• Higgsfield Blog (rss) · 4 gundur sessiz
• HuggingFace Blog (rss) · 4 gundur sessiz

DeepSeek-V4.1-Flash doubles model size while cutting KV cache memory 4x through a redesigned encoder-decoder and sparse attention architecture.

• DeepSeek-V4.1-Flash is a 552B MoE model (8B active params for input, 16B for output) using a Causal Encoder-Decoder split (20 encoder + 20 decoder layers) so prompt tokens skip the full backbone.
• KV cache footprint dropped from 3,514 bytes/token (V4-Flash) to 890 bytes/token, cutting HBM needs to 1/4 and SSD cache storage to 1/8 of the predecessor.
• Sliding-Window Attention with 'Bounded Replay' lets the model cheaply reconstruct local context after a session pause, avoiding costly full recomputation.
• New Compressed Sparse Attention 2 (CSA2) lets attention layers share KV state and retrieval results via Full/Reindex/Reuse layer modes, cutting redundant memory across layers.
• A Hierarchical Sparse Indexer bounds search cost for 1M-token contexts by narrowing candidates from 2,048 blocks down to 512 positions, regardless of total context length.
• On Artificial Analysis Intelligence Index, V4.1-Flash scores 40 (vs Gemini 3.8 Flash High's 41) at roughly 1/4 the cost per task ($0.30/M input, $1.20/M output tokens).

Generating running routes with GPT-6 Astra and ChatGPT Work

• <p>Here's a neat thing I had <a href="https://simonwillison.net/2026/Aug/30/understanding-chatgpt-work/">ChatGPT Work</a> with GPT-6 Astra (Max) do this morning:</p> <blockquote> <p><code>I live at &lt;my address&gt;. Figure out 5K and 10K running routes from me that loop from my house. Use OSM data

shot-scraper 1.12

• <p><strong>Release:</strong> <a href="https://github.com/simonw/shot-scraper/releases/tag/1.12">shot-scraper 1.12</a></p> <p>I've added WebP support to my <a href="https://shot-scraper.datasette.io/">shot-scraper</a> screenshot automation tool. You can now take a WebP screenshot of a web page like t