Archive
新模型、開放權重、多模態、語音、影像與模型架構。
GitHub Repo
An October 8 reproduction on a Transformers development build found that the Gemma 4 response parser truncates tool names containing hyphens or periods, causing argument parsing to fail. A community fix adds regression tests but has not been merged; whether model-side settings will also be updated remains unclear.
GitHub Repo
A reproduction on the October 8 Transformers development build found that when packed-sequence boundaries are marked only with `position_ids`, Qwen3.5’s gated-DeltaNet layers may let a preceding document affect the next document’s scores. The report offers additional boundary parameters as a workaround, but the impact on released versions and an upstream fix remain unconfirmed.
最新模型
JetBrains released Mellum2.1 Thinking on October 8. It retains the 12B MoE architecture and focuses this update on reinforcement learning in real software environments. The weights are licensed under Apache 2.0, but agent evaluation results are self-reported, and quantized deployments require separate validation.
GitHub Repo
Ollama released v0.40.1 on October 7, switching Windows model manifests to regular files and adding an on-read repair mechanism for existing symlinks. This gives community reports of previously unusable model names an official release path, though recovery across different Windows environments remains to be verified.
GitHub Repo
LangChain released 1.4.4 on October 8, enabling the summarization step to retry with trimmed input when it exceeds the model’s context limit. The fix has been merged and published on PyPI, but recovery rates and summary retention still need validation against real workloads.
最新模型
On October 7, Perplexity announced two open-weight late-interaction models for retrieving text, images, and visual documents. The 0.6B and 9B versions share an embedding space, allowing different-sized encoders to be used for document indexing and online queries, though their actual costs and performance on Chinese remain to be validated.
GitHub Repo
A reproduction on the October 8 Transformers development build found that a uniform frame-sampling function may return more indices than requested, including a position beyond the available frames. The case can be checked without downloading a model, but its impact on stable releases and individual model processors remains unclear.
GitHub Repo
A development-branch reproduction on October 7 suggests Whisper timestamp decoding may mishandle the two-dimensional tensors returned by `generate`, merging batch results or failing at runtime. The report includes a minimal reproduction, but the impact on released versions and any upstream fix remain unconfirmed.
最新模型
Anthropic released Claude Haiku 5.5 on October 7, with adjustable reasoning effort and lower per-token prices for short prompts. AWS is offering the model as well, but long-prompt pricing, cross-region routing, and the cost of real-world tasks still need separate evaluation.
GitHub Repo
A community reproduction on October 6 found that importing a Gemma 4 12B GGUF into Ollama under a name without its size may trigger the wrong prompt template. A proposed fix includes a regression test but has not been merged, and its effect on generation quality remains unmeasured.
GitHub Repo
An October 7 Ollama community report describes long-prompt requests on Windows that may complete generation underneath yet return a JSON parsing error. A proposed fix aims to expose masked transport errors, but it has not been merged and does not address the root cause of the disconnect.
最新模型
On October 7, Liquid AI released d1-3B and the experimental d1-omni-600M, which return structured decisions without generating answers token by token. Local and Jetson deployment examples are available, but the low-latency figures depend on warm-up conditions, and evaluation results are still primarily based on the company’s own testing.