Model
Model explorer

Falcon-H1 34B

OPEN
TII · Falcon-H1 family · released May 21, 2025

Flagship of the hybrid Falcon-H1 series; 256K context, rivals ~70B dense models on many tasks.

ReasoningCodingVisionFunction callingTool useAgentic
939.0
Elo · rank #225
Parameters
34B
Active params
Undisclosed
Context
256K tokens
Architecture
Hybrid Transformer + Mamba-2 (parallel attention/SSM heads, independently adjustable ratio)
License
TII Falcon-LLM License 2.0
Languages
18+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath23.8%#138
best: GPT-5.2 · 100.0%
AlpacaEvalHuman preference48.3%#21
best: Tülu 2+DPO 70B · 95.1%
AMCMath69.4%#1
best: this model · 69.4%
ARC-ChallengeReasoning61.0%#86
best: Llama 3.1 405B · 96.9%
BIG-Bench HardReasoning70.7%#52
best: ERNIE 4.5 300B-A47B · 94.3%
GPQA DiamondReasoning41.5%#213
best: GPT-5.6 · 94.6%
GSM8KMath83.6%#71
best: Llama 3.1 405B · 96.8%
HellaSwagReasoning81.9%#67
best: Claude 3 Opus · 95.4%
HumanEval+Coding81.7%#14
best: Mistral Small 3.2 (24B) · 92.9%
HumanEvalCoding87.2%#34
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning89.4%#24
best: Gemma 4 26B A4B · 98.5%
LiveCodeBenchCoding49.7%#120
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math83.8%#59
best: GPT-5 · 99.4%
MBPP+Coding71.4%#14
best: Llama 3.1 405B · 88.6%
MBPPCoding83.9%#9
best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%
MMLU-ProKnowledge58.7%#116
best: Claude Fable 5 · 91.5%
MMLU (STEM)Knowledge83.6%#1
best: this model · 83.6%
MMLUKnowledge84.0%#66
best: OpenAI o3 · 92.9%
MT-BenchHuman preference9.2#5
best: Hunyuan-Large (A52B) · 9.4
TruthfulQAKnowledge65.3%#15
best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%
Run it locally
VRAM @ Q4
19 GB
VRAM @ FP16
68 GB
Fits on (Q4)
RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Throughput data unavailable.
Quantizations
GGUF · MLX
Falcon-H1 family
Elo progression across releases
API price weights · each benchmark row carries its own source badge (see methodology)