A handful of Snapdragon Summit announcements caught my eye, and they all point the same way. Tiny models squeezed below 4 bits per weight are being tuned for the neural processors inside glasses and phones, and the chips are being built with those models in mind. Let’s walk through what was shown, what the numbers tell us, and what I’d watch as a builder.
A few terms before we start
Some quick vocabulary first. An NPU, or neural processing unit, is a block on a chip designed for the matrix math that neural networks run on. It does that work on far less power than a CPU or GPU. Qualcomm’s is called the Hexagon NPU.
Quantization means storing a model’s weights with fewer bits. Models are usually trained in 16-bit formats, and 4-bit versions have become the everyday choice on laptops and phones. A “1-bit” model keeps each weight at roughly one bit. That makes it dramatically smaller, and the hard part is keeping it smart.
Two more terms. Prefill is the phase where the model reads your prompt (and, for vision models, the image) before it writes anything. It largely decides how long you wait for the first word. Decode is the phase where it generates the answer one token at a time.
A 2B vision model inside 4 GB glasses
The one I found most striking came from PrismML. In its September 23 announcement, the company showed a 2 billion parameter vision-language model running locally on smart glasses built on the Snapdragon AR1 Gen 1 platform. The model pairs a 1.7B 1-bit language model with a 0.3B vision encoder at 4 bits.
The footnotes are worth reading. Qualcomm ran the tests on a platform with 4 GB of RAM, 6 TOPS of peak AI compute, and a 1,024-token context. Qualcomm measured 0.43 GB of language model weights for the 1-bit version against 1.66 GB for a 4-bit equivalent, which is about 3.8 times smaller. It also measured 15.36 tokens per second against 7.44, roughly double the speed. PrismML says quality matches 4-bit Qwen3 1.7B on its own evaluations, including MMLU Redux, GSM8K, and HumanEval+.
That’s a lot of capability for something that sits on your face all day. PrismML’s CEO, Babak Hassibi, calls the approach “model-hardware co-design,” which sums up the whole story nicely.
Now for the limits. Bonsai 1.7B came out earlier this year. The fresh work here is the vision model built on it and the port to the Hexagon NPU, using an internal Qualcomm SDK with 1-bit kernel support. No glasses running this model have been announced, and I haven’t seen confirmation that the vision model’s weights will be public.
A shared context layer for on-device agents
The same day, Liquid AI announced Liquid Context with Qualcomm. With your permission, it builds a picture of your routines and preferences from device signals and shares that context with whichever agents you choose, local or cloud. Liquid says it’s optimized for the Hexagon NPU so it can update in the background without a cloud model processing every signal.
Alongside it is Liquid Agent, an embedded agent that OEMs can license and customize. It runs on LFM2.5-2.6B, which Liquid released earlier in September and has now tuned for the NPU. I like the architecture here. A permissioned memory layer that any agent can read is a sensible way to let small local models feel personal without shipping your life to a server.
It’s still a pitch to device makers, though. No OEM has been named, and the companies describe themselves as “exploring additional opportunities” together.
Faster answers from a small vision model
On September 24, Liquid followed up with LFM2.5-VL-3B-DSpark, a drafter for its 3B vision-language model. It uses speculative decoding, a clever trick. A small, fast “draft” model guesses the next several tokens, and the big model checks all of those guesses in a single pass. Correct guesses give you several tokens for the price of one, and wrong ones get corrected, so with greedy decoding the output matches the big model running alone.
The drafter has about 280M parameters, adding roughly 9% to the deployed model. Liquid reports decode speedups of up to 3.13x and end-to-end gains of up to 2.62x with MLX-VLM on an M5 Max. With llama.cpp on an M3 Ultra, it reports 1.57x to 2.14x faster decoding. Those runs used 16-bit weights, and Liquid says speeding up quantized models is outside the scope of this release.
Liquid is refreshingly candid about the catch. Speculative decoding only speeds up decode. Vision encoding and prompt prefill stay the same, and on edge hardware they take a big share of total time, so a large decode speedup becomes a smaller end-to-end gain. The drafter is available as GGUF on Hugging Face, with support in llama.cpp and MLX-VLM, under Liquid’s LFM license, which Hugging Face lists as “other.”
Chips built for very low precision
The hardware side came a day earlier. Qualcomm announced the Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6 on September 22. According to press coverage from RCR Tech, the Hexagon NPU supports a precision ladder from INT2 and INT4 through INT8, INT16, FP8, and FP16, and on the Extreme Qualcomm claims up to 50% faster prefill for INT4 models. RCR also reports 50% more shared memory for the Extreme’s NPU.
INT2 support in a flagship phone NPU is the hardware counterpart to models like Bonsai. Chip designers clearly expect very low precision models to be a real workload.
Reading the numbers honestly
I’m genuinely excited about this direction, and I want to be careful with it too. Every number above is vendor-reported: Qualcomm ran the Bonsai tests, PrismML the quality evaluations, and Liquid the DSpark benchmarks. None of it has been independently reproduced yet.
These are also demos. The glasses aren’t on sale, Liquid Context doesn’t have a named device maker, and the new Snapdragon phones haven’t reached reviewers. Bonsai 1.7B and LFM2.5 both came out earlier this year, and the fresh work is the NPU port, the Bonsai vision model, and the DSpark drafter.
What I’d watch as a builder
Three things will tell us whether this becomes something you can build on. First, whether PrismML publishes the weights for the Bonsai vision model, so the community can test the quality claims. Second, whether a named OEM ships Liquid Context in a real device. Third, independent tokens-per-second and quality numbers from 8 Elite Gen 6 phones when they arrive in Q4.
In the meantime, I’d encourage you to read the primary sources yourself, especially the PrismML footnotes and the limitations section of the DSpark post. And if you have a Mac or a decent machine handy, grab the DSpark drafter and try it: the GGUF pairs with the LFM2.5-VL-3B GGUF in llama.cpp, and there’s a Safetensors build for MLX-VLM. Watch the acceptance rate, time prefill on your own images, and compare with the published numbers. That hands-on feel for where the time goes is great preparation for when these chips and models reach people’s hands.