Pinned Tweet
LibertAI Self-Sustaining AI Agents are here! 🤖💰
Your agent can now:
- Hold its own wallet with USDC on @Base
- Pay for LLM calls individually via #x402 (no API keys)
- Monitor & auto-replenish Aleph Cloud credits
- Interact with any onchain protocol via #Coinbase AgentKit
No billing dashboards. No API keys. Just autonomous AI that pays for itself.
Built with:
- Coinbase AgentKit
- LibertAI API (OpenAI-compatible)
- x402 payments
- Aleph Cloud @aleph_im
👉 Start building: docs.libertai.io/agents
LibertAI 🕶️ retweeted
Thank you @VitalikButerin for making this point on stage
This is what we work on at @Libertai_DAI for years: quantizing models so people can run it locally, and for bigger ones we develop ways to have full verifiability (verifiable programs and os images on TEE with signatures) so you don't sent a signle prompt if you can't verify the model and inference stack yourself.
LibertAI 🕶️ retweeted
AI agents? They give us superpowers.
I poked at holes and walls on time travel theories, see where they hold and where they don't.
If you build a time machine, does the universe stop you?
If you build a time machine, does the universe stop you?
LibertAI Labs spent billions of tokens on it. Nine papers, all public with DOIs.
On the Ori time machine, Hawking's quantum safety net shows up. It doesn't hold.
Preprints → labs.libertai.io/priority
DeepSeek-V4.1-Flash-1017 on 2 DGX Sparks via @Libertai_DAI 2.9 bpw for exllamav3: 196 GiB of weights.
Running @ 30 tok/s .
LibertAI 🕶️ retweeted
I have tested nvidia/GLM-5.3-Flash-FP4-Local and one of the best DSV4.1 Flash EXL3 quads of the LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw on a 2 x DGX Spark cluster on real coding and math problems. Better output was given by the QAD NVFP4 quants of the GLM 5.3 Flash distilled by NVIDIA. No looping, better instruction following, working fine in Deepseek-Harness. It is my #1 choice for Local AI on 2 DGX Spark machines now.
Thanks @NVIDIAAI for supporting local AI.
LibertAI 🕶️ retweeted
DeepSeek v4.1 Flash 1017 Fully Uncensored - for you 2x dgx spark people!
I got lazy and cba to make my own exl3 quant atm (wanted to get this out for u guys asap) so go thank @Libertai_DAI for the exl3 2.9bpw quant. In the next few days I’ll be making my own ~2.8-2.9bpw quant thats strongly calibrated towards cybersec and reasoning.
Fully complies with all requests, issue is that -4% difference is a bit lower than I’d personally like it to be (not blaming libertai, I just need to do better on the second iteration even though this is quick A/B/C/D answering without any reasoning)
Crazy that when I got started with all this like half a year ago, models like MiniMax m2.5 was nowhere close to how capable this model is, even when its been compressed down to all hell. The world is changing so fast.
huggingface.co/dealignai/Dee…
LibertAI 🕶️ retweeted
For Science I am going to load this up. TP4 vllm and the Tensorfold
1017, part 2: EXL3.
DeepSeek-V4.1-Flash-1017 at 2.9 bpw for exllamav3: 196 GiB of weights. The first EXL3 of 1017 on Hugging Face.
Perplexity vs the base model: +1.8% on code, +0.4% on prose.
🤗 huggingface.co/LibertAIDAI/D…
LibertAI 🕶️ retweeted
もうDeepSeek-V4.1-Flash-1017のEXL3量子化が爆誕してる……
huggingface.co/LibertAIDAI/D…
196.3 GiBだからDGX Spark ×2で動かせるサイズではあるな
Deepseek v4.1 Flash 1017 is now available on LibertAI inference.
Code at up to 800 tok/second. Prose at up to 500.
6
557
LibertAI 🕶️ retweeted
Same prompt, same 2 DGX Sparks, same speed.
-DeepSeek V4.1 Flash: 46,355 tokens, 20.6 min
-V4.1 Flash 1017: 38,146 tokens, 16.9 min
18% fewer tokens for the same pagoda. Both 37.6 tok/s, thinking on, no output cap.
EXL3 2.9bpw quants by @Libertai_DAI (1017)
@deepseek_ai
1017, part 2: EXL3.
DeepSeek-V4.1-Flash-1017 at 2.9 bpw for exllamav3: 196 GiB of weights. The first EXL3 of 1017 on Hugging Face.
Perplexity vs the base model: +1.8% on code, +0.4% on prose.
🤗 huggingface.co/LibertAIDAI/D…
DeepSeek dropped V4.1-Flash-1017 overnight.
Less than 7 hours later it was on Hugging Face in NVFP4: 475 GiB → 400 GiB.
The first NVFP4 of 1017. And the routed experts are bit-exact. 👇
🤗 huggingface.co/LibertAIDAI/D…
Now at 3.5 bpw too: 240 GiB.
Same sets, code / prose:
base 1017: 1.8749 / 6.4277
3.5 bpw: 1.8895 / 6.4341 (+0.8% / +0.1%)
2.9 bpw: 1.9094 / 6.4520
Kit users: point KMAP_HOST at this repo's exl3_k_map.json.
🤗 huggingface.co/LibertAIDAI/D…
Measured on held-out, unpublished text, 32 × 2048 tokens per set (code / prose):
base 1017: 1.8749 / 6.4277
EXL3 2.9 bpw: 1.9094 / 6.4520
Switch Engram off in the same quant and it jumps to 4.16 / 9.06. Those n-gram tables do real work.
1
1
378
DeepSeek dropped V4.1-Flash-1017 overnight.
Less than 7 hours later it was on Hugging Face in NVFP4: 475 GiB → 400 GiB.
The first NVFP4 of 1017. And the routed experts are bit-exact. 👇
🤗 huggingface.co/LibertAIDAI/D…
Open weights. Open method. Zero gatekeeping.
This is what LibertAI does: frontier models you can actually own, not rented access to someone's API.
→ libertai.io
🤗 huggingface.co/LibertAIDAI/D…
1
1
111
Don't want the 400 GiB download?
1017 is already live on the LibertAI API as deepseek-v4.1-flash (and deepseek-v4.1-flash-1017). The -thinking variant runs at max reasoning effort.
→ libertai.io
84

















