Now Live in Pickaxe: GLM 5.2, DeepSeek V4 Flash, and Kimi K2.5

[chrisfrompickaxe]

Three new open models just joined the Pickaxe lineup: GLM 5.2, DeepSeek V4 Flash, and Kimi K2.5. They’re live in the model picker now, on every plan.

• GLM 5.2: a 1M-context open model built for complex coding, tool use, and long-running agent workflows, with strong capability at substantially lower cost than many flagship models
• DeepSeek V4 Flash: a fast, 1M-context open model built for reasoning, tool use, and everyday AI workflows, offering strong performance at exceptionally low cost
• Kimi K2.5: a 1-trillion-parameter multimodal open model built for general reasoning, visual tasks, coding, and tool-driven workflows at a competitive cost

Open models bring serious capability at a fraction of flagship cost, which makes them a great pick for high-volume agents where every credit counts.

Pick one in the builder, or compare them side by side in the model directory.

Which one will you try first?

[thomasumstattd]

Don’t sleep on DeepSeek 4v4 Flash y’all.

This model is almost free to run and almost as smart a Sonnet 5 at a tiny fraction of the price.

$0.14 per 1M input tokens

$0.28 per 1M output tokens

I am testing it right now as a replacement for some of my cheaper models and I’m shocked at how good it is.

[brandoo]

Yes would be good to have the new v4Flash 0731 alongside the basic v4 Flash

[thomasumstattd]

From what I’ve read the DeepSeek Flash-0731 version is now the official version. So if Pickaxe is calling “deepseek-v4-flash” in the API direct from DeepSeek it is calling the “Flash-0731” version. The previous Deepseek-v4-flash version was a preview that looks to be getting rapidly deprecated.

A lot depends on how Pickaxe is sourcing the model. If Pickaxe is using OpenRouter, Together AI, Fireworks, or DeepInfra, they will need to manually upgrade to the newer model.

I suspect Pickaxe is using a US based provider which means the upgrade is not automatic.

[brandoo]

The inference provider is a question here. I still have few minutes generating response on pickaxe while it’s instant on Gemini/Claude/Gpt/etc..

I use regularly duck.ai for some questions and the responses is always ultra fast, this speed on pickaxe would make me leave Hobbyism.

Also, are we sure the DeepSeek v4 Flash available on Pickaxe is the 0731 ?

@chrisfrompickaxe

[briang]

Thanks for asking, to answer your question: We are currently updating DeepSeek v4 Flash to become the 0731 model, we will also keep the other model as well if some users wish to keep using it.

As for the speed, Pickaxe adds a lot of functionality like a RAG knowledge base, user memories, and actions meaning there latency can be higher than the default providers. For the fastest responses, my recommendation is to:

1. Lower reasoning to none or low
2. Use fast response models like Gemini Flash, Haiku, or GPT 5.4 nano.

Let me know if this makes sense.

[thomasumstattd]

That is great news!

If I pick this model:

will it be the 0731 model?

[thomasumstattd]

@briang While you are in there making changes, can you add all of the reasoning options from the API.

Here is what I see

Here are the options from the API.

• none
• low
• high
• xhigh
• max

[briang]

@brandoo @thomasumstattd Thanks for letting us know about these releases.

We’ve now released the new DeepSeek along with two other exciting open-weight models. Let me know if you have any questions!

[thomasumstattd]

Thanks Brian! I just tried the new model and wow. It is so cheap! It opens up whole new use cases for Pickaxe that were not previously financially viable.