Archive
GitHub 熱門專案、開源工具、框架、runtime 與重要 release。
代理工具
The UK AI Security Institute has open-sourced optstop, which uses posterior uncertainty at the item and task levels to decide when to stop repeated evaluations. Across nine shadow tests, it saved an average of 81.1% of trials, though performance varied with score type, item ordering, and the exchangeability assumption.
推論系統
FreeBalance uses the previous layer’s residual state to predict expert load in the next layer, overlapping weight transfers with attention computation. The method does not alter the final routing decisions or model outputs, but the evidence so far covers only two MoE models, a single eight-GPU topology, and prefill workloads.
代理工具
TWIN requires an agent to first encode an unknown environment as an executable Python world model. The agent may take real actions only after the model can replay its entire history exactly. Using the same base model, this raises the ARC-AGI-3 score from 7.8 with direct interaction to 93.3.
推論系統
The new release fixes two compatibility gaps in the model preprocessing layer: WebP images are now transcoded, and the Qwen renderer no longer requires system messages to appear first. Although the changes do not introduce new models or kernels, they can prevent multimodal and agent conversations from failing before they reach the inference core.
代理工具
The new release allows a `UIEventStream` to be created before `run_input` is available and lets `AGUIEventStream` hold its own thread and run IDs. It also fixes Temporal sandbox compatibility and tracing attribution for fallback models, reducing diagnostic blind spots in long-running agent workflows.
最新模型
Rather than comparing only local differences between adjacent denoising steps, GCache estimates the cumulative impact of cache errors on the final image or video. It uses dynamic programming to select cache refresh points, then applies Bayesian optimization to align theoretical weights with actual generation quality.
代理工具
SkillShapley treats the semantic steps in `skill.md` as players in a cooperative game and uses Shapley values to measure each step’s contribution to task success. Its BAES approximator prioritizes skill configurations that can yield multiple single-step comparisons, reducing the cost of running agents individually.
AI 安全
The new release can use Linux memory cgroups to cap memory for builds and tests run by the agent, preventing runaway processes from crashing the entire session. It also fixes a Windows special-path validation bypass, an MCP reconnection loop, and a sandbox issue that consumed an entire CPU core while idle.
代理工具
OmniScientist carries perception through hypothesis generation, experimentation, and writing instead of first compressing images, waveforms, or trajectories into text summaries. In blinded comparisons, the full system achieved an 85% win rate against a version that received only precomputed features, though evaluations still relied primarily on model judges.
LLMOps
The new release can copy a small sample of real requests to a candidate router, then use a blinded judge to compare responses without affecting production results. Complexity classification also no longer treats a fixed agent system prompt as a signal of each request’s technical difficulty.
推論系統
The new release adds support for the Kimi K3 text model, hybrid KDA/MLA attention, latent MoE, and tool-call parsing. The converter directly reorders the MXFP4 bit layout, eliminating the need to first expand the entire 2.8T-parameter model to BF16.
最新模型
The previous LoRA loading path did not verify whether a tensor’s offset plus its size exceeded the GGUF file bounds, allowing corrupted or incompletely downloaded adapters to appear to load successfully. b10451 now fails immediately and confirms that the tensor index actually exists.