Model

INTELLECT-3

OPEN

Prime Intellect · INTELLECT family · released Nov 26, 2025

Prime Intellect (decentralized/open RL training startup) post-trained this 106B/12B-active MoE on top of the openly-licensed GLM-4.5-Air-Base checkpoint, then open-sourced not just the weights but the full RL stack (PRIME-RL framework, verifiers library, 500+ agentic RL environments) under permissive licenses. Served via Prime Intellect's own chat.primeintellect.ai and Inference API; hosted pricing not publicly confirmed.

ReasoningCodingVisionFunction callingTool useAgentic

1719.0

Elo · rank #106

Parameters

106B

Active params

12B (MoE)

Context

98K tokens

Architecture

MoE, post-trained via SFT + large-scale asynchronous RL on top of Zhipu's open GLM-4.5-Air-Base

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath88.0%#54

best: GPT-5.2 · 100.0%

GPQA DiamondReasoning74.4%#113

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning14.6%#82

best: Claude Opus 5 · 64.7%

LiveCodeBenchCoding69.3%#74

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math98.1%#6

best: GPT-5 · 99.4%

MMLU-ProKnowledge81.9%#46

best: Claude Fable 5 · 91.5%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA72.1 GB1× A100 80GB

LoRA225.8 GB2× H200 141GB

Full fine-tune1701.3 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $8.44 (1× A100 80GB)

Plan a fine-tune →

INTELLECT family

Elo progression across releases

INTELLECT-3Nov 20251719.0

Resources

HFPrimeIntellect/INTELLECT-3 DOCShttps://www.primeintellect.ai/blog/intellect-3 APIAvailable via Prime Intellect API

API price weights · each benchmark row carries its own source badge (see methodology)