DeepSeek's Post-Outage Surprise: Open-Weight V4.1 Flash Hits 300 Tokens/s, Rivals GPT-5.6 Sol

Outage, Compensation, and a Sudden Model Drop

DeepSeek recently suffered a service outage, apologized to users, and issued compensation. Two days later the company quietly released a new model, DeepSeek-V4.1-Flash-1017, with its weights open-sourced. The release upgrades coding, reasoning, and Agent capabilities across the board.

Toolchain Upgrade and Ecosystem Integration

Alongside the model, DeepSeek Harness was upgraded to version 0.2.1-rc.1 (or later) to fully exploit the new capabilities. The model has been specifically optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, promising a more complete experience after upgrading.

Community Reaction

"I knew it — that downtime was you secretly leveling up. Too ambitious."

"Progress plus a refund? Can't handle this honesty."

Comments like these flooded the official post, with thousands of likes, reflecting a mix of amusement and appreciation for the rapid turnaround.

Benchmark Results: Speed and Intelligence

Overseas users measured 244 tokens per second , with some runs approaching 300 tokens/s .

This throughput exceeds GPT 6.1 Sol's Ultrafast mode .

Intelligence level is reported to be on par with GPT 5.6 Sol .

One user remarked: "Don't understand how they deliver this quality and speed at such a low price."

Real-World Developer Tests

Video generation tests show results very close to Opus 5.5 , especially in 3D production and frontend design .

A Blender + Three.js scene test led a developer to call it "the first open-weight model that feels this close to Opus" , with generated video quality described as stunning.

Training Methodology: Self-Feedback Reinforcement Learning

The improvement is not merely a benchmark chase. The model employs self-feedback reinforcement learning , which strengthens the ability to build complete projects from scratch. In end-to-end scenarios like game development and 3D modeling, the model requires fewer interaction rounds, produces more complete results, emits shorter outputs, and runs faster — directly lowering the comprehensive cost per task. For users billed by token, this is more practical than leaderboard chasing.

Search and Fetch APIs Opened to All

DeepSeek's self-developed Search and Fetch interfaces are now publicly available.

Backed by an in-house search engine serving hundreds of millions of requests daily .

Pricing: $0.75 per 1,000 calls per interface.

When invoked via DeepSeek Harness, Claude Code, Codex, or other server-side tooling, the APIs are free for DeepSeek API users until October 31 ; billing starts after that date.

Strategic Narrative: A Coordinated Combo

From outage to compensation to a new open-weight model plus toolchain adaptation in two days, the sequence resembles a deliberate combination punch. Observers note this demonstrates long-termism: acknowledging problems openly, restoring trust with concrete actions (compensation, new model, open source, tool integration), and immediately pushing the product forward rather than just talking about user experience.