Tag

#deepseek

329 posts25 followers

Explore Hashnode

Alternatives

RONI DAS1d ago · 9 min read

DeepSeek open-sourced its inference engine. The three parts it left out are the real system design.

Near the top of the DeepInfer README there is a short list with a plain heading: "Required for a complete inference service (not included in this release)". It has three items. A KV cache system. "VE/

systemdesignacademy.hashnode.dev

Evan2d ago · 4 min read

Typed Plugins for DeepSeek Harness: A Tour of DSHX

DeepSeek Harness plugins are ordinary TypeScript: a Host side that registers tools and APIs, and a Client side that renders slots in the UI. The difficult part is not writing one. It is keeping both s

liyown.hashnode.dev

UplinkSwitch3d ago · 6 min read

DeepSeek API cost per request, October 2026: a chat turn and a long document

How a DeepSeek API request is billed DeepSeek bills API calls per token, at three rates per million tokens: input, cached input and output. The Flash models cost cents per million tokens and Pro costs

uplinkswitch.hashnode.dev

Yong Yu3d ago · 7 min read

DeepSeek V4.1-Flash-1017 Is Live Under the Same Name: First Call, Ignored temperature, and the Nov 1 Search Bill

Attributed compile + two small offline scriptsPrimary sources (DeepSeek, 2026-10-17): DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real (and the Chinese version); the Hugging Face model card

yongboyu.hashnode.dev

Neeraj Sharma4d ago · 5 min read

DeepSeek now charges 0.75 USD per 1,000 searches. On cheap models, tools cost as much as thinking.

DeepSeek released V4.1-Flash-1017 on 17 October: an MIT-licensed update to its Flash model with a 1M token context, list prices of 0.15 USD per million input tokens and 0.60 USD per million output tok

axionry.hashnode.dev

Shivam Kumar4d ago · 2 min read

DeepSeek V4.1-Flash: The Efficiency Update That Changes Agent Economics

The release October 17, 2026: DeepSeek shipped V4.1-Flash (snapshot 1017), a distilled update to its V4 line optimized for three production concerns — output brevity, reasoning density per token, and

shivaysinghrajput.hashnode.dev

LifeHack Nexus5d ago · 14 min read

The Local DeepSeek-R1 Reasoning Blueprint: Running 14B & 32B Models on Consumer GPUs for $0 (VRAM Math & Setup)

⚡ Benchmarking Summary & Local Hardware Telemetry: The release of DeepSeek-R1 has shattered the proprietary AI monopoly. By utilizing Large-Scale Reinforcement Learning (RL) with test-time compute, o

lifehacknexus.hashnode.dev

AnandOct 11 · 10 min read

An Attention Window Is Not a Checkpoint: Inside vLLM's 128-Token Replay

You ask a coding agent to fix a retry bug: Fix the failing retry test. Keep the public API unchanged. It reads the client code, opens the test, makes a small edit and runs it. The test still fails.

2 anandms101.hashnode.dev

RONI DASOct 11 · 10 min read

DeepSeek V4.1 Flash reads your prompt with half the model. I ran its prompt code and found one small tool adds 266 tokens to every request.

On 10 September 2026 DeepSeek, an AI company based in China, released DeepSeek V4.1 Flash. You can call it on DeepSeek's API (the web service where DeepSeek runs the model for you) or download the wei

systemdesignacademy.hashnode.dev

Neeraj SharmaOct 10 · 9 min read

DeepSeek V4.1 Flash agent loops: your cache-hit rate is the bill

The post behind this week's Hacker News thread, "Why isn't the industry freaking out about DeepSeek 4.1 Flash?", credits DeepSeek's smaller KV cache for keeping all-day sessions under a dollar. I want

axionry.hashnode.dev

Search Hashnode

Explore Hashnode

Trending tags this week

DeepSeek open-sourced its inference engine. The three parts it left out are the real system design.

Typed Plugins for DeepSeek Harness: A Tour of DSHX

DeepSeek API cost per request, October 2026: a chat turn and a long document

DeepSeek V4.1-Flash-1017 Is Live Under the Same Name: First Call, Ignored temperature, and the Nov 1 Search Bill

DeepSeek now charges 0.75 USD per 1,000 searches. On cheap models, tools cost as much as thinking.

DeepSeek V4.1-Flash: The Efficiency Update That Changes Agent Economics

The Local DeepSeek-R1 Reasoning Blueprint: Running 14B & 32B Models on Consumer GPUs for $0 (VRAM Math & Setup)

An Attention Window Is Not a Checkpoint: Inside vLLM's 128-Token Replay

DeepSeek V4.1 Flash reads your prompt with half the model. I ran its prompt code and found one small tool adds 266 tokens to every request.

DeepSeek V4.1 Flash agent loops: your cache-hit rate is the bill