AI Briefing#long-context
The latest AI and developer news about #long-context, with the original source and a short summary.
The latest AI and developer news about #long-context, with the original source and a short summary.
Huawei has unveiled KVarN, a new KV-cache quantization backend that improves both capacity and throughput for vLLM at the same time.
KOLongDoc, a VLM benchmark for evaluating comprehension of long Korean public institution documents, has been released.
MiniMax has previewed its M3 model, which introduces a new sparse attention mechanism to achieve 15.6x faster long-context response speed.
To overcome the limitations of existing LLM judges, an agentic approach capable of verifying external systems — the **Agent Judge** — is needed.
RTPurbo, a technique that converts a Full Attention LLM into a highly efficient Sparse model with only a few hundred short training steps, has been unveiled.
The latest open-weight LLMs have adopted KV cache compression and attention optimization techniques for long-context efficiency
NVIDIA has released a 2.6B open-source world model for 1-minute 720p video.
Open-weight LLMs including Gemma 4 and DeepSeek V4 are lowering long-context costs through KV sharing and compressed attention.
Fixed-size online memory improves LLM long-term memory performance.
Explicitly stating source boundaries was needed for LLMs to distinguish text usable as grounds for an answer.
Ovis2.6-80B-A3B has been released as an MoE-based multimodal model.
Selective hierarchical attention was about 17x faster than standard attention at 512K context.
MSA has unveiled memory sparse attention that scales up to 100M tokens.
DeepSeek released previews of V4-Pro and V4-Flash, emphasizing price competitiveness.
Artificial Analysis's Long Context benchmark compared FOSS and SOTA models.
NVIDIA released an NVFP4 quantized checkpoint of the Gemma 4 26B model.
AutoSP is a DeepCompile-based compiler that automates Sequence Parallelism for long-context training.
The open-source coding agent Dirac scored 65.2% on Terminal-Bench-2 and 8/8.
DeepSeek cut V4-Pro pricing by 75% and also lowered API cache-hit costs to one-tenth of the previous rate.
Qwen3.6-27B was run up to a 100K context on an A5000 16GB using IQ4_XS.
DeepSeek has released the V4 model family, which supports a 1M token context.
A DeepSeek-V4 preview has been released, supporting a 1.6T MoE and 1M context.
DeepSeek-V4 unveiled a 1M-token context and agent optimizations.
Sessa is a decoder architecture for long context that puts attention into a recurrent feedback path.