Refuseless launched an abliterated version of GLM 5.3, an open model with refusals removed, served through OpenAI-compatible endpoints with zero prompt retention for cybersecurity teams. The provider …
Google restricted live access to its Gemini 4 Argon foundation model to vetted enterprise organizations and cybersecurity defenders under the Fairwind Program, citing the model's proficiency in revers…
Together AI Staff AI/ML Engineer Zain Hasan told Let's Data Science that the company's reported 83% coding-task success rate at $3.35 per task for a DeepSeek-first, GPT-second cascade was reconstructe…
Xiaomi released MiMo 2.6 this week with an unusually open reinforcement-learning process, publishing an RL dashboard, a technical report, details of how it built its RL environments, and the RL enviro…
Contrastive Language Models (CLMs) introduced CLM-8B, an 8-billion-parameter System One model trained with a contrastive learning objective that connects states and actions, pre-trained on 60M Nemotro…
DeepSeek released V4.1 Flash under MIT open weights, priced at $0.003 per 1 million cached tokens, and the model scores 74.2% on DeepSWE, beating Claude Opus 5. The release covers what changed in the …
Jacky Kwok announced Contrastive Language Model (CLM), an 8-billion-parameter "System One" model trained with a contrastive learning objective linking states and actions, which delivers up to 9× faste…
GitHub released Project HydraFusion, a research preview that routes GitHub Copilot CLI coding tasks across multiple models via three execution patterns — Single, Cascade, and Critique — instead of a s…
Xiaomi's MiMo-V2.6 Pro ranks No. 1 on open-weight benchmarks by weighted average despite using a classic Grouped Query Attention architecture with Sliding Window Attention at a 128-token window, accor…
Developer 0xSero released model-toolkit, a GitHub repository of standalone Python tools for evaluating, observing, pruning, and quantizing language models, drawn from the REAP and EXL3 experiments inc…
An audit of DeepSWE v1.1 found defects or ambiguous requirements in 37 of the benchmark's 113 tasks (32.7%), based on a review of all 372 recorded failures for the Opus 5, Sol and Fable 5 models. The …
Together published a migration guide for companies moving from closed-source to open-source AI models, arguing the process can take weeks to months rather than months to years when a managed service i…
Specific Labs' Real-SWE benchmark, published in September 2026, found that the best frontier coding agent resolves only 38.8% of tasks on private enterprise codebases, meaning roughly six in ten agent…
DeepSeek released DeepSeek V4.1 Flash, a 552 billion parameter mixture-of-experts model under an MIT license that activates only 8 billion parameters during prefill and 16 billion during decode, cutti…
A blog post examines three categories of benchmarks — performance "napkin math" estimates, AI model evals including DeepSWE and Senior SWE-Bench, and winter tire comparisons — arguing each is misused …
Deepseek released V4.1-Flash, a multimodal model with 552 billion parameters that cuts KV cache memory to a quarter of its predecessor's requirement, with only 16 billion parameters active per token. …
A new approach to evaluating AI coding agents, called behavioral evaluations, is gaining traction as a more reliable alternative to end-to-end benchmarks, according to a technical article on harness e…
GitHub has released Project HydraFusion, a research preview in GitHub Copilot CLI that builds a per-request execution plan across multiple models and providers, selecting from Single, Cascade, or Crit…
Mouse, an agent harness built on the OpenCode engine, scored 24/30 on the FrontierHarness benchmark, the highest pass rate among 13 harnesses tested with Kimi K3 on 30 coding tasks. Mouse passed 18/21…
OpenAI released GPT-6 Astra on 3 September 2026, its most capable model to date, with agentic scores of 74.1% on DeepSWE v1.1 and 72.6% on OSWorld 2.0. The model is the first to reach OpenAI's 'Critic…