1. H Company Releases NeoMME Single-Tower Multimodal Encoders
H Company has released NeoMME, a family of 260M and 800M parameter bidirectional encoders that function as single-tower multimodal models. By processing multilingual text and raw 32x32 RGB image patches through a single Transformer, NeoMME eliminates the need for a separate vision tower or causal decoder. Despite its small size, the NeoMME-Retriever-260M matches the performance of the 3.75B ColQwen2.5 model on the ViDoRe v3 benchmark. The models are released under an Apache 2.0 license, support a 16,384-token context, and can compress index storage from 1.5 MB to just 6 kB per page.
- • NeoMME models are available in 260M and 800M parameter sizes under an Apache 2.0 license with day-zero Hugging Face support.
- • The single-tower architecture processes multilingual text and raw 32x32 RGB image patches through a single Transformer, dropping the separate vision tower.
- • NeoMME-Retriever-260M matches the performance of the 3.75B-parameter ColQwen2.5 model, scoring 0.523 nDCG@10 on the ViDoRe v3 benchmark.
- • On an NVIDIA L40S, the 260M model indexes 51.3 pages per second, while query encoding takes 78.3 ms on a CPU-only host.
- • Index storage requirements can be reduced from 1.5 MB to 6 kB per page using hierarchical token pooling and asymmetric quantization.
Developers can build highly efficient multimodal search and RAG pipelines using compact models that run fast on CPUs and require minimal index storage.
2. XHToken Releases Spark-X2.5 Models with 1M Context Window
XHToken has released Spark-X2.5-4B and Spark-X2.5-1.7B, two compact language models designed for long-context tasks. Despite their small parameter sizes, both models support over 200 languages and feature a native context window of up to 1 million tokens, enabled by a hybrid attention architecture combining one full-attention layer with three sliding-window attention layers. The models are built for agentic workflows and offer day-one integration with frameworks like Claude Code and Codex, as well as inference engines including vLLM, SGLang, llama.cpp, and Ollama.
- • The Spark-X2.5-4B and Spark-X2.5-1.7B models support over 200 languages and feature a native context window of up to 1 million tokens.
- • The architecture uses a hybrid attention design consisting of one full-attention layer and three sliding-window attention layers.
- • The models are integrated with agent harnesses including Codex, Claude Code, OpenClaw, and Hermes.
- • Supported inference frameworks include vLLM, SGLang, llama.cpp, MLX, Ollama, and LM Studio.
- • The models are compatible with hardware platforms from NVIDIA, Huawei, Hygon, and HOUMO.AI.
Developers can run highly compact, long-context models locally with native support for popular agent frameworks and inference engines.
3. UC Berkeley Releases CUA-Lite for Computer-Use Agents
Researchers from UC Berkeley have released CUA-Lite, an open platform designed to simplify the development and evaluation of computer-use agents (CUAs). The platform unifies agents, environments, traces, and evaluation frameworks under a single action space and data schema. A key component is Lite.OSWorld, which replaces heavy virtual machines with Docker containers, shrinking the environment footprint from 4.1 GB to 0.9 GB while maintaining identical performance scores across 13 tested models. CUA-Lite supports Python 3.12 and can be deployed on any Docker host without nested virtualization.
- • CUA-Lite provides a single action space and data schema across desktop, browser, and mobile environments.
- • The platform includes Lite.OSWorld, which runs OSWorld tasks in Docker containers instead of virtual machines, reducing the footprint from 4.1 GB to 0.9 GB.
- • Performance scores in the Lite.OSWorld container match the original OSWorld VM across 13 tested models.
- • The platform features over 30,000 verifiable tasks, 15+ benchmarks, 10+ agents, and 20+ datasets on Hugging Face.
- • CUA-Lite is compatible with Python 3.12 and deploys on any Docker host without requiring nested virtualization.
Developers can build and test computer-use agents in lightweight, Docker-based environments that reduce disk footprint by over 75% without sacrificing evaluation accuracy.
4. Open-Source 'lm-eval-ledger' Harness Compares Model Answers Visually
A developer has released `lm-eval-ledger`, an open-source benchmark harness designed to help developers inspect and compare how different models answer specific benchmark questions. The tool is YAML-driven, allowing users to configure multiple models and tasks to run with a single command. It stores detailed per-question data—including system prompts, raw generations, extracted answers, and stop reasons—in a SQLite database, which can then be explored and compared side-by-side using an included Flask web interface. The tool supports vLLM, SGLang, Hugging Face, and llama.cpp backends.
- • The tool is YAML-driven, allowing users to define multiple models and tasks in a configuration file to run benchmarks with a single command.
- • It stores benchmark data in a SQLite database and uses a Flask application to visualize and compare model answers.
- • The harness records per-question data including system prompts, model generations, extracted answers, ground truth, and stop reasons.
- • It supports backends including vLLM, SGLang, Hugging Face, and llama.cpp on both Linux and Windows.
- • Initial benchmarks run on a single 5090 GPU compared Qwen3.5-9B, NVIDIA-Nemotron-3.5-Lightning, and Gemma-4-12B-it.
Developers can visually inspect and compare exact model outputs across different benchmarks to understand why models succeed or fail beyond raw scores.
5. Open-Source 'cache-pressure' Tool Validates Local LLM KV Cache Eviction
A new open-source tool called `cache-pressure` has been released to help developers validate and measure KV cache management performance on local LLM deployments. The tool works by calibrating cache expectations, filling the cache with stable contexts, and validating cache hits in reverse order to see exactly how old contexts get evicted. Using the tool, its creator successfully identified and resolved cache management bugs in vLLM, boosting retained tokens under pressure from roughly 1 million to over 3 million. The tool is compatible with vLLM, ninfer, llama.cpp, and SGLang.
- • The tool functions by calibrating cache expectations, filling the cache with stable contexts, and validating cache hits in reverse order.
- • Using the tool, the author identified and fixed cache management issues in vLLM, increasing retained tokens under pressure from 1,052,025 to 3,000,048.
- • The tool is compatible with multiple inference engines, including vLLM, ninfer, llama.cpp, and SGLang.
- • Running the tool will evict all existing cached context on the target deployment.
Developers can identify and resolve silent context eviction bugs in local inference engines to ensure reliable long-context performance under heavy load.
6. Evaluation Compares Eight Abliterated Qwen 3.8 27B Variants
A detailed study evaluating eight abliterated (uncensored) variants of the Qwen 3.8 27B model highlights the trade-offs of different uncensoring techniques. Conducted over 11 days using 167 GPU hours, the study found that surgical edits generally outperform heavy edits, which often cause models to fail to terminate their thinking loops (as seen in the 'obliteratus' variant, where 44.8% of responses failed to complete). The 'orcarouter' variant achieved the highest HarmBench attack success rate at 82.2%, while the 'apostate' variant offered a strong balance of low KL divergence and near-identity capabilities.
- • The evaluation compared eight abliterated variants of Qwen 3.8 27B over 11 days, utilizing approximately 167 GPU hours.
- • The methodology included weight comparisons, KL divergence measurements, 13 benchmarks, and the HarmBench 400 classic.
- • The 'orcarouter' variant achieved the highest HarmBench attack success rate (ASR) at 82.2% and was the only model with fully verified weights.
- • The 'apostate' variant achieved a 78.7% ASR and is noted for its low KL divergence and near-identity capabilities.
- • The 'obliteratus' variant should be avoided, as aggressive editing caused 44.8% of responses to fail to complete their thinking process.
- • The 'blackfrost' variant was found to contain an undisclosed 1457-character jailbreak system prompt within its chat template.
Developers deploying uncensored local models can select variants that maintain reasoning capabilities and avoid broken thinking loops caused by over-editing.