SURL AI News
繁中English日本語

Archive

推論系統

推論 runtime、量化、 serving、KV cache、成本與效能。

GitHub Repo

Nanbeige4.2-3B Compresses an Agent Model With a Looped Transformer, Supporting a 256K Context With 3B Non-Embedding Parameters

Nanbeige Lab has released Nanbeige4.2-3B, pretrained from scratch on 28 trillion tokens. It repeatedly applies the same set of Transformer layers to increase computational depth without adding parameters. In official agent benchmarks, it generally outperforms the larger Qwen3.5-9B, but the results depend heavily on Nanbeige's own execution framework and still require independent reproduction.
最新模型

SGLang 0.5.16 Dynamically Adjusts Speculative Verification Length Based on Confidence and Overhauls Caching for Hybrid Architectures

The new SGLang release introduces DSpark, which changes the verification window for each round based on the draft model’s confidence, and makes UnifiedRadixTree the default cache for sliding-window, Mamba, and sparse-attention models. The release also removes several quantization backends and changes the rollout transport format, so compatibility must be checked before upgrading existing services.
最新模型

LLaDA 2.2 Lets Diffusion Models Add and Delete Tokens Mid-Generation, Targeting Iteratively Correctable Agent Loops

inclusionAI has open-sourced the 100B-class LLaDA2.2-flash, using Levenshtein Editing to let parallel decoding insert, delete, and replace content. The model outperforms the company’s own autoregressive baseline in some agent benchmarks and throughput tests, but still trails in code repair and function-calling accuracy, while serving support remains incomplete.
推論系統

AMD Open-Sources Hyperloom to Automatically Rewrite and Validate GPU Inference Optimizations Through a Multi-Agent Closed Loop

AMD has released ROCm Hyperloom 1.0.0a1, enabling agents to handle everything from performance profiling to kernel and host-code modifications, then decide whether to retain changes based on correctness and speed tests. The initial release supports only three Instinct GPUs, and the claim of reducing optimization time “from weeks to hours” remains AMD’s own; adopters will need to reproduce the results with their own models.
AI 研究與訓練

VIGOR Dynamically Allocates RLVR Rollouts Based on Reward Variance, Cutting Sampling for Mathematical Reasoning Training by Up to 2.3×

A new study replaces the fixed number of reasoning trajectories generated for every problem with a strategy that concentrates additional compute on samples exhibiting greater reward disagreement. The authors report higher rollout efficiency on mathematics and coding tasks, but have not released their code, while actual GPU time and scalability to larger models remain to be verified.
最新模型

AMD Opens the Full Instella-MoE Training Pipeline, With the 16B Model Activating Only 2.8B Parameters per Token

AMD has released six Instella-MoE checkpoints spanning pretraining, mid-training, long-context extension, SFT, DPO, and reinforcement learning, along with the training frameworks, data recipes, and inference code. The model uses Gated MLA and FarSkip-Collective to reduce attention and cross-GPU communication costs, but a formal technical report and independent evaluations are still pending.
最新模型

UK–US Joint Evaluation: Kimi K3 Completes a Simulated Enterprise Intrusion, but None of 41 Exploit Attempts Achieves Arbitrary Code Execution

Preliminary testing by the UK AISI and the US CAISI shows that Kimi K3 scored 32% on ExploitBench, outperforming GLM-5.2 but trailing the tested US frontier models by a wide margin. The model also completed a 32-step simulated enterprise attack once, but the limited evaluation conditions and sample size do not support direct conclusions about real-world attack success rates.