DeepSeek V4 Flash is at the center of attention again this week. This time it's not about the price list or a leak. It's about speed. On the afternoon of August 8, 2026, Ollama's official X account said that DeepSeek V4 Flash, served on its cloud, has passed 200 tps (tokens per second) in output speed. The post is short and plain: "DeepSeek V4 Flash on Ollama Cloud is now running at 200+ tokens per second output with zero data retention (ZDR). Have a great weekend." The tweet, which quoted an earlier announcement, drew about 128,000 views, 93 reposts, and about 1,600 likes, and spread quickly.
From "120 tps" to "200 tps" Within a Week of Launch: What Happened?
The earlier post this tweet quoted makes the context clearer. Ollama had previously announced: "DeepSeek-V4-Flash-0731 is now fully rolled out as the default deepseek-v4-flash model on Ollama Cloud. This model combines speed, efficiency, and frontier-level performance. Fast: 120+ tokens per second output on Ollama Cloud. Private: zero data retention hosting in the US and Europe." Just a few days later, the same account put out a new figure for the same model: 200+ tps instead of 120. But both numbers are announcements from Ollama's own account, and nothing in either post explains the hardware configuration, number of concurrent requests, or prompt length under which they were measured. Before asking "what optimization happened," the first thing to confirm is that this number is a vendor's own announcement.
Why Speed and Privacy Right Now? DeepSeek During a Price Hike
Studio Huuush has covered several DeepSeek stories over the past few days: DeepSeek signaling an API price increase, hitting No. 1 in usage on the coding agent Cline, and rumors of a leaked V4 Pro. This Ollama announcement reads best as an extension of those. With DeepSeek's own API handling trillions of tokens a day and a price hike on the way, third-party clouds like Ollama press the point that "you can use the same model faster and more safely." The emphasis on ZDR (zero data retention) fits the same picture. For companies and developers who are uneasy sending sensitive data directly to DeepSeek's API, which is run by a Chinese company, hosting in the US and Europe with no data retention is as strong a selling point as speed. Below, this announcement's figures are set next to the per-provider speeds for DeepSeek V4 Flash tracked by the third-party benchmarking firm Artificial Analysis.
| Source | When announced | Speed (tps) | Measured by |
|---|---|---|---|
| Ollama Cloud (0731 rollout announcement) | Early August 2026 | 120+ tps | Ollama's own announcement |
| Ollama Cloud (this tweet) | August 8, 2026 | 200+ tps | Ollama's own announcement |
| DeepSeek official API | Ongoing measurement | About 115–121 tps | Artificial Analysis (third party) |
| Other cloud providers, typical range | Ongoing measurement | About 83–150 tps (some in the 205 tps range) | Artificial Analysis (third party) |
200 tps Will Vary by Environment
In practical terms, the appeal of Ollama Cloud is clear. You can run a 284B-parameter-class model with one line, ollama run deepseek-v4-flash:cloud, without putting it on your own GPU, and it connects right away with CLI tools like Claude Code, OpenClaw, and Hermes. Billing is based on GPU usage time rather than tokens, so costs are relatively easy to predict at small scale. But a few things need flagging before you use this "200 tps" announcement as the basis for a real decision.
- Confirmed: The Ollama Cloud service and the DeepSeek V4 Flash model actually exist, and you can use them with external tools such as Claude Code.
- Confirmed: The X post exists, was posted on August 8, and is spreading with more than 128,000 views.
- Confirmed: Even by third-party benchmark measurements, DeepSeek V4 Flash runs at 83–150 tps depending on the provider, with some reaching the 200 tps range. In other words, a figure in the "200s" isn't unusual in the industry.
- Not confirmed: The measurement conditions behind Ollama's 200 tps (number of concurrent requests, prompt length, whether batching was used), and whether the figure is an average or a peak.
- Not confirmed: Whether this figure has been cross-checked by an independent benchmarking firm such as Artificial Analysis. So far, no third-party measurements other than Ollama's own announcement have been found.
In the end, this news is better read not as a confirmed fact that "DeepSeek V4 Flash got faster" but as a sign that competition in cloud serving, built around speed and privacy, has begun in earnest. If you're a developer who needs this speed for real work, the surest approach is a short test with your own workload (concurrent requests, prompt length) rather than relying on the vendor's number.