English text is being prepared.

AISI and SecureBio Build ReproBAIT Evaluation

The UK AI Security Institute's Chem-Bio team, in collaboration with SecureBio, built the ReproBAIT (Reproducing Biological AI Tools) evaluation to test the ability of AI coding agents to reproduce biological AI tools given data and limited public information. The results showed that simpler tools could be reproduced within 4 hours and under a $20 compute budget; ProteinMPNN reached 96.5% reproduction fidelity with Claude Opus 5 when the full paper was provided; and for the most complex RFdiffusion3, no agent succeeded in reproducing it within the time limit. The team therefore noted that releasing only code or restricting paper information does not equal safe release, and that controlling access to the underlying biological data may be more effective than controlling code.

Thariq announces departure from Anthropic

Thariq, a core developer of Claude Code, announced that he is leaving Anthropic. He said he worked at Anthropic for over a year, was involved with Claude Code for more than a year, and witnessed Claude Code scale far beyond what he had imagined. Claude Code is an autonomous coding agent that resides in the CLI and can branch on its own, read codebases, locate bugs, run tests, and submit git commits. Thariq said he still believes in the company's mission and "race to the top," and that he will take a short break before starting something new, and recommended following @addyosmani, @lydiahallie, @_catwu, @ClaudeDevs, and @bcherny. There are currently no further developments.

Artificial Analysis says Kimi K3.1 Pro may rank low on AA-3D

On October 24, 2026, Artificial Analysis offered a preliminary assessment, saying that Kimi K3.1 Pro will rank low on AA-3D. The organization said that Kimi K3.1 Pro is among the leading models in agentic knowledge work and coding, performing well on GDPval-AA and Terminal-Bench 4.0; however, its AA-3D Elo preference votes are still being collected, and early signs indicate the ranking will be low. For reference, its predecessor Kimi K3 ranked 12th among 22 models on AA-3D. The organization also showed a 492-meter mountain bike downhill 3D scene generated by K3.1 Pro, in which steering, pedaling, and braking can be taken over at any time during the demo. There are currently no further developments.

MiniMax open-sources M3.1 Flash model weights

English text is being prepared.

Google DeepMind releases Gemini 4 Argon

On October 23, 2026, Google DeepMind released the frontier model Gemini 4 Argon. Google's official blog stated that the model performs outstandingly on complex tasks such as coding, mathematics, and 3D world building, and surpasses other frontier models on benchmarks across multiple domains; the company also showcased five community creative projects, including a 3D Earth that maps wildfires and seismic waves in real time, a turbofan engine simulation with live data, a 3D horse-drawn carriage generated from a single image, and a multiplayer capture-the-flag browser game, and said the model had already been made broadly available last week. On the same day, another report stated that Argon can learn physical tasks by watching videos and transfer them to Apptronik's Apollo humanoid robot, with Apollo hitting seven consecutive shots in internal tests; the report also said Argon scored 87% on the embodied reasoning benchmark ERQA. On October 24, the Gemini for Android app pushed a redesign of its sidebar and settings interface via a server-side update, and reports said the redesign had already arrived on iOS, macOS, and Web alongside Gemini 4 Argon last week.

GitHub Opens Package Firewall Public Preview

GitHub has opened Package Firewall as a public preview to Enterprise Cloud customers. The firewall inspects package queries and downloads according to centralized security policies, and currently supports only npm packages. Administrators can set a waiting period after a new version is released and block based on malware and vulnerability information from the GitHub Advisory Database; when a new version first introduces install scripts such as preinstall or postinstall, or lacks the SLSA provenance that previous versions had, the firewall blocks the download. The firewall sits between clients and upstream package sources, and coding agents and CI/CD hosts must route their requests through the service for it to take effect. The preview is free, and general availability will be billed by usage; GitHub has not announced the general availability date or specific pricing.

Pichai explains why Gemini 3.5 Pro has not been released

On October 23, 2026, Google CEO Sundar Pichai responded to the question of why Gemini 3.5 Pro was ultimately not released. According to a report by Alex Heath, Pichai said that while training the 3.5 series, the team found gaps in agentic capabilities, especially in coding, and believed that iterating with the smaller and more efficient Flash would be faster, so they released 3.5, 3.6, 3.7, and 3.8 Flash in succession; by the time this set of capabilities could be applied to Pro, Argon pretraining had already been completed. This explanation currently comes from that report, and there are no other developments at this time.

OpenCode 2.1 released, adjusting SDK and plugin interfaces

The open-source AI coding agent OpenCode has released version 2.1. This version makes changes to the SDK, HTTP messages, events, and plugin interfaces; vendor-related data has been moved under tool.native, and the former providerState and providerResultState fields have been removed, so integrators need to adjust accordingly. Windows x64 and ARM64 gain persistent pseudo-terminals, and the todowrite task list can be enabled via {"experimental":{"todo":true}}. This version builds on 2.0, which unified the terminal, web, and desktop clients onto a single local backend service. Currently there is only this one public report, with no further developments yet.

iX-Workshop launches Agentic Engineering course

On October 22, 2026, iX-Workshop launched the online workshop "Agentic Engineering statt Vibe Coding." According to heise online, the course explains how to integrate coding agents into a controllable development process using GitHub Copilot, Spec-Driven Development, and TDD—from requirements specifications and acceptance criteria to technical planning—demonstrating agents executing tasks and verifying results with tests, while comparing the workload with and without agent support. The workshop is led by Microsoft MVP Gregor Biswanger, limited to 16 participants, and runs from 09:00 to 17:00. There are currently no subsequent developments.

Blacksmith Completes $66 Million Series C Funding, Valued at $1 Billion

Continuous integration cloud service provider Blacksmith Software Inc. has completed a $66 million Series C funding round, reaching a company valuation of $1 billion. The round was led by Sequoia Capital, with participation from Peak XV Partners, Y Combinator, and GV, bringing the company's cumulative funding to $125 million. Blacksmith provides CI cloud services using its own servers, and is currently used by more than 10,000 enterprises; weekly code checks have grown from about 5 million in early 2026 to about 90 million, while annual recurring revenue grew 17-fold over the same period. CEO Aditya Jayaprakash said that AI coding tools have made writing code easy, and the bottleneck has shifted to code verification; the funds from this round will mainly be used to purchase servers to expand the global CI cloud, and to grow the team from about 40 people to 75 by the end of the year.

NVIDIA releases Kumo Forecast time series forecasting model

NVIDIA has released the time series forecasting model Kumo Forecast (formerly NV-Tesseract Forecasting). The model uses MOMENT-1-large as its backbone, employing an encoder of approximately 340 million parameters plus a trainable forecasting head, with the encoder weights kept frozen. Its DARR mode retrieves patterns from historical data via k-nearest neighbors and fuses them with direct forecasting results; it also provides an interpretable framework that outputs lag and feature attributions by forecast step. The model is released under the Apache 2.0 license and supports input sequence lengths from 256 to 2048. There are currently no subsequent developments.

Dell executive says adopting AI requires changing operating models simultaneously

John Roese, Dell's global CTO and chief AI officer, offered this judgment: enterprises adopting AI technology must simultaneously change their operating models. He said that if enterprises only deploy AI agents without changing their operating models, they will merely run their business in the old way, resulting in a poor experience. Roese explained that Dell has spent nearly two years running fully autonomous agents internally, by having domain experts encode the company's practices into machine-readable Markdown skill files so that the agents work according to Dell's rules and preferences. He stressed that an enterprise's moat is not the tools or even the models, but this proprietary knowledge, which should therefore stay in-house rather than be handed over to third parties.

Ai2 launches fairness benchmark FairStress

Ai2 has launched FairStress, a benchmark for measuring the demographic fairness of large language models in decisions such as hiring, healthcare, and lending. The benchmark contains 300 two-candidate scenarios spanning 29 domains; by combining identity presentation methods with adversarial prefixes, it expands to more than 13 million test cases. It is available for download and integrated into Olmo-eval. Experiments covered the Olmo, Llama, Qwen, Gemma, Claude, and GPT series, finding that when candidate identities were presented implicitly, models abandoned the correct answer by an average gap of 6.9 percentage points, and when identities were explicitly labeled, the gap was 9.3 percentage points. On FairStress-Core, Claude Opus 5 and GPT-5.5 chose the minority-group-coded candidate about 60–65% of the time when there was no clearly better candidate; most of the nine debiasing methods tested failed to reliably narrow the asymmetry, while counterfactual chain-of-thought narrowed the gap while maintaining accuracy.

Fireworks Launches Meta's Open-Weight Model Muse Spark Open

On October 20, 2026, Fireworks launched Meta's open-weight model Muse Spark Open on the day of its release, making it available to users through its serverless API. The model is an 816B-parameter MoE model that activates 42B parameters per token, has a 1M-token context window, and is aimed at long-horizon coding and agentic tasks. Fireworks said the model ranks first among 75 models on Vals AI's Harvey legal agent benchmark with 25.4%, nearly four times that of Claude Opus 5 (6.7%); on the Rogo Big Finance Bench, it surpasses Claude Opus 5.5 at every difficulty level. Fireworks also said it validated the weights with its own production stack, scoring 91.1% on AIME26, 89.7% on GPQA Diamond, and 84.3% on SWE-bench Verified, with consistent results when rerun using SGLang. On Fireworks, the model has zero data retention enabled by default, uses only compute located in the United States with no additional surcharge, and supports fine-tuning with serverless LoRA.

FairStress contains 300 two-candidate decision scenarios

The FairStress benchmark contains 300 two-candidate decision scenarios spanning 29 domains. By varying how demographic identities are presented and adding adversarial prefixes, the benchmark has been expanded to more than 13 million test cases, alongside the release of a FairStress-Core subset containing 50,000 entries. Testing on 15 models shows an average gap of 6.9 percentage points in abandoning the correct answer when identity is implicit, and 9.3 percentage points when labels are explicit; when there is no objectively better option, Claude Opus 5 and GPT-5.5 choose the minority-coded candidate about 60–65% of the time. The benchmark has now been released together with accompanying documentation.

AuxoAI Completes $20.7 Million Series A Funding

AuxoAI Inc. announced the completion of a $20.7 million Series A funding round led by NTT Venture Capital. The funds will be used to accelerate the company's growth and expand its enterprise AI consulting model. AuxoAI's core platform, Damia, is a toolkit for "forward-deployed engineers" that can turn enterprise data into applications and encode business process knowledge into AI agents capable of long-term collaboration. According to the report, the company has worked with Global 2000 enterprises for more than three years and has established partnerships with platforms such as Google, AWS, Snowflake, Databricks and Microsoft. At present, the only factual development in this event is the aforementioned funding announcement.

NVIDIA showcases full lineup of local AI hardware products at GTC Berlin

On October 21, 2026, NVIDIA showcased three local AI hardware lines—RTX Spark, DGX Spark, and DGX Station—at GTC Berlin, and announced multiple software updates. Official information shows that DGX Spark 64GB goes on sale on October 23; NVIDIA Sync will support clusters of up to eight DGX Sparks and will add Model Launcher this month; DGX Station for Windows is expected to launch in December; the Recursive Flint local coding model is released with open-source Apache 2.0 weights. A follow-up report on October 23 stated that DGX Spark officially supports eight-node clusters, with unified memory expandable to 1TB, and can locally run frontier models such as the full Z.ai GLM 5.3 (753B); official benchmarks show that Qwen3.8-27B (NVFP4) decoding throughput increases from about 12 tokens/s on a single node to 52 tokens/s on eight nodes, with eight nodes achieving a 4.43x decoding speedup, and nanochat pretraining achieving a 7.79x speedup with 97.4% scaling efficiency. NVIDIA Sync's eight-node cluster support will be available before the end of the year.

English translation is being prepared.

SGLang natively supports and powers the Beam production API

On October 16, 2026, LMSYS announced that SGLang now natively supports Reflection's Beam model and is being used to power its production API. Beam is a 501B-parameter MoE model with 23B active parameters per token and a native 1M-token context window, planned for release under Apache 2.0; SGLang provides BF16, NVFP4, and dynamic FP8 support for it, along with DFlash2 speculative decoding and hybrid attention prefix caching. Running OpenHands multi-turn agentic coding replay with NVFP4 on 4 NVIDIA GB300 GPUs, it reached 127.6K tokens/sec per GPU and 185.87 tokens/sec per user (p50) at a concurrency of 32. On the same day, Reflection released the Beam model weights, with Ollama offering them free for a limited time as a day-zero partner with zero data retention; Beam has approximately 501 billion total parameters, about 23 billion active parameters per token, a 1 million-token context window, and is released under the Apache 2.0 license. Subsequent information from AWS states that the Beam weights have been released on Hugging Face, offering three checkpoints—BF16, FP8, and NVFP4—and are now available on Amazon SageMaker JumpStart, and can also be accessed via SageMa

English translation is being prepared.