MobileLLM-R1
Meta's sub-billion reasoning specialists: 140M to 950M models that do math and code on device, not small talk.
About
MobileLLM-R1 is Meta FAIR's attempt to answer whether reasoning can emerge in sub-billion-parameter models, and the answer, at least on Meta's numbers, is yes. The family spans 140M, 360M and 950M parameters in base and final variants, distilled in part from Llama-3.1-8B-Instruct and fine-tuned specifically for math, Python and C++ coding, and science problems. These are not chat models; they are narrow reasoning specialists meant to run entirely on device.
Meta reports the 950M model at roughly 5x the MATH accuracy of Olmo 1.24B and 2x SmolLM2 1.7B, while matching Qwen3-0.6B on MATH (74.0 vs 73.0) and beating it on LiveCodeBench, all on about a tenth of Qwen3's training tokens. Training recipes and data sources are published, which is admirable. The wrinkle is licensing: Meta shipped it under the FAIR non-commercial license at launch, so it is a research artifact, not a production asset, unless the terms have since changed.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Meta (HF model card): 950M ~5x MATH accuracy vs Olmo 1.24B, ~2x vs SmolLM2 1.7B; MATH 74.0 vs Qwen3-0.6B 73.0; LiveCodeBench 19.9 vs 14.9 (provider-reported)
- MobileLLM-R1.5 refresh (Nov 2025) claims further gains; figures not independently verified
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (verified tier, metrics-v1). Evidence: September 12, 2025 release, 140M/360M/950M sizes, 32K final context and benchmark claims confirmed on the facebook/MobileLLM-R1-950M model card; non-commercial FAIR license reported by VentureBeat at launch. Strengths: remarkable reasoning efficiency from a top lab, fully published training recipes, honest scoping as non-chat specialists. Docked for: (1) FAIR non-commercial license blocks commercial use, a hard ceiling for most adopters; (2) all benchmark claims are Meta-reported with conflicting third-party license reports, so re-verify terms on the model card. License: FAIR Non-Commercial (at launch; verify current terms before commercial use).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Meta FAIR; released 2025-09-12
- · 140M/360M/950M reasoning specialists
- · Conflicting third-party license reports (some say Apache 2.0) - flag