INTELLECT-3

Prime Intellect (Prime Intellect, Inc.) · analysed 2026-08-10 · skill v0.3.5

sparse-MoE full-attention text-only open-weights-open-data frontier reasoning-specialised

Design overview

INTELLECT-3 is a 106B-parameter sparse MoE (12B active, 46 layers, hidden 4096, per-expert MoE dim 1408) that Prime Intellect built entirely by post-training Z.ai's GLM-4.5-Air base — the report discloses no block-level detail of its own, since every architectural attribute is inherited. The contribution is the training system around that fixed block. Two SFT stages establish behavior: a general chat-and-reasoning stage (~33M tokens/step at 65K context, Muon, warmup to 5e-5) and a smaller agentic stage at 98K context via context parallelism, using a Qwen3/GLM-style chat template that always prepends a <|think|> token and preserves reasoning across turns. Then a large-scale RLVR stage (batch 256 × 16 rollouts, 65,536 max context) optimizes the model across math, code, science, logic, deep-research (Serper search tools) and software-engineering environments with verifiable rewards — rule-based math-verify plus a CompassVerifier-7B LLM judge, sandbox test execution, and binary search rewards. The RL loop is the engineering centerpiece: prime-rl runs trainer (FSDP2) and inference (vLLM) disaggregated, with continuous batching, in-flight weight updates (max_off_policy_steps 8), a multi-client orchestrator for linear inference scaling, online difficulty filtering, and an IcePop-style double-sided masked importance-sampling objective (α=0.5, β=5) that the authors found critical for stable long-horizon training. Execution happens in Prime Sandboxes (gVisor containers, Rust gateway, image streaming, up to 256 sandboxes per node) at 4,000+ concurrent rollouts, on a 512-H200 cluster with distributed all-to-all Muon. The report frames the model as evidence for an open, reproducible RL stack as much as for the model itself.

What the report claims is novel

Atlas assessment: INTELLECT-3 contributes no new model architecture: the report states plainly that it is an SFT+RL post-train of Z.ai's GLM-4.5-Air base, and every block-level attribute (attention, norm, activation, tokenizer, expert routing) is inherited and undisclosed. Its genuine novelty is the open RL infrastructure — prime-rl's disaggregated async trainer/inference with in-flight weight updates and multi-client orchestration, the verifiers/Environments Hub ecosystem, and Prime Sandboxes' control-plane-bypassing execution path — plus the RL algorithm detail of double-sided masked token-level importance sampling (α=0.5, β=5, 1e-5 floor) with ablations against GSPO/CISPO. vs deepseek-r1 (the atlas's other open RLVR post-train): INTELLECT-3 is built on an MoE base instead of R1's dense V3-base, uses Muon rather than AdamW, and emphasizes multi-turn agentic environments (SWE, deep research) alongside math/science. vs openai-gpt-oss: both are open post-trained reasoning MoEs, but INTELLECT-3's base (GLM-4.5-Air, ~355B-class MoE family) and full infrastructure disclosure differ. The headline claims (SOTA in weight class, matching GLM-4.6) rest on the report's own harness (Appendix A) and are not independently verifiable here; the report itself flags that its AIME numbers are conservative (no LLM judge).

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyINTELLECT-3[1]
OrganisationPrime Intellect (Prime Intellect, Inc.)[2]
Release datenot disclosed
Report URLhttps://storage.googleapis.com/intellect-3-paper/INTELLECT_3_Technical_Report.pdf[3]
Licensenot disclosed
Open weightsyes[4]

Scale

Total parameters106B[5]
Active parameters / token12B[6]
Layers46[7]
Hidden dimension4096[8]
FFN inner dimension1408[9]
FFN ratio0.34
Query headsnot disclosed
KV headsnot disclosed
Head dimensionnot disclosed
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention variantnot disclosed
Attention layer patternnot disclosed
Depth mixingsequential-residual
Activationnot disclosed
Expert countnot disclosed
Experts per tokennot disclosed
Shared expertsnot disclosed
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksnot disclosed
Softcappingnot disclosed
Othernot disclosed

Context

Trained context length65,536[10]
Deployed context lengthnot disclosed
Extension methodnot disclosed

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokensnot disclosed
Data compositionSFT (Table 1): OpenReasoning-Math (2M/78.1B), OpenReasoning-Code (1.9M/94.3B), OpenReasoning-Science (310K/32B), OpenReasoning-Tool (800K/3.8B), AM General Chat (952K/8.4B), AM Instruction Following (54K/400M), SWE Swiss (10.3K/700M), Toucan Tool (116K/700M), Environments Mix (38.4K/1.9B); main sources are NVIDIA's Nemotron-Post-Training-Dataset-v1 and AM-DeepSeek-R1-0528-Distilled (synthetic reasoning traces from DeepSeek-R1-0528). RL environments: math (21.2K problems from Skywork-OR1, Acereason-Math, DAPO, ORZ-Hard), code (8.6K Python problems, DeepCoder-inspired, SYNTHETIC-2), science (29.3K from MegaScience), logic (11.6K from SynLogic), deep research (DeepDive: 1K SFT + 2.2K RL samples, Serper web search), software engineering (R2E-Gym and mini-swe-agent-plus scaffolds)[11]
Curriculum / stagingtwo-stage SFT (general chat-and-reasoning at 65K context, then agentic SFT at 98K context) followed by RL with online difficulty filtering (problems sorted into easy/normal/hard pools by observed solve rate, trivial rollouts discarded, easy-pool prompts with pass rate 1 removed); max_off_policy_steps = 8[12]
OptimizerMuon (distributed via an all-to-all-based shard reshuffle for FSDP; Newton-Schulz update computed per rank on gathered gradient shards); weight decay 0.01 in SFT stage 1[13]
LR scheduleSFT stage 1: linear warmup from 1e-8 to 5e-5 over 300 steps; SFT stage 2: 5e-8 with linear decay over 800 steps; RL: Muon learning rate 1e-6[14]
Batch scheduleSFT: ~33M tokens per step, FSDP world size 64 with DP replicate size 8 across 512 GPUs; RL: batch of 256 prompts with 16 rollouts per prompt[14]
Precisionnot disclosed
ParallelismFSDP2 trainer (with tensor/context/expert parallelism support in torchtitan; expert parallelism evaluated and NOT enabled because grouped-GEMM kernels already saturate at the training's sequence length/hidden dim); vLLM inference backend; context parallelism used to reach 98K in agentic SFT; FSDP degree 32 at 48K sequence length; activation offloading to CPU for long sequences[15]
Hardware512 NVIDIA H200 GPUs across 64 nodes (400Gbps NDR InfiniBand, ConnectX-7; Lustre + NVMe-NFS storage; Slurm with Cgroup v2); RL used 60 nodes (16 training / 44 inference at ~1:3)[16]
Disclosed computenot disclosed

Post-training

SFTyes[17]
Preference optimisationother[18]
Reasoning traininglarge-scale RLVR across math, code, science, logic, deep-research (web search) and software-engineering environments with verifiable rewards (rule-based math-verify plus CompassVerifier-7B LLM-judge; sandbox test execution; binary search rewards); long chain-of-thought with an always-reasoning chat template (<|think|> token, reasoning retained across turns); multi-turn agentic tool use (up to 200 turns in SWE scaffolds); online difficulty filtering curriculum[19]
Distillationnone

Modality

Typetext-only
Attachmentn/a[20]

Inference efficiency

KV-cache designnot disclosed
Quantisation shippednone disclosed[21]
Speculative decodingnone disclosed[22]
Serving optimisationsvLLM OpenAI-compatible inference backend used for rollout generation, with custom /update_weights and /reload_weights endpoints for policy updates (§2.1.1); no other serving optimizations disclosed[23]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
AIME 202490.8Table 2
AIME 202588.0Table 2
LiveCodeBench v669.3Table 2
GPQA (Diamond)74.4Table 2
HLE14.6Table 2
MMLU-Pro81.9Table 2

Provenance references

  1. stated · Title
  2. stated · Title page (Prime Intellect Team; correspondence to [email protected])
  3. stated · pipeline
  4. stated · Abstract ('We open-source the model together with the full infrastructure stack'); §1 (footnote 2: huggingface.co/PrimeIntellect/INTELLECT-3); §5 ('By open sourcing INTELLECT-3, the environments, and the complete training framework')
  5. stated · Abstract ('a 106B-parameter Mixture-of-Experts model (12B active)')
  6. stated · Abstract ('12B active')
  7. stated · §2.1.6 ('the activation memory for a sequence length of 48k, hidden size 4096, and 46 decoder layers is Memact = 46 × (48,000 × 4,096) × 2 bytes ≈ 18 GB')
  8. stated · §2.1.6 ('hidden size 4096'), Figure 5 caption ('hidden dim 4096')
  9. stated · Figure 5 caption ('Execution time and TFLOPS of torch._grouped_mm with hidden dim 4096 and MoE dim 1408 on H200 SXM')
  10. stated · §3.3 ('a maximum context length of 65536' during RL), §3.2 (general SFT at 65K context; agentic SFT scaled to 98K context via context parallelism), §2.1.6 (sequence lengths up to 72k with activation offloading)
  11. stated · §3.1, Table 1, §3.2
  12. stated · §3.2, §3.3, §2.1.5
  13. stated · §3.2, §2.1.7
  14. stated · §3.2, §3.3
  15. stated · §2.1.1, §2.1.6, §2.1.8, §3.2
  16. stated · §2.4, §3.3
  17. stated · §3.2 (general chat-and-reasoning SFT stage plus agentic SFT stage; chat template with <|think|> token and Qwen3/GLM-style control tokens; always-reasoning model)
  18. stated · §3.3 (RLVR objective J_IcePop: masked token-level importance sampling with double-sided masking M(k) for k∈[α,β], α=0.5, β=5, group-relative advantages Âi,t = Si − mean({Si}G); early ablations compared GSPO vs CISPO)
  19. stated · §3.1, §3.2, §3.3
  20. n/a · §1, §4 (text-only model)
  21. stated · §A.2 (the report only notes that evaluations avoid quantization loss by using creator APIs where available)
  22. stated · §2 (no speculative decoding mentioned)
  23. stated · §2.1.1