepInfer, its inference framework for DeepSeek-V4.1-Flash

AIDeepSeek has open-sourced DeepInfer under the Apache-2.0 license, the in-house framework it says runs DeepSeek-V4.1-Flash at scale with high throughput and low latency. DeepSeek says it is sharing its production stack to help lower compute costs and energy use. A repost of DeepSeek's announcement gives 325 tokens per second for a single chat and about 1.7M tokens per second across 64 Nvidia Blackwell GPUs. The repost cites no independent measurement. Meng Shao says the release covers offline inference, is built primarily on vLLM with code from SGLang and TensorRT-LLM, and supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Shao says batch-invariant deterministic output works only on NVIDIA. Rohan Paul says the release omits a KV cache system, prefill and decode disaggregation, and load balancing across replicas. vLLM says it will apply ideas from DeepInfer to improve DeepSeek V4.1 Flash inference, and SGLang says it plans to bring some of its large-batch and multi-GPU ideas into SGLang.

24-hour heatPeak 153 · Now
24 hours agoNow
sources
11
participants
17

In the past 48 hours

AILatest follow-upPost from wh

wh, Yuchen Jin and 9 more sources ·17 participants in 48h
150
Heat index
No.02Rising↑ 28%

DHS orders US civilian federal agencies to remove DeepSeek products and stop buying them

AIThe Department of Homeland Security (DHS) has ordered all US civilian federal agencies to remove DeepSeek products and third-party services built on DeepSeek implementations from their IT systems, and to stop buying them. CNBC reports the order was originally issued July 24 and made public on a Monday. It excludes the Department of Defense, and national security systems and secure compartmented information systems are also outside it, according to other reports. DHS cites the 2018 Federal Acquisition Supply Chain Security Act and warns the technology could transfer sensitive data to China or track users, according to SiliconANGLE and The Guardian. Agencies can request exceptions by submitting a written justification to DHS within 60 days of the order's issuance, according to CNBC. China's Foreign Ministry opposed the move, saying it overstretches national security and politicizes technology trade, according to TechCrunch. The Commerce Department and the U.S. Navy had restricted DeepSeek earlier, and New York State banned it from government devices, according to SiliconANGLE.

109

7 sources · 16 participants in 48h

No.03↓ 16%

MiniMax releases M3.1 Flash at $0.15/$0.60 per million tokens; weights due October 23

AIMiniMax releases M3.1 Flash on its API at $0.15 per million input tokens and $0.60 per million output tokens, with cache reads at $0.015 per million. MiniMax says the model costs one-third as much as its M3 model and runs three times faster. Artificial Analysis scores it 42 on its Intelligence Index, 13 points above M3, at about $0.22 per task against M3's $0.51. MiniMax plans to release the weights on October 23. Before launch, the model ran as the stealth model Space Bunny on OpenRouter and OpenCode. Several platforms offered free access for limited periods: through October 22 on Vercel AI Gateway when served by MiniMax, 72 hours on OpenRouter, three days on OpenCode, and a limited time in Cline and Command Code. Ollama Cloud hosts the model in the US with zero data retention, and Ollama says it is free there for three days. Testing Catalog, citing MiniMax's API reference docs, describes a 1M context window and tunable thinking depth. MiniMax's launch post does not mention these.

102

17 sources · 20 participants in 48h

Keep reading No.04–09

  1. 04

    Former OpenAI engineer Alfozan held pending trial over hidden office recorder

    AIA federal magistrate judge, Alex Tse, ordered former OpenAI engineer Abdulrahman Alfozan held in custody pending trial on one count of using a device to intercept oral communications. Prosecutors say he hid a recorder in an OpenAI executive conference room in San Francisco, which OpenAI found on Sept. 30, and that he used his wife's ChatGPT account to research building it. The charge carries up to five years in prison and a $250,000 fine, and prosecutors say additional charges are possible. Prosecutors cite one-way flight bookings to Saudi Arabia and a Saudi royal's attempt to recruit him to a Saudi AI initiative as flight-risk evidence. Alfozan's lawyers say he bought a one-way ticket on Sept. 13, before OpenAI found the device, and that the Oct. 12 flight matched the date his wife had to leave the US under immigration rules. Tse refused a $1 million bond offered by the Saudi government. OpenAI says it fired Alfozan, referred the matter to the FBI, and has seen no sign that model weights, source code or customer data were accessed.

    4 sources in 48h56↓ 16%
  2. 05

    DeepSeek V4.1-Flash-1017: MIT-licensed open weights, unchanged API pricing, cheaper Search API

    AIDeepSeek releases DeepSeek-V4.1-Flash-1017, an update to V4.1 Flash, with MIT-licensed weights on Hugging Face and ModelScope and unchanged API pricing of $0.30 per million input tokens and $1.20 per million output tokens at peak, half off-peak. DeepSeek says the model beats its predecessor on all 11 benchmarks it reported, including 77.0 on DeepSWE v1.1. Artificial Analysis scores it 45 on its Intelligence Index, six points above V4.1 Flash. Thomas Wolf reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and questions whether those numbers hold outside DeepSeek's own eval setup. DeepSeek also opened its Search and Fetch APIs to all developers at $0.75 per 1,000 successful requests each. Server-side calls through DeepSeek Harness, Claude Code or Codex are free until October 31, 2026, according to The Decoder, and November 1, 2026, according to MarkTechPost. Unsloth says a 1-bit quantized GGUF cuts the model from 508GB to 250GB and needs a device with 128GB+ RAM, retaining about 88.2% accuracy by its own estimate. OpenRouter, Vercel's AI Gateway, LM Studio and OpenCode have added the model, and Databricks has launched it as a new endpoint.

    8 sources in 48h49↓ 16%
  3. 06

    SGLang Low-Latency Mode raises DeepSeek-V4.1-Flash token speed on NVIDIA and AMD

    AILMSYS Org announces SGLang Low-Latency Mode, built with NVIDIA and AMD. In SemiAnalysis's AgentX trace replay on DeepSeek-V4.1-Flash with DSpark, P90 per-user token speed rises from 671 to 928 tok/s on NVIDIA GB300 (1.38x) and from 332 to 516 tok/s on AMD MI355X (1.56x). SemiAnalysis says merged megakernel support brings GB300 close to 1000 tok/s per user and works out of the box on SGLang Docker images. SGLang says each vendor ships its kernels in a separate repo, with a plugin system that falls back to its general path for steps without a custom kernel. LMSYS says low-latency kernels for GLM-5.3, K3, and MiniMax M3 are still in progress and will be added next.

    3 sources in 48h21↓ 16%
  4. 07

    Meta releases Muse Spark Open, an 816B-parameter open-weight coding modelRising

    AIMeta has released the weights of Muse Spark Open, an 816-billion-parameter model with 42 billion active per token and a 1-million-token context window, built for coding and agentic work. Artificial Analysis scores it 39 on its Intelligence Index, the highest of any US open-weights model, but says it trails Chinese open-weights leaders on agentic evaluations. Meta reports 89.2 on MCP-Atlas and says Chinese open models lead on Terminal-Bench 2.1. The license is free but requires products to display "Built with Muse Spark Open." Providers with more than $20 million in annual revenue that sell it as an API need a separate agreement with Meta. Hosts including OpenRouter, Together AI, Fireworks AI, Baseten and Modal list it at $1.25 per million input tokens and $4.25 per million output tokens. Meta and Apple say two M5 Ultra Mac Studios can run the model locally with MLX, while another report says local use needs at least 1 terabyte of RAM.

    2 sources in 48h13↑ 131%
  5. 08

    Kling AI releases Kling 4.0 video model; API and 4K output still coming soon

    AIKling AI says Kling 4.0 is now fully available through its creation interface at kling.ai and through MCP and CLI access. Leiphone reports the model generates up to 30 seconds of native video, accepts up to 15 mixed reference materials and 10 keyframes, and outputs 4K, 1080p and 21:9. The Decoder reports two-channel stereo audio, lip sync across more than ten languages, and a lighter Kling 4.0 Flash variant capped at 720p and 20 seconds. Kling AI says 10-bit HDR output at 4K and 1080p, and API access for enterprise customers and developers, are coming soon, with no date given. Fal says Kling 4 is coming soon to its platform. Runware says Kling 4.0 and a Kling 4.0 Flash variant are coming soon to its platform, with 4K output and stereo audio, and no dates were given for either platform. Runware's Flash variant is not confirmed in Kling AI's own posts.

    2 sources in 48h12↓ 16%
  6. 09

    Goodfire's Cammarata claims most neurons in large AI models are legible

    AINick Cammarata, a former OpenAI interpretability researcher now at Goodfire, says most neurons in models up to trillions of parameters are legible, with about 90% of neurons in Qwen-32B having one job, compared with about 2% in his earlier OpenAI project using primitive tools. He argues that ambitious interpretability may be solved in time, possibly soon, as automated scientists improve. Researchers disagree on how much this changes the field. Neel Nanda of Google DeepMind says the work offers no real breakthrough and doubts it addresses language models, and Chris Potts of Stanford and Jacob Hilton also call it a hunch that does not solve core problems. Zvi Mowshowitz calls it a real update but not a solution, citing a neuron that responds to both Colorado and Tornado. The companion technical paper had not been released in the reports reviewed.

    2 sources in 48h10↓ 16%
How is heat calculated?About this ranking

Heat counts independent accounts and organizations discussing the same story. Duplicates count once, and activity decays with a 24-hour half-life. It measures discussion, not article quality.

The ranking and the source and participant counts on each card cover the past 48 hours; a story’s own page counts all of its articles since the first one. Trends compare only sources monitored throughout the period; they reflect our coverage, not the whole internet. No trend appears without comparable history.

Source lists show only publicly readable coverage. Participants also include accounts and organizations that contribute to heat alone. Channels from one organization may be grouped, so participant counts may be lower than source counts. Open a story to read its coverage.

Surging
Discussion is growing fast
New
First reported within 6 hours
Rising
Discussion is still growing