Falcon (TII)
B Tier · 7.1/10
UAE's Technology Innovation Institute open-weights family -- Falcon 3 optimized for efficient sub-10B deployment on consumer hardware
Score Breakdown
Benchmark Scores
Benchmarks for Falcon 3 10B
| Benchmark | Description | Score | |
|---|---|---|---|
| MMLU | Knowledge across 57 subjects | 73.1% | |
| GPQA Diamond | Graduate-level science questions | 42.5% | |
| HumanEval | Python code generation | 73.8% | |
| MATH | Math problem solving | 55.4% |
Last updated: 2026-04-13
Personality & Tone
The TII research release
Tone: Workmanlike and neutral. Falcon reads more like an academic reference than a chatbot -- answers are straight, structured, and unremarkable in voice.
Quirks: Built as a research artifact from UAE's TII, not a consumer product. Less instruction-tuning polish than Llama 4 or Qwen and a smaller community of fine-tunes, so the base model is effectively what you use.
The Good and the Bad
What we like
- +Apache 2.0 license -- fully permissive for commercial use
- +Sub-10B sizes run on any consumer GPU or even CPU with acceptable speed
- +Falcon 3 Mamba variant offers state-space architecture for cheap long-context inference
- +Backed by UAE government funding -- long-term viability is strong
- +Strong multilingual performance including Arabic (a gap in most Western open-weights models)
What could be better
- −Not frontier quality -- Falcon 3 10B is B/C-tier vs. Qwen3, Gemma 4, Llama 4 in the same size class
- −Smaller community than Llama, Qwen, Mistral -- fewer fine-tunes and tools
- −Original Falcon 180B (2023) was hyped but quickly obsoleted -- damaged reputation somewhat
- −Falcon 3 release cadence has slowed since 2025
- −No flagship frontier-size model in 2026 -- TII is focused on efficient small models
Pricing
Self-hosted (Free)
- ✓Apache 2.0 with Acceptable Use Policy
- ✓Commercial use permitted
- ✓Weights on Hugging Face
API (Hugging Face Inference, third-party)
- ✓Hosted via HF Inference Endpoints
- ✓Together.ai partial support
- ✓Small community of API hosts
System Requirements
Hardware needed to self-host. Min = smallest viable setup (usually heavy quantization). Max = full-precision / production-grade.
| Model variant | Min | Max |
|---|---|---|
| Falcon 3 7B / 10B (dense) | 4 GB VRAM (Q4) | 16 GB VRAM FP16 |
| Falcon 3 Mamba 7B (state-space hybrid)Mamba architecture gives cheap long-context inference | 4 GB VRAM (Q4) | 16 GB VRAM FP16 |
Known Issues
- Falcon 3 10B trails similarly-sized Qwen3 and Gemma 4 on most benchmarks -- pick it for licensing/multilingual, not peak qualitySource: Artificial Analysis, Hugging Face discussions · 2026-03
- Falcon 3 Mamba 7B has limited llama.cpp support vs. standard transformer variantsSource: GitHub ggerganov/llama.cpp issues · 2026-02
Best for
Developers who need a genuinely Apache-2.0 small model for on-device or edge deployment, or who need strong Arabic/multilingual support.
Not for
Anyone chasing peak benchmark quality -- Qwen3, Gemma 4, Llama 3.3 all beat Falcon 3 in their respective size classes. Also not ideal for agentic or tool-use workflows.
Our Verdict
Falcon is the niche-but-viable choice in 2026. TII has carved out a sensible position: efficient sub-10B Apache-2.0 models with strong Arabic support. It's not trying to compete with DeepSeek or Qwen at the frontier, and that's fine. If you need a small permissively-licensed model for edge deployment and the multilingual mix matters, Falcon 3 is a real option. For most other use cases, Qwen3 or Gemma 4 in the same size class outperform it.
Sources
- Falcon LLM official site (TII) (accessed 2026-04-13)
- Hugging Face blog: Falcon 3 (accessed 2026-04-13)
- Hugging Face tiiuae collection (accessed 2026-04-13)
- Artificial Analysis open-weights leaderboard (accessed 2026-04-13)
Explore more Falcon (TII) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for Falcon (TII).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to Falcon (TII)
Llama 4 (Meta)
Meta's open-weights family -- Scout (10M context), Maverick (multimodal 400B MoE). NOTE: Meta's frontier work moved to the proprietary Muse Spark line in April 2026; Llama remains downloadable and supported but is effectively in maintenance mode
Mistral AI
European AI lab with open and commercial models -- Le Chat is now **Vibe** (May 28 2026): one agent across Work Mode + Code Mode with a VS Code extension and CLI, powered by Mistral Medium 3.5 (128B dense, 256k context, 77.6% SWE-Bench Verified). Earlier 2026 line: Small 4 (119B MoE Apache 2.0), Medium 3, Voxtral TTS
DeepSeek
DeepSeek V4 shipped 2026-04-24: V4-Pro (1.6T/49B active MoE) + V4-Flash (284B/13B active), 1M native context, Hybrid Attention Architecture, open-source on HF. Trails only Gemini 3.1 Pro on world knowledge
Gemma 4 (Google)
Google DeepMind's open-weights model family -- multimodal, 256K context, runs on edge devices
DiffusionGemma (Google)
Google DeepMind's experimental open-weights TEXT-DIFFUSION model (June 10, 2026) -- 26B MoE (3.8B active), Apache 2.0, generates 256-token blocks in parallel with bidirectional attention for up to 4x faster output (1,000+ tok/s on H100). Trades some quality vs Gemma 4 for raw speed
Qwen (Alibaba)
Alibaba's open-weights + API family -- Qwen 3.7 Max flagship GA (May 20 2026: SWE-Bench Pro 60.6%, Terminal-Bench 69.7%, GPQA 92.4%, $2.50/$7.50 per 1M with 50% promo until 6/22), Qwen3.7-Plus multimodal API (Jun 2), Qwen3.6-27B dense Apache 2.0 (beats the 397B MoE on coding from one consumer GPU)
GLM / Z.ai (Zhipu AI)
Zhipu AI's open-weights flagship -- GLM-5.2 (launched 2026-06-13) is a ~753B-parameter MoE with a 1M-token context and the new IndexShare sparse-attention architecture (~2.9x lower per-token FLOPs at 1M context), MIT licensed. Vendor benchmarks put SWE-Bench Pro at 62.1 (up from GLM-5.1's 58.4) and it tops the Artificial Analysis open-weights Intelligence Index; VentureBeat reports it beats GPT-5.5 on several long-horizon coding benchmarks at roughly 1/6 the cost. Drop-in for Claude Code / Cline / OpenCode. Still trained outside the Nvidia stack on Huawei Ascend silicon
Kimi K2.6 (Moonshot)
Moonshot's 1T-parameter MoE open-weights flagship -- Kimi K2.6 (GA 2026-04-20) is #1 open-weights on Artificial Analysis Intelligence Index v4.0 (score 54, ranked #4 overall). Native video input, 256K context, Modified MIT license
Nemotron (Nvidia)
Nvidia's open-weights family -- hybrid Mamba-Transformer MoE architecture, optimized for efficient reasoning on Nvidia hardware. Nemotron 3 Ultra (550B total / 55B active) shipped 2026-06-04 as the family flagship, joining Super (120B/12B, March) and Nano
MiniMax M3
MiniMax's coding/agent flagship -- M3 (June 1 2026): 1M-token context, MSA sparse attention (>15x decoding speedup at long context), SWE-Bench Pro 59.0%, Terminal-Bench 66.0%. OPEN WEIGHTS LIVE on HuggingFace since June 12 (~428B total / ~23B active, native multimodal, minimax-community license)
gpt-oss (OpenAI)
OpenAI's FIRST open-weight models -- gpt-oss-120b (single 80GB GPU, near parity with o4-mini on reasoning) and gpt-oss-20b (runs on 16GB edge devices). Apache 2.0. Launched 2025-08-05. gpt-oss-safeguard ships in 2026 as the safety-tuned variant
IBM Granite 4.0
IBM's enterprise-focused open-weight family -- Granite 4.0 hybrid Mamba-2 + transformer architecture (70-80% memory reduction vs pure transformer), 3B to 32B sizes, Apache 2.0. First open model family to secure ISO 42001 certification. Nano 350M runs on CPU with 8-16GB RAM. 3B Vision variant landed 2026-04-01
Arcee Trinity-Large-Thinking
Arcee AI's US-made open-weight frontier reasoning model -- launched 2026-04-01. 398B total params, ~13B active. Sparse MoE (256 experts, 4 active = 1.56% routing). Apache 2.0, trained from scratch. #2 on PinchBench trailing only Claude 3.5 Opus. ~96% cheaper than Opus-4.6 on agentic tasks
Olmo 3 (AI2)
Allen Institute for AI's fully-open frontier reasoning models -- Olmo 3 family (2025-11-20) includes 7B and 32B sizes, four variants (Base, Think, Instruct, RLZero). Apache 2.0 with fully open data + checkpoints + training logs. Olmo 3-Think 32B matches Qwen3-32B-Thinking at 6x fewer training tokens
AI21 Jamba2
AI21 Labs' hybrid SSM-Transformer (Mamba-style) open-weight family -- Jamba2 launched 2026-01-08. Two sizes: 3B dense (runs on phones / laptops) and Jamba2 Mini MoE (12B active / 52B total). Apache 2.0, 256K context, mid-trained on 500B tokens
StepFun Step 3.7 Flash
StepFun's (China) agent-focused open-weight family -- Step 3.7 Flash (May 28 2026): 198B sparse MoE vision-language model, ~11B active, 256K context, Apache 2.0, ~400 tok/s, SWE-Bench Pro 56.3. Supersedes Step 3.5 Flash (Feb 2026) as the flagship
Cohere Command A
Cohere's enterprise-multilingual flagship -- 111B params, 256K context, runs on 2x H100. 23 languages. CC-BY-NC 4.0 on weights (research / non-commercial), commercial requires Cohere enterprise contract. Follow-ups: Command A Reasoning + Command A Vision
LongCat-2.0 (Meituan)
Meituan's open-source 1.6T-parameter MoE (~48B active) with native 1M-token context, MIT license -- trained entirely on domestic Chinese AI ASICs and revealed as the stealth 'Owl Alpha' model that had been topping OpenRouter