MiniMax has released M3.1 Flash and confirmed that Space Bunny, which anonymously processed more than 200 trillion tokens, was an early version. The company reports a nearly 3-fold increase in online inference speed and a 70% reduction in overall usage costs for Agent tasks compared with the previous generation.
MiniMax has released M3.1 Flash and made it available through MiniMax API, M Plan, and MiniMax Code. API pricing for input, output, and cache reads is 1 yuan, 4 yuan, and 0.1 yuan per million tokens, respectively. The company confirmed that the model that anonymously processed more than 200 trillion tokens under the name Space Bunny was an early version of M3.1 Flash. According to MiniMax, the new model improves capabilities in Agent tasks, programming, Computer Use, and cybersecurity, with a nearly 3-fold increase in online inference speed and a 70% reduction in overall usage costs for Agent tasks compared with the previous generation. Model weights are scheduled for release on October 23; the source does not specify the year. The company also said that a Pro version aimed at pushing the upper limits of intelligence will be released later.
US orders federal civilian agencies to stop procuring and remove DeepSeek
要闻
The U.S. Department of Homeland Security issued an order barring federal civilian agencies from procuring technology products and services developed by DeepSeek and its affiliates and requiring removal of related products from their systems. The removal requirement also covers third-party products and services developed using DeepSeek implementations.
The U.S. Department of Homeland Security published an order notice on October 19 requiring federal civilian agencies to stop procuring and remove DeepSeek-related products and services; the source does not specify the publication year or effective date. The procurement ban covers information technology products and services developed by Hangzhou DeepSeek and its affiliates, while the system removal requirement also extends to third-party products and services developed using DeepSeek implementations. According to the notice, the secretary issued the order under the Federal Acquisition Supply Chain Security Act of 2018 after reviewing recommendations from the Federal Acquisition Security Council and related evidence. The U.S. Commerce Department, the Navy, and states including New York, Tennessee, and Virginia had previously prohibited the use of DeepSeek on government devices.
OpenCode announced that it has completed an update to DeepSeek V4.1 Flash, moving the model in both the Go subscription plan and Free to version 1017, with no action required from users.
OpenCode announced that it has updated DeepSeek V4.1 Flash in the Go subscription plan and Free to version 1017. According to the company, users do not need to take any action for this update.
Qoder offers free access to new DeepSeek model for a limited time
开发生态
Qoder is offering DeepSeek V4.1-Flash-1017 free for a limited time in its international and CN editions, covering both free and paid users of its individual plans with a billing multiplier of 0x. The offer ends on October 24; the source does not specify the year.
Qoder announced the launch of its 1024 warm-up promotion, making DeepSeek V4.1-Flash-1017 free in its international and CN editions from the day of the announcement through October 24; the source provides neither the announcement date nor the year of the promotion. According to the company, the offer covers both free and paid users of its individual plans, with the model’s billing multiplier set to 0x and no Credits deducted during the promotion. The company states that the model supports a 1M context window and allows thinking intensity to be set to the maximum.
Command Code has added MiniMax M3.1 Flash, codenamed Space Bunny, with free access for users on all subscription plans for a limited time. The model offers a 1 million token context window.
Command Code has launched MiniMax M3.1 Flash, and users on all subscription plans can now use the model free of charge for a limited time. The newly available model is codenamed Space Bunny and provides a 1 million token context window. Regarding its programming capabilities, Command Code officially states that MiniMax M3.1 Flash offers frontier-level coding capabilities.
TRAE announces Programmers' Day event with a million-point grand prize
开发生态
TRAE has announced its 1024 Programmers’ Day activities, with check-ins through its China-edition desktop app granting entry to a draw for a single prize of 1 million points. The program also includes in-person gatherings across 56 cities, online project showcases and sharing sessions, and a developer-themed film.
TRAE has announced a series of 1024 Programmers’ Day activities, with check-ins for the prize draw running from October 19 at 00:00 to October 24 at 18:00; the source does not specify a year. According to the company, users must log in to the latest version of the TRAE China-edition desktop app and check in to receive that day’s reward and qualify for the draw for 1 million points. There will be only 1 winner, scheduled to be drawn during an official livestream on October 24 at 19:00. During the same period, the community will host online project showcases, sharing sessions and other activities, while in-person TRAE Friends events will cover 56 cities. TRAE also plans to release the film 《致 AI 时代的开发者》 on October 23.
Devin introduces automatic usage limit increases based on usage efficiency
开发生态
Devin announced three enterprise features—efficiency scoring, efficiency-based automatic usage approvals, and Devin Coach—to help allocate spending based on recent usage performance. Scores range from 0–100, while Coach provides feedback on prompts before tasks run.
In this announcement, Devin introduced three enterprise features: Efficiency Score, efficiency-based automatic usage approvals, and Devin Coach; the source does not specify a publication date. According to the company, Efficiency Score uses recent sessions to identify useful outcomes, estimate the hours a human engineer would need for the same work, and combine efficiency with output volume into a score of 0–100. Useful outcomes extend beyond merged PRs to investigations, analysis, and other work that a classifier determines has made progress. Costs are measured in Agent Compute Units (ACUs), which include LLM and cloud infrastructure costs. Efficiency determines the starting score, while volume can add at most one-fifth of the gap between that score and 100. Devin Coach checks prompts before tasks run and may suggest planning first or switching to the default mode, but it does not silently rewrite prompts. Admins can view suggestion acceptance rates, daily trends, and breakdowns by type; engineers can ignore the default inline suggestions.
Ant Bailing has made five Ming-3.0 speech models available for download, covering speech generation, audio understanding, and speech recognition, with parameter counts ranging from 0.5B to 126B. Flash uses the MIT license and comes with two deployment options.
Ant Bailing has open-sourced five Ming-3.0 models and made them available on Hugging Face and ModelScope, covering TTS, audio understanding, and ASR. The TTS lineup includes Ming-3.0-tts-0.5B, with 0.5B parameters, and Tiny, with 7.9B total parameters and 1.3B active parameters. According to the company, the TTS models support voice design, control over vocal expression, and long-form speech generation, providing functions from configuring a voice to producing long-form content.
The audio understanding lineup includes Flash and Audio Tiny. According to the company, Flash has 126B total parameters and activates 6.5B at each inference step, supporting analysis of long recordings, spoken question answering, and understanding of music and environmental sounds. It can also generate tool requests for an App to execute. Flash uses the MIT license and offers deployment options through vLLM-Omni and SGLang. Audio Tiny has 9.3B total parameters and 2.7B active parameters, with a focus on efficient dialogue. ASR Tiny has 8.6B total parameters and 2B active parameters. According to the company, it supports recognition of dialects and accents and can use hotwords and previous dialogue to assist recognition without additional training.
fal has launched FLUX 3 Max, a video model developed with Black Forest Labs. According to the company, it scored 1,139 Elo, tying for #1 in a preliminary image-to-video evaluation, and generates a 10-second video in 12.45 seconds.
fal has made FLUX 3 Max available, although the source does not specify a launch date. Jointly developed by fal Research and Black Forest Labs on FLUX 3, the post-trained model supports both text-to-video and image-to-video generation at 720p. According to the company, FLUX 3 Max and H3 Max are both post-trained by fal Research and served through the same API, so switching requires only a change to the model ID. Both generation modes have the same pricing, billed per second of output video, with no minimum charge or subscription.
Preliminary Artificial Analysis AA-Video-I2V v1.1 results published by the company give FLUX 3 Max a score of 1,139 Elo. The company describes it as tied for #1 because its confidence interval overlaps those of the leading models; MiniMax H3, listed first on the board, scores 1,141 Elo. Seedance 2.5 also scores 1,139 Elo and takes 277.80 seconds to generate a 10-second video, compared with 12.45 seconds for FLUX 3 Max, which the company describes as about 22 times faster. Generation times are averages for 10-second, 16:9 videos at 720p/768p. The company also says post-training raised FLUX 3's score from 1,089 to 1,139 Elo, while inference optimization reduced generation time from 126.44 seconds to 12.45 seconds.
Inception has launched the diffusion-based LLM Mercury 3 with a 1M-token context and introductory input and output prices of $0.10 and $0.35 per million tokens, respectively. The company reports a speed of 1,068 tokens per second on NVIDIA GPUs.
Inception launched Mercury 3 with access available at launch through the Inception API Platform, OpenRouter, Baseten, and Vercel. The model is compatible with the OpenAI API, and its context increases from Mercury 2.5’s 260K tokens to 1M tokens. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens; a 50% launch discount brings those prices to $0.10 and $0.35, respectively. Cached tokens receive a 90% discount on the input price, and the API trial includes 100 million free tokens.
According to the company, Mercury 3 has more than 40 times the total parameters of Mercury 1 and runs at 1,068 tokens per second on NVIDIA GPUs. It says the model outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4–8 times faster. Early-access partners already use it for voice and chat support, real-time tutoring, data engineering, and web navigation. TinyFish AI cofounder Shuhao Zhang said Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in the company’s production web-agent harness. Enterprise deployments support dedicated capacity, autoscaling, compliance controls, and configurable data retention.
Decision 3.0 multimodal decision-making model released
模型发布
The vLLM Semantic Router team has released Decision 3.0, a multimodal decision model family with six sizes ranging from 0.6B to 27B. It outputs decisions and probabilities rather than generated text and can now run on users’ own infrastructure through the System One API.
The vLLM Semantic Router team has released the Decision 3.0 (D3) family, with deployment recipes for all six sizes merged into vLLM Recipes and self-hosting now available through the System One API using vLLM nightly. The family includes 0.6B, 0.8B, 2B, 4B, 9B, and 27B models, accepts text, JSON, image, and video inputs, and returns selections, yes-or-no judgments, scores, and probabilities for each option without generating text. On performance, vLLM says the 27B model ranks first on both the text and vision leaderboards of Jev Decision Index 0.3.2, scoring 66.3 and 71.3, respectively. According to vLLM, in a test using a single AMD Instinct MI325X and 64 concurrent clients under the same mixed workload, serving this model with vLLM delivered 1.8 times the request throughput of the server bundled with its checkpoint.
Xing4.0-29B-Intent goes open source, combining intent recognition and parameter extraction
模型发布
China Telecom Artificial Intelligence Technology Co., Ltd. has open-sourced Xing4.0-29B-Intent, a 29B MoE model with 4B activated parameters for intent recognition and parameter extraction, alongside a dataset containing approximately 2.75 million samples.
China Telecom Artificial Intelligence Technology Co., Ltd. has open-sourced Xing4.0-29B-Intent, although the source does not specify a release date. The model weights, public data, and code, including the Python SDK, are released under the Apache License 2.0. Built on Xing4.0-29B-A4B, the model uses a MoE architecture with 29B total parameters and 4B activated parameters. According to the project, it follows candidate categories, field names, and descriptions defined in an input schema to convert natural language into intents and parameters in a unified JSON format. Its applications include intelligent assistants, request routing, and document information extraction; args returns [] when no parameters can be filled.
The companion Xing4.0-29B-Intent-oss dataset combines 61 public datasets with China Telecom's in-house smart device assistant data in a unified schema-guided representation. It contains approximately 2.75 million samples covering 4 task types, 19 domains, and 52 languages, along with a standard smart assistant label set covering 170+ intents and 148 slots. The companion XingIntent SDK calls the service through an OpenAI-compatible API and automatically assembles prompts and parses results. The project provides the base model's deployment documentation for frameworks including vLLM and SGLang, and states that performance in specific use cases needs validation with the target data and schema.
