All Articles · SURL AI News

Transformers Community Reports Tokenizer Configuration Mismatch That May Silently Drop Chinese Input in DeepSeek Models

A reproduction on Transformers 5.8.1 suggests that certain DeepSeek models may load the wrong tokenization pipeline, causing Chinese input to disappear and English spaces to be lost. Current development-branch documentation describes a fallback for configuration mismatches, but whether this case is fixed across versions remains unconfirmed.

2026-10-04

GitHub Repo

Ollama Community Reports Missing Text in LFM2 Output; Certain Python Strings May Disappear Silently

A community reproduction on Ollama 0.35.1 suggests that `lfm2:24b` may omit certain Python strings, altering JSON values and document content. Upstream has yet to confirm the issue, and there is not enough evidence to determine whether it stems from tokenizer decoding, model packaging, or the hardware path.

2026-10-04

Ai2 Releases AstaBrief 8B, Simplifying Citation-Backed Scientific Reporting in a Single Generation

Built on Qwen3-8B, AstaBrief releases model weights and training data, combining research questions and retrieved passages into reports. Ai2 reports that the end-to-end workflow is about 3.5× faster, though the comparison mainly uses models and evaluations from 2025.

2026-10-04

Aleph Alpha Releases Kolibri-1, a 78B MoE Model with Hybrid Attention for Million-Token Context

Kolibri-1 is released with open weights under Apache 2.0. It activates about 3.46B parameters per token and targets German-English bilingual use, document processing, and tool calling. Its million-token window extends beyond the model’s native 256k training length; the company still recommends shorter contexts for complex tasks.

2026-10-04

GitHub Repo

Ollama Community Reports Qwen GGUF Thinking Levels Not Taking Effect; Low-Intensity Settings May Still Use the Default

A community report involving Ollama 0.35.0 says a specific Qwen3.8 GGUF may retain the template’s default `xhigh` setting after receiving a lower-intensity thinking request. Official documentation says to choose a supported value based on model metadata; whether template capabilities and API controls align remains unconfirmed.

2026-10-04

GitHub Repo

vLLM Community Reports Silent Loss of Image Input; Per-Request Thinking Settings May Affect Visual Understanding

A community report involving vLLM 0.30.0 suggests that adding `chat_template_kwargs` to image requests may leave the model with text only, while the server still returns success. The issue has yet to be confirmed upstream; deployments combining visual input with thinking-mode controls should verify the actual input.

2026-10-04

GitHub Repo

Hermes Community Reports Desktop and Docker Version Mismatch, Update May Prevent Remote Conversations

After a Hermes Desktop update, the client may send RPC fields that an older Docker backend rejects, causing new conversations to fail. Official documentation confirms the two use different update mechanisms; the report highlights the need for protocol compatibility checks in agent deployments.

2026-10-03

GitHub Repo

Dify Community Reports MCP Resource Links Being Dropped, Leaving Empty Results Despite Successful Tool Responses

Dify’s MCP integration has been reported to mishandle resource_link content, potentially leaving tools that return only resource links with blank observations. A community patch to preserve links and metadata has been proposed but not merged; the impact on released versions remains unconfirmed.

2026-10-03

GitHub Repo

vLLM Community Reports CPU Sampling Bias; Fixed Random Table May Make Some Tokens Unselectable

A reproduction test with a large vocabulary suggests that the vLLM V1 CPU sampler may deviate from the model’s original probability distribution. The issue involves a fixed Gumbel random table; the impact on real models and an official fix remain to be confirmed.

2026-10-03

GitHub Repo

vLLM Community Reports Ignored Per-Layer LoRA Scaling, Potentially Degrading Fine-Tuning Results in Deployment

A report says vLLM 0.30.0 does not apply PEFT’s per-layer rank and alpha settings, which could silently change adapter behavior. The community has provided a reproduction and a weight-conversion comparison; an official fix and the scope of impact remain unconfirmed.

2026-10-03

GitHub Repo

Ai2 Open-Sources Olmo-core 3, Redesigning Expert Placement and Data Routing for MoE Training

The new version keeps experts resident on GPUs and uses data routing to reduce the cost of repeatedly aggregating weights. Initial tests reported about 2.7× throughput. A trillion-parameter configuration has passed systems testing, but it used random routing, so training quality and stability remain unverified.

2026-10-03

Cloudflare Open-Sources Clef Decision Models, Producing Option Probabilities in a Single Forward Pass

Clef and Clef-flash turn agent classification and routing into direct scoring of specified options, and release their weights under Apache 2.0. Official tests show latency advantages, but task accuracy and probability calibration still need to be validated against deployment data.

2026-10-03