Kinetic
Open-source engine for faster local language model inference with bit-for-bit identical output
- Upvotes
- 0
- DR
- 0 no change since last check
- Followers
- 0
Kinetic runs language models from GGUF files faster than llama.cpp on CPU and GPU, with all optimizations preserving exact output.
About Kinetic
Kinetic is an open-source inference engine developed by Velocity that executes language models faster than llama.cpp while maintaining bit-for-bit identical outputs. Written in Rust under the Apache-2.0 license, it runs models including Gemma 4 and Qwen3.5 with images from unchanged GGUF files on CPUs, NVIDIA GPUs via CUDA, and other GPUs via Vulkan.
The engine implements nine default-on optimization techniques including speculative decoding, prompt caching, continuous batching, and others that are all exact: greedy output remains token-for-token identical to plain decoding, and seeded sampled answers are the same with and without optimizations. On an RTX 3060 Laptop GPU with 6GB, editing requests run 4.5× faster on LFM3-1.7B and 2.9× faster on Qwen3.5-4B compared to llama.cpp.
Kinetic ensures reproducibility through receipts that record all request parameters, allowing verification that answers are identical when rerun. Every technique can be disabled individually for measurement, and performance is validated on the same GGUF files used by llama.cpp.
- inference-engine
- local-inference
- rust
- gpu
- llm
- gguf
- speculative decoding
Features worth knowing
- Speculative decoding for faster token generation
- Prompt caching for context reuse
- Real-time prompt processing while typing
- Concurrent request batching
- Bit-for-bit identical output verification
- Multi-GPU and multi-backend support (CUDA, Vulkan)
- Vision model support with image encoding
- Mixture-of-experts model support
- RAD-1 automatic draft token generation
- Cross-platform support for Windows and Linux
Pricing
Pricing model: Open source
Follow Kinetic and hear about their milestones, new launches and demo videos by email. Stop any time.
Kinetic metrics
| Metric | Value |
|---|---|
| Domain Rating | 0 no change since last check |
| Trust Flow | 0 no change since last check |
| Citation Flow | 11 no change since last check |
| Referring domains | 2 no change since last check |
| External backlinks | 4 no change since last check |
Ahrefs Domain Rating scanned 11 Nov 2026, updated 10 Nov 2026. Majestic scanned 10 Nov 2026, updated 10 Nov 2026. Link stats via Majestic. Trust Flow, Citation Flow and Topical Trust Flow are trademarks of Majestic-12 Ltd.
Kinetic domain facts
Reveal the domain facts
| Domain age | 6 months |
|---|---|
| Registered | |
| Expires | |
| Registrar | Key-Systems GmbH |
| DNSSEC | Not signed |
Kinetic launch results
| Measure | At launch, then latest |
|---|---|
| Domain Rating (Ahrefs) | 0 first measured 10 Nov 2026 |
| Trust Flow (Majestic) | 0 first measured 10 Nov 2026 |
| Referring domains (Majestic) | 2 first measured 10 Nov 2026 |
Day 0 since launch. Next receipt: day 30, 10 Dec 2026. The figures report a change; they do not say what caused it, and many things can. No figure here is a guess.
Launching something soon?
Free tips from Best Digital Launch for getting ready, launch day, the week after and what to do next.
- Set up
- On launch day
- The days after
- Beyond launch
Compare and explore
FAQ
- How does Kinetic maintain bit-for-bit identical output with optimizations enabled?
- Kinetic ensures exact output by implementing one rule: every row of a batch is summed in the same order as that token would be on its own. On the CPU, this comes from the block structure of GGUF weights where products inside a block are added in 32-bit integer arithmetic (exact in any order), and only the fold of block sums is done in floating point always in the same order. On the GPU, Kinetic resumes a cached conversation only from states that a cold run of the new prompt also reaches, bit for bit.
- What hardware did Kinetic benchmark against llama.cpp?
- All benchmarks were measured on an RTX 3060 Laptop GPU with 6 GB memory and a Ryzen 5 6600H processor against llama.cpp build b12001.
- How quickly does Kinetic return the first token for long documents?
- After pasting a 22,631-token document and typing an 80-character question, the first token arrived 0.13 seconds after pressing Enter with Kinetic's Zero Wait optimization, compared to 10.95 seconds with llama-server.
- What models does Kinetic support?
- Kinetic runs language models including Gemma 4 and Qwen3.5 with images from unchanged GGUF files on CPU, NVIDIA GPUs through CUDA, and other GPUs through Vulkan.
- Where is Kinetic slower than llama.cpp?
- A 512-token prompt on LFM3-1.7B runs at 0.89× llama.cpp's speed, and a 23-token prompt on Qwen3.5-4B at 0.55×. Without Zero Wait, a long prompt takes longer. When no source can guess the next tokens, RAD-1 gives nothing and speed is plain generation.
Leave a review for Kinetic
Nobody has reviewed Kinetic yet. Start things off.
A review is always free and can never be paid for. Go to the reviews and discussion
Reviews and chat · Product news · Send it on · Run this company? Claim it here · Report it
Kinetic has a free listing. The owner can unlock the full profile with a paid launch for $5:
- Founder story (locked)
- Milestones (locked)
- Demo video (locked)
- Dofollow website link (locked)
Makers
- Jakub KozicaCEO & Founder of Velocity
Alternatives to Kinetic
YABANIN
Self-hosted AI infrastructure with measurable returns
AI inference & GPU cloud0 votes. Log in to add your vote for YABANINQuantum Computing Inc
Putting photons to work
AI inference & GPU cloud0 votes. Log in to add your vote for Quantum Computing IncKaeso
Open-source AI gateway you run yourself
AI inference & GPU cloud0 votes. Log in to add your vote for KaesoLoops
Python SDK for LoRA fine-tuning with supervised fine-tuning and reinforcement learning
AI inference & GPU cloud0 votes. Log in to add your vote for Loops
