Latent Space highlights Fable 5’s agentic lead and Tencent’s Hy3 as open frontier shifts to deployment robustness

A Latent Space AINews roundup highlights Anthropic’s Fable 5 as the current leader on a new agent benchmark, AutomationBench, which evaluates agents across 657 tasks and 40 simulated SaaS apps with explicit objectives and guardrails. Fable 5 achieved 48.6% success, narrowly ahead of Opus 4.8 at 48.5%, with Gemini 3.5 Flash at 42.6% and GPT-5.5 xhigh at 42.1%. Open-weight models lagged, with GLM-5.2 max at 27.8%. The benchmark underscores that every model still violates business rules, and Gemini performed notably well on objective-per-guardrail-violation and cost efficiency.

Tencent released Hy3, a 295B Mixture-of-Experts model with 21B active parameters, 192 experts with top-8 routing, GQA, and a 256K context window, under the Apache 2.0 license. Community posts emphasized its competitive performance on reasoning, coding, and agentic tasks, with particular attention to reliability improvements such as tool-calling stability and anti-hallucination measures. Inference support was unusually mature at launch, with vLLM providing native tool-call and reasoning parsers, MTP speculative decoding, and validated support on NVIDIA and AMD hardware.

Tencent also upstreamed production kernels into vLLM main, including load-balanced decode scheduling and fused FP8 MoE serving, reporting gains up to 2.95x on mixed-length decode and latency reductions of roughly 24% time-to-first-token and 17% tokens-per-output-time versus default backends. Community reaction led Nous Portal to offer Hy3 free on its platform for two weeks.

Artificial Analysis introduced six domain-specific capability indices—Finance & Accounting, Legal, Healthcare & Medical, Strategy & Operations, Engineering, and Economics—arguing that single scalar benchmark scores are increasingly meaningless without cost per task. The indices show that rankings reshuffle sharply by domain and that the price/performance frontier has steepened, aligning with critiques that cost-efficiency must be part of model evaluation.

Two recent papers focus on memory and retrieval bottlenecks for persistent agents. A-TMA targets ‘ghost memory’ by improving conflict accuracy by +0.240 absolute on the LTP benchmark when added to Graphiti, while ReContext is a training-free long-context inference harness that replays model-internal evidence before answer generation, improving evidence utilization across eight 128K datasets. BlockSearch is cited for enabling million-token in-context retrieval.

Anthropic released mechanistic interpretability research claiming a global-workspace-like internal structure in Claude, centered on a small subset of activations they call J-space. The core claim is the identification of a privileged internal representational substrate available for report, modulation, and flexible reasoning, which researchers described as stronger evidence for a model ‘working memory’ or internal workspace than prior public work. Anthropic also shipped a Neuronpedia demo for open-weight models.

The interpretability results drew debate over consciousness framing, with supporters arguing for a functional analog of access consciousness and critics contending the company overclaimed by conflating privileged latent activation with consciousness. Some researchers highlighted practical safety angles, suggesting the workspace could surface hidden concepts, detect prompt injections, and expose internal sabotage-related features before they are verbalized.