最新模型

新模型、開放權重、多模態、語音、影像與模型架構。

Ollama Pre-Release Reuses Verified Models, Cutting Warm Embedding Request Latency by More Than 40%

On October 12, Ollama merged a model lookup optimization and included it in v0.40.4-rc0, reducing repeated processing for subsequent requests to the same model. Maintainer tests showed warm-request latency reductions of 41% to 56% for two embedding models, though the results still need validation in real deployments.

2026-10-08

GitHub Repo

Transformers Community Reports Mamba2 Grouping Gap in Normalization; Non-Fused Path May Differ from Training Core Computation

An October 12 report to the Transformers community says non-fused Mamba2 RMSNorm does not normalize by configured groups, potentially producing different outputs and gradients from the fused training path. The report includes CPU and CUDA comparisons and a local patch check, but no official fix or full-model quality evaluation.

2026-10-12

GitHub Repo

Ollama prerelease fixes MLX rollback state, reducing memory retention after tool calls

Ollama merged an MLX memory fix on October 12 and included it in v0.40.4-rc0. The change addresses large buffers retained after speculative decoding rollbacks. In a 12-tool-call comparison, maintainers observed 3.52 GiB less memory held in the final round, though results across models remain unverified.

2026-10-06

GitHub Repo

Ollama Community Proposes Tool Argument Precision Fix After Go Client May Silently Rewrite Large Integers

An October 11 report from the Ollama community says the Go client may round large integers and high-precision decimals when decoding tool calls, then send the altered values back in conversation history. The proposed fix adds streaming and desktop tool tests but has not been merged; the numeric type change also requires compatibility handling.

2026-09-25

GitHub Repo

Transformers Community Reports HunYuan-VL Batch Generation Bug That May Pair Candidate Sequences with the Wrong Images

A Transformers development report on October 11 said HunYuan-VL may mishandle image feature expansion when generating multiple candidate answers. The report includes a minimal reproduction, and another participant said they could reproduce it; the impact on production models and a fix remain unconfirmed.

2026-10-12

GitHub Repo

Dify Community Proposes MCP Pagination Fix After Tool Discovery May Silently Miss Tools Beyond the First Page

A report from the Dify community on October 10 found that the MCP client does not track pagination cursors, so it may retrieve only the server’s first page of tools. The proposed fix includes partial regression tests but has not been merged; the full deployment and verification process remains to be confirmed.

2026-10-12

GitHub Repo

Dify Community Proposes Workflow Fix After Snippet Node IDs May Send Document Variables to Models Unresolved

A Dify community report on October 10 said nodes inserted through Snippet may create unparseable variable references, causing models to receive placeholders instead of document content. The proposed fix includes regression tests but has not been merged and will not migrate existing workflows.

2026-10-11

GitHub Repo

Transformers Community Investigates Mamba2/Zamba2 Cached Decoding Discrepancy; Time-Step Limit May Not Be Applied

A Transformers community report on October 10 found that cached decoding in Mamba2/Zamba2 may omit the time-step limit; another user reported reproducing a logits discrepancy. A fix has been proposed, but its impact on released versions, validation with accelerated kernels, and effects on generation quality remain unconfirmed.

2026-10-10

GitHub Repo

Ollama community investigates rnj-1 loading failure; sliding-window attention metadata may trigger assertion

An October 10 report to Ollama said rnj-1 may abort during loading in newer runtimes, returning API error 500. The following day, a contributor traced the cause to an attention-check change in llama.cpp and offered an alternative GGUF, but an official fix and the full scope of impact remain unconfirmed.

2026-10-11

GitHub Repo

Dify Community Reports a Gap in Agent History Replay Caching; Random Tool Call IDs Could Raise Inference Costs

A Dify community report on October 9 said that rebuilding agent conversation history generates new tool call IDs, potentially reducing prompt-cache reuse by model services. The report proposed using stable IDs, but the actual cost impact and status of any official fix remain unconfirmed.

2026-10-11

GitHub Repo

Transformers Community Proposes Sequence Isolation Fix After Missing Acceleration Packages May Change Model Outputs

A Transformers development-branch report on October 9 found that some PyTorch fallback operators ignore document boundaries, allowing packed sequences to affect one another. The community has proposed a fix and comparison tests, but the proposal has not been merged, and the impact on released versions remains to be confirmed.

2026-10-09

GitHub Repo

Langflow Community Reports Prompt Caching Gap: Second-Level Timestamps May Block Reuse of Multi-Turn Conversation History

A Langflow community report on October 9 says a date field in the Agent’s default prompt injects a second-level timestamp on every run, which may prevent later turns from hitting the prompt cache. The report suggests using a date instead, but public performance measurements and verification of an official fix are still lacking.

2026-10-10