DeepSeek quietly rewrote the cost curve of long-context inference, Cognition pushed coding agents onto a 2.8T-parameter base, and the week's two most-trending papers both say the same thing: the next gains won't come from bigger models but from sharper interfaces — semantic action spaces that let VLMs play robots, and interventions that let multi-agent crews optimize their own prompts.
Frontier & Text Models
DeepSeek's V4.1 Flash is a 552B-parameter multimodal MoE that activates just 8B–16B parameters per token — and squeezes its global KV cache to 890 bytes per token. Built on a Causal Encoder-Decoder architecture (20-layer encoder + 20-layer decoder) and pretrained from scratch on 45T multimodal tokens, V4.1 Flash extends context to 1M tokens through CSA2 (Compressed Sparse Attention 2), which assigns each attention layer a Full, Reindex, or Reuse mode and shares keys across layers via a Hierarchical Sparse Indexer. FP4 KV caching (one E4M3 scale per 16 channels) cuts the global cache to roughly a quarter that of DeepSeek-V4-Flash and about 437x smaller than DeepSeek-V1. The routing scheme pairs 1 shared expert with 384 routed experts per layer (6 activated per token), vision enters through DeepSeek-ViT plus a 2-layer MLP projector under joint text+image pretraining, and a controllable reasoning-effort dial runs 1–100. For input-heavy agentic workloads this is the model to watch.
Read more → https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
Cognition's SWE-2 is a 2.8T-parameter coding model built on a Kimi K3 base, tuned to push the Pareto frontier of agentic coding across multiple effort levels. Announced today as the company's most advanced coding model to date, SWE-2 deliberately refuses to bet on a single operating point, instead delivering highly competitive agentic coding performance at several effort tiers — a recognition that agent economics are about the whole curve, not the peak. The move up the parameter scale to 2.8T signals Cognition's answer to a market where coding agents are the hottest commodity and efficiency across budgets is the real differentiator.
Read more → https://www.cognition.ai/blog/swe-2
Robots & Embodied AI
Show-Harness lets vision-language models “play” robots through a compact semantic interface of discrete action units, grounding even closed-source frontier VLMs into zero-shot robot control. Instead of translating pixels straight to motors, embodiment-specific interpreters ground the semantics into local robot actions, keeping the VLM responsible for fine-grained physical decisions. That design unlocks two things at once: closed-source frontier VLMs become usable as zero-shot robot policies, and small open VLMs can be adapted for low-cost deployment with just a few GPU-hours of fine-tuning. A companion GUI Manipulation Interface (GUMI) collects demonstrations without specialized teleoperation hardware. Across tasks, embodiments, and environments, the harness outperforms representative agentic and VLA paradigms — evidence that the right interface can unlock embodied capability without more model capacity.
Read more → https://arxiv.org/abs/2609.10522
Papers & Research
AgentGrad fixes two weak stages of textual-gradient prompt optimization for multi-agent systems — extraction and aggregation — and cuts wall-clock optimization time by 2.5x. Where prior methods pick a target prompt without verifying it actually resolves the failure, AgentGrad uses sequential intervention to modify one agent at a time and pinpoint the exact agent whose change fixes the failure; the resulting output becomes the supervision for a fine-grained gradient. A semantic abstraction layer then clusters similar gradients so unrelated failure modes don't get concatenated into prompts that don't generalize. The result is state-of-the-art performance across five multi-agent benchmarks, in roughly half the time of the next-fastest baseline — a reminder that in the agent era, the prompt pipeline itself has become the bottleneck worth optimizing.
Read more → https://arxiv.org/abs/2609.08572