NEW Liquid AI Models are WILD!

What Liquid AI Deaspark Is and How It Works

Liquid AI Deaspark is a speculative decoding engine that makes local AI models run up to 3.18 times faster without making them bigger. It does this by using a tiny draft model called Deaspark, which is about 300 million parameters, to guess a block of tokens ahead in one shot. Then the main model verifies those guesses.

If the guesses are correct, the main model keeps all of them. If one guess is wrong, the main model corrects from that point and continues. The result is much faster token generation while the output stays identical to the non-accelerated version. The main model stays in charge of quality, so you do not lose accuracy for speed.

Main Features of Liquid AI Deaspark

• Up to 3.18x faster on an H100 server GPU and up to 2.87x faster on an M4 Max MacBook Pro.
• Works with three Liquid Foundation Models: LFM 1.2B, LFM 2.6B, and LFM 8B (mixture of experts).
• Output is identical because the main model verifies every token. The draft model only speeds up the process.
• Compatible with llama.cpp, MLX, VLLM, SGLANG, and ONNX, so it fits into existing setups without rebuilding.
• No need for larger hardware — Deaspark works on consumer laptops and edge devices.

How to Set Up Liquid AI Deaspark

1. Download one of the LFM models: LFM 1.2B, LFM 2.6B, or LFM 8B from Liquid AI's official site.
2. Install a compatible inference engine like llama.cpp, MLX, or VLLM on your machine.
3. Download the Deaspark draft model (around 300 million parameters).
4. Configure your inference engine to use Deaspark as the speculative decoding drafter alongside the main LFM model.
5. Run a test prompt to verify the speed increase. You should see token generation rates that are up to 3x faster than without Deaspark.
6. Integrate the accelerated model into your application or agent workflow using the same API you normally use.

How to Use Deaspark in Real Agent Workflows

In this episode, Julian Goldie walked through two example workflows for an AI community called the AI Profit Boardroom. First, a lead qualification agent. When someone fills out a form or sends a DM about the community, the LFM 2.6B agent running locally picks it up. You give it a prompt to read the inquiry, identify the business type, the AI problems they want to solve, their readiness, and score them 1 to 10. Then the agent writes a personalized reply under 150 words, matching the community's features to the person's needs, and logs it to your CRM.

Second, a content research agent. For weekly videos on new AI tools, the agent uses a prompt to read through source material, pull the most important points, and produce a 200-word video brief. Deaspark speeds up the token generation for every reasoning step, so what used to take hours now takes minutes. Both agents run locally and privately.

Results and Speed Benchmarks Shown

Liquid AI released benchmark numbers for three LFM models with Deaspark enabled. On an H100 server GPU, the LFM 8B model went from 428 tokens per second to 1362 tokens per second, a 3.18 times improvement. On the same hardware, the LFM 2.6B model went from 323 to 864 tokens per second, a 2.67 times speedup.

On an M4 Max MacBook Pro, the LFM 1.2B model went from 136 to 389 tokens per second, a 2.87 times speedup. The 2.6B model also showed a 2.87 times improvement on the MacBook. All benchmarks were run on consumer hardware that most people already own. The output quality stayed identical because the main model still verifies every token. No trade-off between speed and accuracy.

Tips, Limits and Mistakes to Avoid

• Deaspark only works with supported LFM models. You cannot use it with other large language models like Llama or GPT unless they support speculative decoding.
• The draft model is small (300M parameters), so it works best with main models that have a compatible vocabulary. Make sure you use the correct pairing.
• Do not expect a 3x speedup on every task. The improvement depends on batch size, prompt length, and hardware. Benchmarks are averages.
• For agent workflows, the speed gain matters most when you have many sequential reasoning steps. A single quick query may not show the full benefit.
• Always verify output quality during testing. Even though quality should be identical, edge cases can occur if the draft model guesses are consistently wrong.

Who Should Use Liquid AI Deaspark

Liquid AI Deaspark is for business owners, marketers, coaches, and agency owners who want to run AI agents locally without spending money on expensive cloud GPUs. If you already use local AI models for tasks like lead qualification, content research, member onboarding, or support automation, Deaspark lets you do them faster.

It is also for developers building agentic workflows with LFM 2.6B, which is designed for planning, tool calling, and multi-step reasoning. The speed boost makes agents feel more responsive. And because everything runs locally, your data stays private. If you want to automate parts of your business on autopilot and save time, this tool removes the speed bottleneck of local AI.

Frequency of Updates and Community Support

Liquid AI regularly releases model updates and improvements to Deaspark. The latest LFM models are designed specifically for agent workloads. The community mentioned in this episode, the AI Profit Boardroom, has 3,800 members who share setups, troubleshoot together, and swap prompts. There are also 75,000 members in the free AI Success Lab.

Both communities offer daily tutorials, live coaching calls, and a 30-day roadmap to implement these workflows. If you need help setting up Deaspark or building agents, you can find step-by-step guides and support from other users who are already running these systems locally.