Falcon (TII)
Abu Dhabi's open Falcon-H1 hybrid models punch above their size
Overview
Falcon is TII's (Abu Dhabi, founded 2020) open-source LLM family, built on the Falcon-H1 hybrid Mamba-Transformer architecture (0.5B-34B), with a Falcon Arabic line and reasoning variant Falcon-H1R. Weights are free on Hugging Face; a hosted Falcon Chat playground is also available.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 12 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with Falcon (TII)
- Falcon-H1 hybrid Mamba-Transformer models, 0.5B-34B
- Falcon-H1R reasoning variant
- Falcon Arabic / Falcon-H1-Arabic (up to 256K context)
- Free weight downloads under Apache 2.0-based license
- Hosted Falcon Chat playground
- Hugging Face integration (transformers, vLLM support)
ASEAN Perspective
Falcon (TII) in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
Falcon's hybrid architecture — Mamba state-space layers running in parallel with transformer attention inside every block — is the real story here: TII's own benchmarks show Falcon-H1-34B-Instruct matching or beating Qwen3-32B, Qwen2.5-72B and Llama3.3-70B at roughly half the parameter count, and the 7B reasoning variant (Falcon-H1R) reportedly beats models several times its size on math reasoning (83.1% on AIME 2025, ahead of a 15B and a 32B rival). Long-context support up to 256K tokens on the 7B/34B Arabic models is genuinely useful for legal and document-heavy use cases.
The catch: hybrid-architecture tooling is younger than pure-transformer tooling, so quantization and serving support (vLLM and others) lag, and some users report slower-than-expected throughput on certain configurations versus comparably-sized Qwen models. General knowledge benchmarks still favor larger pure-transformer models. For self-hosters who want an open, sovereign model with strong Arabic support and no license fee, Falcon-H1 is one of the better options on Hugging Face right now.
What people say
Falcon doesn't have the retail review footprint of a SaaS product — it's evaluated by developers on Hugging Face downloads, benchmark leaderboards, and technical blog posts rather than G2 stars, so this summary leans on that evidence instead.
The headline technical claim, independently discussable via TII's published benchmarks and covered by VentureBeat: Falcon-H1R 7B scored 83.1% on the AIME 2025 math-reasoning leaderboard, ahead of Apriel-v1.6-Thinker (15B, 82.7%) and OLMo 3 Think (32B, 73.7%) — a smaller model beating larger ones on a genuinely hard benchmark, which is the kind of result that gets attention in the open-model community precisely because it's unusual. On throughput, TII reports the hybrid design becomes more efficient than pure transformers as context grows, citing up to 4x faster input processing and 8x faster output generation at long sequence lengths, though short-context performance slightly favors traditional transformers.
The rough edges show up in ecosystem maturity rather than model quality. A GitHub issue on the vLLM project documents Falcon-H1 7B running significantly slower than Qwen 7B in that serving framework specifically — a reminder that hybrid Mamba-attention architectures aren't yet first-class citizens in every inference stack. TII itself acknowledges general-knowledge and MMLU-style benchmarks still favor larger, pure-transformer models trained on bigger datasets.
Falcon-H1-Arabic (published January 2026, 3B/7B/34B) is the more distinctive release: context length jumped from the earlier Falcon-Arabic's 32K to 128K (3B) and 256K (7B/34B), enough to process hundreds of pages of legal or medical documentation in one pass, and it's positioned as the most capable open Arabic-language model TII has shipped. Everything ships under the TII Falcon License (Apache 2.0-based) and is downloadable without a paywall — a genuine sovereignty and cost advantage for developers who don't want a metered API dependency, even if the English-language ecosystem still gets more third-party tooling attention than the Arabic line.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including Falcon (TII)'s official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Falcon (TII) unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Falcon (TII) directly →
Spotted something out of date? Suggest an update →
Alternatives to Falcon (TII)
More in LLMs & Chat
