The flash model just 5x'd agentic throughput. deepseek-v4.1-flash on vllm, out oct 7, uses decoder…
the flash model just 5x'd agentic throughput. deepseek-v4.1-flash on vllm, out oct 7, uses decoder replay with cuda graphs to cut prefill compute by 30 to 40 percent.
the small model now outruns the queue.
Context
Verified event date: the vLLM Blog post 'DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0' is dated Oct 7, 2026 (Inferact and the vLLM Team). Its summary says that in the three weeks after the model's release, Inferact and the vLLM team worked on agentic throughput.
A vLLM pull request, 'Compact large decoder prefills with a dependency-preserving halo', says DeepSeek V4.1 Flash runs the full prompt through decoder layers 21 to 39 even though those layers reuse the encoder's global KV and have only a 128-token causal window. The paper on the model (arXiv 2609.19969) describes a 552B-parameter multimodal mixture-of-experts with 16B active parameters and up to one million tokens of context.
The Oct 7 date and the 5x agentic throughput match the vLLM blog title. The 30 to 40 percent prefill compute cut, the decoder replay and the CUDA graph detail were not seen in the sources read, so unsupported here, not refuted. The sources read do describe a decoder prefill compaction change in a pull request, which is the nearest match. 'The small model now outruns the queue' is the author's opinion, and a 552B backbone is not small by parameter count, though 16B parameters are active per token.
Related work
- vLLM Recipes: DeepSeek-V4.1-Flash ↗Deployment recipe; describes the 20-layer encoder and 20-layer decoder split.
- Lambda: DeepSeek-V4.1-Flash inference ↗Throughput table on 4x B200 with vLLM TP4.
- Nous Research raises $90M ↗Oct 7, 2026, another open model lab story from the same day.
Watch next
- Find the 30 to 40 percent prefill figure and the decoder replay detail in the vLLM blog.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 11 October 2026 at 15:55 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →