What it is
Falcon-H1R-7B is a reasoning model from the Technology Innovation Institute (TII) in Abu Dhabi, released on January 5, 2026. It has 7 billion parameters and a hybrid design that mixes standard Transformer attention with Mamba2 layers, a type of state-space model that handles long sequences with less memory. It is built on Falcon-H1-7B-Base, fine-tuned on long reasoning traces and then trained with reinforcement learning.
It reads and writes text, shows its reasoning in think tags, and serves a 262,144-token context by default in vLLM. TII says it runs at up to 1,500 tokens per second per GPU at batch size 64, nearly twice Qwen3-8B.
The weights use TII's Falcon LLM licence. It is royalty-free for commercial use with no revenue cap, but products built on it must state that they use Falcon technology from TII, and all use must follow an acceptable use policy that TII can update.
Our review
Falcon-H1R-7B is a specialist with a good maths record and very little outside confirmation. TII's model card reports 88.1 on AIME 2024, 83.1 on AIME 2025, 97.4 on MATH500 and 68.6 on LiveCodeBench v5-v6, ahead of Qwen3-32B and Phi-4-Reasoning-Plus 14B in its table. gpt-oss-20b still led on AIME 2025 (84.4) and LiveCodeBench (72.0) in the same comparison.
Artificial Analysis has only an estimate: 8 on its Intelligence Index, marked 'independent evaluation forthcoming', which is exactly the median for open models between 4B and 40B. On TII's own figures it is weak at agent work and instruction following, with 4.9 on Terminal-Bench Hard and 53.4 on IFBench against 69.4 for gpt-oss-20b.
Adoption is thin. The main repository logged about 1,700 downloads in the past month and roughly 37,000 in total, a rounding error beside Qwen or Gemma, so expect fewer fine-tunes, quantisation fixes and community answers. The hybrid design also needs recent vLLM, SGLang or Transformers builds; TII pins vLLM 0.11.0.
The licence is the other drag. It is royalty-free with no revenue cap, but all use must comply with an acceptable use policy that TII can revise and users are told to monitor, and any public statement about a derivative must say it is built using TII's technology. The published terms end only the patent licence automatically, if you sue over patents; VentureBeat reads the AUP clause as a wider termination risk. A legal team will treat that very differently from the Apache 2.0 on gpt-oss-20b or Granite 4.2.
Where it earns its place
An ed-tech company running a maths homework helper for secondary-school students
Generate worked solutions to algebra and olympiad-style problems
Self-host on a single GPU and let it reason at length; the hybrid design keeps long traces affordable. Answers still need a verifier, and the product must credit TII's Falcon technology.
A research engineer generating synthetic reasoning data
Produce millions of long chain-of-thought maths traces for fine-tuning a smaller model
TII's claim of about 1,500 tokens per second per GPU at batch size 64 makes bulk generation cheap. The acceptable use policy follows the outputs into any downstream licence.
A procurement lead at a regulated bank evaluating small open models
Decide whether a 7B reasoning model can go through legal review
The maths numbers are strong, but a custom licence with an updatable AUP and attribution clauses costs more review time than Apache 2.0 alternatives.
Before you commit
- Setup
- Serve with vLLM 0.11.0, SGLang or recent Transformers; TII recommends temperature 0.6, top-p 0.95 and up to 65,536 new tokens, with tensor parallelism of 2 for heavy reasoning batches.
- Integrations
- Hugging Face weights plus official GGUF and FP8 repositories; Hermes-style tool calling in vLLM; a Falcon Chat demo but no listed commercial API.
- Data and privacy
- Falcon LLM licence: royalty-free, no revenue cap, mandatory attribution, and an acceptable use policy that TII may update and users must monitor.
- Cost at scale
- Official GGUF files run from 4.6GB at Q4_K_M and 8.1GB at Q8_0 to 15.2GB at BF16, so one 24GB GPU handles production batches; there is no per-token hosted price to compare.
- Limits
- Text only; 262,144-token default context that TII suggests reducing to save memory; prone to long or repetitive reasoning, for which TII advises repetition and presence penalties.
What's good
- 88.1 on AIME 2024 and 97.4 on MATH500, by TII's figures
- Hybrid Mamba2 design keeps long reasoning cheap to serve
- Q4_K_M quantisation is only 4.6GB
- Royalty-free commercial use with no revenue cap
What's not
- Independent Artificial Analysis score is only an estimate
- 4.9 on Terminal-Bench Hard and 53.4 on IFBench
- Licence requires attribution and an updatable acceptable use policy
- Small community: about 37,000 downloads in total
Best for
Teams that need strong competition-style maths reasoning from a model small enough for a laptop, and whose lawyers accept a custom licence.
Alternatives
Falcon-H1R-7B: common questions
- Is Falcon-H1R-7B worth it?
- Falcon-H1R-7B posts maths scores far above its size, but almost nobody has adopted it and the independent tests are still to come. We rate it 3.6 out of 5.
- How much does Falcon-H1R-7B cost?
- Free weights under the Falcon LLM licence (royalty-free; attribution and acceptable use policy required). No public hosted API listed by Artificial Analysis.
- What are the downsides of Falcon-H1R-7B?
- Independent Artificial Analysis score is only an estimate; 4.9 on Terminal-Bench Hard and 53.4 on IFBench; licence requires attribution and an updatable acceptable use policy; small community: about 37,000 downloads in total.
- Who is Falcon-H1R-7B best for?
- Teams that need strong competition-style maths reasoning from a model small enough for a laptop, and whose lawyers accept a custom licence.
- How much does Falcon-H1R-7B cost for a team or in production?
- Official GGUF files run from 4.6GB at Q4_K_M and 8.1GB at Q8_0 to 15.2GB at BF16, so one 24GB GPU handles production batches; there is no per-token hosted price to compare.
- Is Falcon-H1R-7B safe for company data?
- Falcon LLM licence: royalty-free, no revenue cap, mandatory attribution, and an acceptable use policy that TII may update and users must monitor.
- What are the best Falcon-H1R-7B alternatives?
- The closest alternatives we have reviewed are gpt-oss-20b, Granite 4.2, Phi-4-Reasoning-Vision-15B and SmolLM3-3B.
Answers checked Oct 23, 2026.