Stay ahead of AI risks
with continuous
monitoring.

Qwen3.8-Max is a 2.4-trillion-parameter model with 95 billion active parameters. It is the first Qwen Max-class model announced for open-weight release. It demonstrates unusually long autonomous operation: one coding project ran for approximately 16 days and produced 265 commits, 127 pull requests and 151 issues without human assistance. In an AI research task, the model independently reproduced a paper and then ran 18 follow-up experiments, developing a method that improved on the reported baseline. The model can coordinate large agent systems; one financial-research workflow used approximately 330 subagents to conduct around 6,000 backtests. The report provides no dedicated safety evaluation, external red-teaming or assessment of frontier risks. Many results are internal benchmarks or selected demonstrations.

Kimi K3 is an open-weight, multimodal mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated parameters and a one-million-token context window. Its architectural and training changes reportedly improve scaling efficiency by approximately 2.5 times over Kimi K2. It achieves strong long-horizon coding and agentic performance, including 42% on SWE-Marathon, 91.2% on BrowseComp and 76.3% on an internal multi-agent orchestration benchmark. The model can sustain extended autonomous workflows: one case study reports that it designed and verified a small AI chip during a 48-hour run using open-source tools. The report contains a detailed cyber capability assessment, but no equivalent biological-risk, manipulation or comprehensive alignment evaluation.

DeepSeek-V4-Flash is an open-weight mixture-of-experts model with 284 billion total parameters, 13 billion activated parameters and a one-million-token context window. It was trained on 32 trillion tokens and supports multiple reasoning-effort settings, allowing additional test-time computation for harder tasks. At one million tokens, it reportedly uses about 10% of the inference computation and 7% of the key-value cache required by DeepSeek-V3.2. At maximum reasoning effort, it performs strongly on reasoning and coding benchmarks, including 88.1% on GPQA Diamond, 91.6% on LiveCodeBench and 79% on SWE-bench Verified. It performs less well than DeepSeek-V4-Pro on complex agentic tasks. The report provides no dedicated safety evaluation, red-teaming results or assessment of frontier-risk capabilities.

GLM-5.1 is primarily designed for agentic software engineering and long-horizon tasks, with stronger coding, terminal-use, repository-generation, and tool-use performance than GLM-5. The main capability advance is sustained performance over long tasks. In one experiment, GLM-5.1 continued optimizing over 600+ iterations and 6,000+ tool calls, repeatedly changing strategies based on observed results rather than plateauing early. It performs strongly on coding and cybersecurity benchmarks, scoring 58.4% on SWE-Bench Pro, 63.5% on Terminal-Bench 2.0, and 68.7% on CyberGym. On CyberGym, it slightly exceeded the reported scores for Claude Opus 4.6 (66.6%) and GPT-5.4 (66.3%). Long-horizon capability remains imperfect: the report identifies difficulties escaping local optima, maintaining coherence across thousands of tool calls, and reliably evaluating its own progress when tasks lack clear quantitative feedback. GLM-5.1 is open-weight and released under the MIT License, allowing local deployment through several inference frameworks.

M2.7 ran a constrained recursive self-improvement workflow, iteratively modifying its own scaffold over 100+ rounds and achieving a 30% performance improvement on internal evaluation sets. M2.7 handles 30–50% of an internal ML research workflow end-to-end, with human researchers intervening only for critical decisions; the model autonomously manages literature review, experiment pipelines, log analysis, debugging, metric tracking, merge requests, and smoke tests across teams. M2.7 achieved strong software engineering scores (56.22% on SWE-Pro, near GPT-5.3-Codex level) and is described as capable of deeply understanding production systems, reducing live incident recovery time to under three minutes.

Mistral Small 4 is an open-weight, multimodal model combining general instruction-following, reasoning, and agentic coding capabilities. It accepts text and images and has a 256,000-token context window. It uses a mixture-of-experts architecture with 119 billion total parameters, of which 6 billion are active per token. Users can adjust the reasoning level to prioritize either speed or more intensive analysis. Mistral reports that the model matched or exceeded GPT-OSS 120B on three reasoning and coding benchmarks while generally producing shorter outputs. On LiveCodeBench, it reportedly performed better while generating 20% less output. Compared with Mistral Small 3, the model reportedly reduced completion time by 40% or supported three times as many requests per second, depending on the deployment configuration. The model is released under the Apache 2.0 licence and can be fine-tuned or deployed locally. The report does not provide safety evaluations, red-teaming results, or detailed limitations.