Skip to content
Drix10 Blog

DeepSeek V4.1 Flash 1017 Abliterated – 1K+ tps, low token cost

, 8 items in AI Developer Tools, 4 min read

In this digest (8 items)

DeepSeek V4.1 Flash 1017 Abliterated is offered for authorized security testing with over 1,000 tokens per second and token pricing of $0.24 input, $0.012 cached, $0.96 output. The model is abliterated to refuse fewer security-testing requests and retains zero data from any request.

Key points

  • 1K+ tokens per second inference speed.

  • Pricing: $0.24 per input token, $0.012 per cached token, $0.96 per output token.

Sources

Hermes on Upstash Box – always-on customer-success assistant on $8/mo box

Hermes, the open-source agent from Nous Research, now runs continuously on an Upstash Box costing $8 per month. It watches customer signals, prepares meeting briefs, and triages inbox tasks while staying online when a laptop sleeps.

Key points

  • Runs on an always-on Upstash Box for $8/mo, keeping scheduled jobs active even when the laptop is asleep.

  • Built about 30 custom skills over several months, covering meeting prep, inbox triage, churn analysis, and upgrade outreach.

Sources

DeepSeek Harness Creator Mode announced

DeepSeek Harness introduces Creator Mode, letting users turn ideas into plugins for design workspaces, sandbox games, or 3D LLM visualizers. The post invites creators to build with the new mode.

Key points

  • Creator Mode enables turning ideas into plugins.

  • Plugins can be used in design workspaces, sandbox games, or a 3D LLM visualizer.

Sources

SIE Serverless launch – managed inference for small models

SIE Serverless is a new managed inference product for small models, announced today. It targets spiky workloads on 20+ task-specific workloads across Search, OCR and Agentic execution and offers a $500 launch credit grant.

Key points

  • Launch: SIE Serverless, a managed inference product focused on small models.

  • Scope: supports spiky workloads on 20+ task-specific workloads across Search, OCR and Agentic execution.

  • Offer: $500 credits launch grant available for sign-up.

Sources

CreaoAI Video Studio enables 8-second TikTok video replication

CreaoAI Video Studio claims it can replicate an 8-second TikTok video that received over 10,000 likes, offering higher video quality and smoother movements. The post provides open-source white-model dance videos and prompts for immediate use.

Key points

  • Feature: Replicate an 8-second TikTok video with higher quality and smoother movements.

  • Resource: Open-source white-model dance videos and prompts are available for download.

Sources

Tencent Hy4 added to Chutes API with new model ID

Tencent Hy4 is coming to Chutes. It will be available through the existing Chutes API using the same key and a new model ID.

Key points

  • Partnership with Tencent Cloud to bring Hy4 to Chutes.

  • Use the same API key, but a new model ID for Hy4.

Sources

OpenDesign Creator Mode open-sourced in OpenDesign

OpenDesign open-sourced a DeepSeek Harness-style Creator Mode. It lets users turn ideas into plugins directly in the platform.

Key points

  • Open-sourced DeepSeek Harness–style Creator Mode in OpenDesign.

  • Create plugins with a prompt, import from GitHub, or upload locally.

Sources