Meta's tiny reasoning model punches above its weight

Meta just dropped MobileLLM-R1, a lightweight edge reasoning model family on Hugging Face that ranges from 140M to 950M parameters. It is built specifically to handle math, code, and scientific logic directly on edge hardware.

Unlike general-purpose conversational models, this family skips open-ended chit-chat to squeeze maximum reasoning performance out of sub-billion parameter budgets. The weights are live now on Hugging Face across base and post-trained variants.

How it squeezes reasoning into edge devices

Fitting serious logic into sub-1B architectures usually leads to heavy accuracy drops. Meta tackled this by tweaking the Transformer setup inside MobileLLM-R1-950M, the family's largest model.

It runs 22 Transformer layers with an embedding dimension of 1536 and a hidden dimension of 6144. It relies on Grouped-Query Attention (GQA) with 24 query heads and 6 KV heads, which keeps memory bandwidth and compute in check on smaller devices.

To trim down parameters without triggering latency spikes, Meta used block-wise weight sharing alongside SwiGLU activation. The vocabulary sits at 128K with shared input and output embeddings. The base model supports a 4K context window, while the post-trained reasoning checkpoint stretches to 32K.

Data efficiency is another major highlight. MobileLLM-R1 trained on roughly 4.2T tokens in total. For perspective, Qwen3-0.6B digested around 36T tokens. Meta matched or beat comparable performance using less than 12% of the training budget, followed by targeted supervised fine-tuning across math, code, and reasoning sets.

The benchmark numbers and the catches

On structured benchmarks, the 950M variant punches well above its weight class.

On the MATH500 evaluation, MobileLLM-R1-950M scored 74.0. That is nearly five times higher than Olmo-1.24B (19.2) and double SmolLM2-1.7B (19.2), while nudging ahead of Qwen3-0.6B at 73.0.

It also reached 19.9 on LiveCodeBench, compared to 14.9 for Qwen3-0.6B, and scored 15.5 on AIME'24 and 16.3 on AIME'25.

There are definite tradeoffs, though. GSM8K came in at 67.5, trailing behind Qwen3-0.6B's 79.2. Because Meta tuned it almost exclusively for structured tasks, the model is noticeably weaker at common-sense dialogue and general creative text.

Hardware constraints also apply. While a 32K context is great on paper, sustaining long sequences inflates the KV cache footprint quickly on phones. More importantly, Meta released the weights under a FAIR NC (non-commercial) license, blocking any direct production deployment.

You can pull the models right now from Hugging Face if you want to inspect the checkpoints or run local tests.

Why it matters

In our view, MobileLLM-R1 proves that edge reasoning does not require endless parameters or brute-force data hoarding. Clever architectural trimming and focused post-training can turn a sub-1B model into a genuinely sharp calculator and code assistant.

If you build local developer tools or tinker with edge AI research, it is absolutely worth running locally. But for startups hoping to ship commercial features on phones, the non-commercial license means this release serves mostly as a technical blueprint rather than a plug-and-play solution.