推論系統

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

Muse Spark 1.3 Reduces Agent Tool Calls, While Its Highest Reasoning Mode Still Awaits Safety Testing

Meta has updated Muse Spark with a focus on long-horizon agents, software development, and task tracking in mixed-topic conversations. Internal comparisons show reductions in both tool calls and token usage. The model is now available in Muse Code and via the API, but its highest reasoning mode, open weights, and fully reproducible evaluations have yet to arrive.

2026-09-03

最新模型

Gemini 3.8 Flash Shifts Toward High-Intensity Agentic Reasoning, While Cyber Version Offers Vulnerability Patching Through a Vetted-Access Program

Google has made Gemini 3.8 Flash generally available for production, offering a one-million-token context window, adjustable reasoning intensity, and new API migration requirements. Built on the same core model, Flash Cyber focuses on vulnerability discovery and automated patching but is available only to vetted defenders through the Fairwind program.

2026-09-02

最新模型

RWKV7-G1j Scales an Attention-Free Recurrent Model to 13.3B While Maintaining Fixed State During Long-Sequence Decoding

The RWKV team has released the 13.3B-parameter G1j base model, along with Transformers, LoRA, and TileLang inference support. Its fixed-size recurrent state avoids a KV cache that grows with context length, but the model has no published safety or capability evaluations and should not be treated as a production-ready chat assistant.

2026-09-02

最新模型

Qwen3.8-Max-0902 Strengthens Long-Horizon Coding Agents, but Its Lead Remains Within the Margin of Error

Alibaba Cloud has updated its hosted Qwen3.8-Max snapshot, focusing on production-grade software development, multi-tool collaboration, and visual understanding while retaining a 1-million-token context window. It currently ranks first on Code Arena, but the smaller sample and overlapping ranking intervals are not yet sufficient to demonstrate that it consistently outperforms other flagship models.

2026-09-02

最新模型

BenchMIRT Uses Multidimensional Item Response Theory to Dissect LLM Evaluations, Revealing That Safety Scores Can Conflate Reasoning Ability

Ai2 has open-sourced BenchMIRT, which separates latent capabilities such as safety and general reasoning based on model response patterns across more than 34,000 questions. Experiments show that some safety benchmarks primarily reflect reasoning performance and that a carefully selected 10% of questions can largely preserve capability rankings, though the data covers only models released by early 2025.

2026-09-02

最新模型

Gemini Turns Long-Video Understanding Into a Tool Loop, Dynamically Selecting Clips and Sampling Rates Based on the Query

Google has added agentic video processing to three Gemini Flash models, allowing them to load frames, audio, or transcripts on demand instead of processing an entire video at a fixed frame rate in advance. The company claims token savings of up to 88% and cost reductions of 66%, though time to first token for short videos and evaluation reproducibility remain to be tested.

2026-09-02

推論系統

llama.cpp b10731 Adds Recurrent State Rollback for Qwen3.8-Flash-Next, Eliminating Repeated Full-State Transfers During MTP Speculative Decoding

llama.cpp fixes an issue in Qwen3.8-Flash-Next that prevented SSM and convolutional states from being rolled back correctly when draft tokens were rejected. Testing on a single RTX PRO 6000 showed code decoding performance rising from 123 to 183 tokens/s, though the results remain limited to a specific quantization, draft head, and single processing slot.

2026-09-02

推論系統

LMDeploy 0.17 Introduces DeepEPv2 and Mooncake KV Storage, Expanding Cross-Node MoE Inference Data Paths

LMDeploy 0.17.0 adds the DeepEPv2 expert-parallel backend to TurboMind and connects its KV connector to Mooncake Store. The release also supports Kimi K2.6, non-power-of-two page sizes, and structured responses, but provides no end-to-end performance data across hardware configurations.

2026-09-02

最新模型

Claude Fable 5.1 Cuts Cache Costs for Long-Running Agents, but Upgrading Involves Breaking Changes to Thinking Blocks and Tool Calls

Anthropic has launched Claude Fable 5.1 with a 1 million-token context window, 75% lower cache-read pricing, per-message reasoning effort, and progress updates between tool calls. Existing Fable 5 workflows cannot simply swap in the new model: forced tool use, thinking-block reuse, and safety routing all require renewed testing.

2026-09-02

代理工具

Codex CLI 0.152 Adds Per-Tool MCP Output Quotas and Fixes Lost Instructions in Long-Running Sessions

OpenAI has added output token limits for individual MCP tools and now allows app-server shell commands to run for more than an hour. The release also fixes an issue where user instructions and authorizations could be lost after conversation history was compacted, while the built-in planning tool is now disabled by default.

2026-09-01

最新模型

Flower Endeavor 1.0 Opens Up Private Enterprise Deployment, but Model Specifications and Independent Evaluations Are Still Missing

Flower Labs has introduced Endeavor 1.0 for reasoning, software development, and long-running agent tasks, available either as a managed service or deployed in customer environments. Its initial results are close to those of some frontier models, but with access currently limited to a preview and the benchmarks based solely on vendor self-evaluation, real-world deployment costs and capabilities cannot yet be verified.

2026-09-01

代理工具

VS Code Agent Host Moves the Agent Loop Out of Extensions, Using AHP to Synchronize Sessions Across Windows

VS Code 1.132 through 1.135 connect agents such as Copilot and Claude to a standalone Agent Host, allowing sessions to continue across windows, remote connections, and headless environments. The open Agent Host Protocol unifies session state, but not the reasoning, permissions, or tool behavior of individual models.

2026-09-01