Archive
推論 runtime、量化、 serving、KV cache、成本與效能。
最新模型
A new study increased the generation limit for four open-weight models from 64 to 4,096 tokens and found that the leading model changed on all three reasoning benchmarks. Even after truncated outputs were excluded, 3% to 19% of questions changed from correct to incorrect as the budget increased, showing that a single token setting is insufficient for deployment decisions.
最新模型
Liquid AI has released LFM2.5‑VL‑3B, a 3.1B-parameter model that adds screen grounding, layout OCR, multi-image understanding, and function calling. The company reports 228 tokens/s on an Apple M5 Max and roughly 11,000 tokens/s in high-concurrency throughput on a single H100, though the figures have yet to be independently reproduced.
最新模型
InclusionAI has released the 7.9B-parameter Ling-3.0-tiny, which combines linear attention with sparse MoE and offers BF16, FP8, and INT4 weights. The company says FP8 reaches 86 to 90 tokens per second on an M4 Pro, but vLLM and Ollama support has yet to fully land in stable upstream releases.
推論系統
Rather than merely asking models to think longer, ThinkRetrieve retrieves complete solutions to similar problems at each reasoning boundary and inserts them into the next segment of the reasoning context. All five tested models, ranging from 1.5B to 8B parameters, outperformed sequential test-time compute, but the method requires a large, rigorously decontaminated solution database.
最新模型
Meta has released the Muse Glimmer 30B weights, vision encoder, and DFlash draft model. The 17GB version can run text, image, and speculative decoding workloads simultaneously within 24GB of VRAM. Meta measured an average of 233.4 tokens/s on an RTX 5090, but safety evaluations and deployment details show that the model still requires external safeguards and a newer runtime.
AI 輔助科學研究
Using reasoning models, coding agents, and file-based memory across roughly 240 research sessions, a team ultimately proposed—and the authors verified—`K_G ≥ 6π/11`. The system demonstrated strong local proof capabilities but did not independently recognize that its search had reached a bottleneck; research direction and decisions about which results to pursue still depended on roughly 40 human instructions.
代理工具
ChemWorld uses typed operations, a compatibility compiler, and private chemical laws to let researchers replace a single world rule while retaining the same public interface. The open-source release passed 52 generated combinations and eight end-to-end workflows, but it validates software execution semantics—not real-world chemical accuracy.
AI 基礎設施
Rather than selecting inference nodes based solely on prefix hits, MISA-T also limits how many sessions from different workloads can occupy the KV cache concurrently. The authors measured rollout throughput gains of 43.6% to 53.3% on Step3.7 and Qwen3.6-35B-A3B, but the implementation has not yet been released.
最新模型
Rather than deciding shared experts, FFN width, and the number of activated experts separately, UniF-MoE lets each token first select shared channels and then accumulates routing probabilities according to its remaining requirements. In the authors’ DeiT and BERT experiments, it improves accuracy and MoE inference cost, but it has not yet been validated on generative large language models or distributed expert parallelism.
最新模型
ReRound uses no calibration corpus. Instead, it trains a conditional diffusion model on the model’s own weights to reconsider whether weights near the midpoint of a quantization interval should be rounded up or down. It preserves the existing low-bit format at deployment with no additional inference overhead, but each model requires several hours of model-specific reconstruction upfront.
最新模型
XPeng introduces XCoT-VLA, replacing dozens of natural-language reasoning tokens with a limited set of semantic action tokens, then using flow matching to generate continuous trajectories. Open-loop tests show lower lateral error, but no closed-loop safety evaluation has been reported, and the latency figures do not cover the full in-vehicle hardware and software stack.
應用研究／AI 基礎設施
Through heterogeneous graph attention, physical-variable decoding, and iterative residual correction, GENCO unifies power flow, optimal power flow, and state estimation. Batched inference on large grids is substantially faster than conventional AC solvers, but its zero-shot cross-grid and real-world SCADA accuracy remains insufficient to directly replace existing tools.