Audited on NVIDIA A100-SXM4-40GB • 2.15M to 2.77M Real Tokens • Zero Synthetic Padding

Multi-Million-Token Intelligence.
Bounded in 2,112 Active Tokens.

Conventional Transformers require 57.3 GiB to 137.3 GiB of Key-Value cache memory alone at 2.15M–2.77M tokens. Powered by the ISOM-R2 Paged Virtual SVD Cache and $SO(d)$ Lie-manifold transport, the four models in the Ael Model Family stream and synthesize across entire open-source repositories in 7.42 to 13.94 seconds with +0.01 GB to +0.09 GB streaming KV overhead.

Explore Audited Hardware Logs Author: Prannessh K. V. A. →

01 • Model Lineup

Four Specialized Architectures. One Bounded State Engine.

Active GPU Attention Window: 2,112 Tokens (64 Sinks + 2,048 Window)

1.54B Dense 28L GQA

Ael-Coder-1.5B

Multi-file PyTorch code synthesis across 82 transformers modules. Synthesizes LoggedGELU across a 1.08M-token inter-file gap with zero numerical error.

Context

2,147,447

Throughput

289.5K/s

Peak VRAM

3.14 GB

Total Time

7.42 s

Open Model Card →

1.54B Dense SO(64) Lie Verifier

Ael-Reasoning-1.5B

Symbolic mathematics & combinatorics across 42 sympy modules. Synthesizes exact DerangedFibonacci recurrences with $O(1)$ Lie-manifold rollback.

Context

2,311,513

Throughput

231.7K/s

Peak VRAM

3.03 GB

Total Time

9.97 s

Open Model Card →

15.71B / 2.36B Act 64-Expert MLA

Ael-Coder-16B-MoE

Repository-scale Mixture-of-Experts with Multi-Head Latent Attention ($SO(192)$). Bridges a 1.39M-token inter-file gap with +0.06 GB streaming KV overhead.

Context

2,774,027

Throughput

227.7K/s

Peak VRAM

30.51 GB

Total Time

12.18 s

Open Model Card →

40.0B Dense 60L 4-Bit NF4

Ael-Pro-40B

Enterprise backend security & cryptographic compliance across 122 django modules. Synthesizes 1.8M-iteration PBKDF2 + HMAC-SHA256 audit classes.

Context

2,399,330

Throughput

172.1K/s

Peak VRAM

24.73 GB

Total Time

13.94 s

Open Model Card →

02 • Hardware Telemetry

Memory Compression & Streaming Speed

Measured on single NVIDIA A100-SXM4-40GB GPU

Total GPU Memory at 2.15M–2.77M Tokens

Ael Audited Peak VRAM (Weights + 2,112 Active KV) vs. Standard Transformer (Weights + Full FP16 KV)

Lower is Better

Ael-Coder-1.5B (2.15M tokens) 19.2× Lower VRAM

Ael (ISOM-R2)

3.14 GB

Standard GQA

60.32 GB

Ael-Reasoning-1.5B-Instruct (2.31M tokens) 21.3× Lower VRAM

Ael (ISOM-R2)

3.03 GB

Standard GQA

64.61 GB

Ael-Coder-16B-MoE (2.77M tokens) +0.06 GB Stream KV

Ael (ISOM-R2)

30.51 GB

Standard MLA

115.05 GB

Ael-Pro-40B (2.40M tokens) 6.4× Lower VRAM

Ael (ISOM-R2)

24.73 GB

Standard MQA

158.91 GB

Audited Telemetry Chart

Switch between Streaming Throughput (tok/s) and Memory Scaling (GB)

Zero OOM Guarantee: Every Ael model completes 2.15M to 2.77M real-token prefill and multi-hop generation well inside a single 40GB A100 GPU, sustaining 172,072 to 289,500 tokens/sec.

03 • Benchmark Matrix

Complete Audited Evaluation Table

Model & Class	Target Domain & Real Corpus	Audited Tokens	Inter-Hop Gap	Time & Speed	Weights / Peak VRAM	Multi-Hop Recall & Live Verification
Ael-Coder-1.5B AelCoder15BForCausalLM • 1.54B	Multi-File PyTorch Synthesis `huggingface/transformers` (82 files)	2,147,447 1,049 chunks	1,077,049 tok 526 chunks	7.42 s 289,500 tok/s	2.98 GB / 3.14 GB +0.01 GB stream	Pages [388, 791, 265] (100%) Synthesized `LoggedGELU` • `max_err = 0.00e+00`
Ael-Reasoning-1.5B-Instruct AelReasoning15BForCausalLM • 1.54B	Symbolic Math & Combinatorics `sympy/sympy` (42 modules)	2,311,513 1,129 chunks	932,802 tok 455 chunks	9.97 s 231,748 tok/s	2.88 GB / 3.03 GB +0.01 GB stream	Pages [501, 956] (100%) Synthesized `DerangedFibonacci` • $SO(64)$ `2.51e-05`
Ael-Coder-16B-MoE AelCoder16BMoEForCausalLM • 15.71B/2.36B	Repository-Scale MoE Synthesis `huggingface/transformers` (82 files)	2,774,027 1,355 chunks	1,394,629 tok 681 chunks	12.18 s 227,686 tok/s	29.28 GB / 30.51 GB +0.06 GB stream	Pages [339, 1020] (100%) Synthesized `LoggedGELU` • $SO(192)$ MLA `4.47e-05`
Ael-Pro-40B AelPro40BForCausalLM • 40.0B NF4	Enterprise Security & Crypto Audit `django/django` (122 modules)	2,399,330 1,172 chunks	1,171,652 tok 572 chunks	13.94 s 172,072 tok/s	21.62 GB / 24.73 GB +0.09 GB stream	Pages [304, 876] (100%) Synthesized `SignedPBKDF2Hasher` • HMAC + Tamper Pass

Model	Real Codebase Corpus	Audited Context	Prefill + Gen Time	Weights / Peak VRAM	Active KV	Retrieved Chunks	Exact Output
Ael-Coder-1.5B	93 Python files (`transformers` + `torch`)	1,052,188 (514 chunks)	6.10 s (172,490 tok/s)	3.09 GB / 3.51 GB	2,112 tok	[250, 249] (GT: 249)	CaptureStd (100%)
Ael-Reasoning-1.5B-Instruct	93 Python files (`transformers` + `torch`)	1,052,188 (514 chunks)	6.07 s (173,342 tok/s)	3.56 GB / 3.98 GB	2,112 tok	[250, 249] (GT: 249)	CaptureStd (100%)
Ael-Coder-16B-MoE	93 Python files (`transformers` + `torch`)	1,089,849 (533 chunks)	15.86 s (68,717 tok/s)	31.49 GB / 32.06 GB	2,112 tok	[258, 257] (GT: 257)	CaptureStd (100%)
Ael-Pro-40B	93 Python files (`transformers` + `torch`)	1,056,780 (516 chunks)	15.39 s (68,670 tok/s)	21.71 GB / 22.84 GB	2,112 tok	[253, 252] (GT: 252)	CaptureStd) (100%)

04 • Audited Execution Logs

Verbatim A100 Hardware Transcripts

Select any of the four benchmarked Ael models to inspect its streaming prefill telemetry, retrieved pages, and live execution verification.

Ael-Coder-1.5B — 2,147,447-Token Multi-File Code Synthesis

82 Python files from huggingface/transformers • Hop 1: CaptureStdout (Chunk 265) • Hop 2: GELUActivation (Chunk 791)

Hugging Face Repo → Notebook (.ipynb)

nvidia-a100-sxm4-40gb — Prannesshkva/Ael-Coder-1.5B

Loaded AelCoder15BForCausalLM | Model Weights VRAM: 2.98 GB
Tokenizing 2.15M+ real-world multi-file corpus (82 files | 9,780,715 chars)...
Hop 1 Target   : class CaptureStdout(CaptureStd) @ Token #543,865 -> Chunk 265
Hop 2 Target   : class GELUActivation(nn.Module) @ Token #1,620,914 -> Chunk 791
Inter-File Gap : 1,077,049 real tokens (526 chunks apart)

  [ISOM-R2 Engine] Streaming 2,147,383 codebase context tokens across 1049 chunks (active GPU buffer < 400 MB)...
  [ISOM-R2] Prefill 428,032 / 2,147,383 tokens (19.9%) | Active buffer: 2112 tokens | VRAM: 2.99 GB
  [ISOM-R2] Prefill 856,064 / 2,147,383 tokens (39.9%) | Active buffer: 2112 tokens | VRAM: 2.99 GB
  [ISOM-R2] Prefill 1,284,096 / 2,147,383 tokens (59.8%) | Active buffer: 2112 tokens | VRAM: 2.99 GB
  [ISOM-R2] Prefill 1,712,128 / 2,147,383 tokens (79.7%) | Active buffer: 2112 tokens | VRAM: 2.99 GB
  [ISOM-R2] Prefill 2,140,160 / 2,147,383 tokens (99.7%) | Active buffer: 2112 tokens | VRAM: 2.99 GB
  [ISOM-R2] Prefill 2,147,383 / 2,147,383 tokens (100.0%) | Active buffer: 2112 tokens | VRAM: 2.99 GB
  [ISOM-R2] Retrieved salient context pages: [388, 791, 265] | Active KV: 2112 tokens

========================================================================================
2.15M-TOKEN MULTI-FILE SYNTHESIS BENCHMARK: Prannesshkva/Ael-Coder-1.5B
========================================================================================
Model Architecture Class  : AelCoder15BForCausalLM
Total Real Source Files   : 82 files (9,780,715 chars)
Total Real Context Tokens : 2,147,447
Distance Between Files    : 1,077,049 tokens (526 chunks apart)
Total Time (Prefill+Gen)  : 7.42 s (289,500 tok/s)
Model Weights VRAM        : 2.98 GB
Peak Total GPU VRAM       : 3.14 GB ( Overhead: +0.16 GB )
Active KV Cache Length    : 2112 tokens
Ground-Truth Chunks       : Hop 1 = Chunk 265 | Hop 2 = Chunk 791
Retrieved Chunk Indices   : [388, 791, 265] (Hop 1 Hit: True | Hop 2 Hit: True)
----------------------------------------------------------------------------------------
MODEL SYNTHESIZED MULTI-FILE CODE:
----------------------------------------------------------------------------------------
class LoggedGELU(GELUActivation):
    def forward(self, x):
        with CaptureStdout(replay=False) as cs:
            out = super().forward(x)
            print(x.shape)
        return (out, cs.out)
----------------------------------------------------------------------------------------
LIVE GPU EXECUTION VERIFICATION OF SYNTHESIZED CLASS:
  • Live Instantiation & Forward Pass : PASSED (Output shape: [1, 4])
  • Captured Stdout via CaptureStdout : 'torch.Size([1, 4])'
  • Numerical Parity vs GELUActivation: max_err = 0.00e+00 (PASSED)
========================================================================================

Ael-Reasoning-1.5B-Instruct — 2,311,513-Token Symbolic Math Benchmark

42 modules from sympy/sympy • Hop 1: subfactorial (Chunk 501) • Hop 2: fibonacci (Chunk 956) • $SO(64)$ Verified

Hugging Face Repo → Notebook (.ipynb)

nvidia-a100-sxm4-40gb — Prannesshkva/Ael-Reasoning-1.5B-Instruct

Loaded AelReasoning15BForCausalLM | Device: CUDA | Weights VRAM: 2.88 GB
Tokenizing 2.31M+ real-world Symbolic Math corpus (42 modules | 6,534,112 chars)...
Hop 1 Target   : class subfactorial(CombinatorialFunction) @ Token #1,026,083 -> Chunk 501
Hop 2 Target   : class fibonacci(DefinedFunction)          @ Token #1,958,885 -> Chunk 956
Inter-Hop Gap  : 932,802 real tokens (455 chunks apart)

  [ISOM-R2 Engine] Streaming 2,311,398 codebase context tokens across 1129 chunks (active GPU buffer < 400 MB)...
  [ISOM-R2] Prefill 460,800 / 2,311,398 tokens (19.9%) | Active buffer: 2112 tokens | VRAM: 2.89 GB
  [ISOM-R2] Prefill 921,600 / 2,311,398 tokens (39.9%) | Active buffer: 2112 tokens | VRAM: 2.89 GB
  [ISOM-R2] Prefill 1,382,400 / 2,311,398 tokens (59.8%) | Active buffer: 2112 tokens | VRAM: 2.89 GB
  [ISOM-R2] Prefill 1,843,200 / 2,311,398 tokens (79.7%) | Active buffer: 2112 tokens | VRAM: 2.89 GB
  [ISOM-R2] Prefill 2,304,000 / 2,311,398 tokens (99.7%) | Active buffer: 2112 tokens | VRAM: 2.89 GB
  [ISOM-R2] Prefill 2,311,398 / 2,311,398 tokens (100.0%) | Active buffer: 2112 tokens | VRAM: 2.89 GB
  [ISOM-R2] Retrieved salient context pages: [501, 956] | Active KV: 2112 tokens

============================================================================================
2.31M-TOKEN SYMBOLIC MATH & MULTI-HOP REASONING BENCHMARK: Prannesshkva/Ael-Reasoning-1.5B-Instruct
============================================================================================
Model Architecture Class     : AelReasoning15BForCausalLM
Real Symbolic Math Modules   : 42 files from sympy/sympy (6,534,112 chars)
Total Real Context Tokens    : 2,311,513
Distance Between Math Hops   : 932,802 tokens (455 chunks apart)
Total Time (Prefill+Deduce)  : 9.97 s (231,748 tok/s)
Model Weights VRAM           : 2.88 GB
Peak Total GPU VRAM          : 3.03 GB ( Overhead: +0.15 GB )
Active KV Cache Length       : 2112 tokens
Ground-Truth Chunks          : Hop 1 (subfactorial) = Chunk 501 | Hop 2 (fibonacci) = Chunk 956
Retrieved Chunk Indices      : [501, 956] (Hop 1 Hit: True | Hop 2 Hit: True)
--------------------------------------------------------------------------------------------
SO(64) LIE-MANIFOLD REASONING & O(1) ROLLBACK AUDIT:
--------------------------------------------------------------------------------------------
Reasoning Trajectory Tokens  : 57 steps x 64-dim SO(64) Lie submanifold
SO(64) Orthogonality Error   : 2.51e-05  (||U_hop^T U_hop - I||_F / sqrt(64))
O(1) Multi-Hop Rollback Err  : 2.68e-05  (||U_hop^T (h_K - Delta_H) - h_0||_2 / ||h_0||_2)
Geodesic Kinetic Drift (k)   : 0.045786
Manifold Confidence Score    : 0.7954  (CyclicManifoldVerifier Valid: True)
--------------------------------------------------------------------------------------------
MULTI-HOP SYMBOLIC MATH DEDUCTION OUTPUT:
--------------------------------------------------------------------------------------------
class DerangedFibonacci:
    @staticmethod
    def evaluate(n):
        fib_value = fibonacci(n).evalf()
        subfactorial_count = subfactorial(int(fib_value)).evalf()
        return (int(fib_value), int(subfactorial_count))
--------------------------------------------------------------------------------------------
LIVE COMBINATORIAL VERIFICATION (n -> (F_n, !F_n)):
  -> n = 3 : Deduced (F_n= 2, !F_n=         1) | SymPy Ground-Truth = (2, 1) | Match: True
  -> n = 4 : Deduced (F_n= 3, !F_n=         2) | SymPy Ground-Truth = (3, 2) | Match: True
  -> n = 5 : Deduced (F_n= 5, !F_n=        44) | SymPy Ground-Truth = (5, 44) | Match: True
  -> n = 6 : Deduced (F_n= 8, !F_n=     14833) | SymPy Ground-Truth = (8, 14833) | Match: True
  -> n = 7 : Deduced (F_n=13, !F_n=2290792932) | SymPy Ground-Truth = (13, 2290792932) | Match: True
  -> Overall Symbolic Math & Manifold Audit : PASSED (100% EXACT)
============================================================================================

Ael-Coder-16B-MoE — 2,774,027-Token Repository-Scale MoE Synthesis

82 files from huggingface/transformers • 1,394,629-token inter-file gap • $SO(192)$ Multi-Head Latent Attention Verified

Hugging Face Repo → Notebook (.ipynb)

nvidia-a100-sxm4-40gb — Prannesshkva/Ael-Coder-16B-MoE

Loaded AelCoder16BMoEForCausalLM (15.71B total / 2.36B active MoE params) | Weights VRAM: 29.28 GB
Tokenizing 2.25M+ real-world multi-file corpus (82 files | 9,779,186 chars)...
Hop 1 Target   : class CaptureStdout(CaptureStd) @ Token #695,058 -> Chunk 339
Hop 2 Target   : class GELUActivation(nn.Module) @ Token #2,089,687 -> Chunk 1020
Inter-File Gap : 1,394,629 real tokens (681 chunks apart)

  [ISOM-R2 Engine] Streaming 2,773,953 codebase context tokens across 1355 chunks (active GPU buffer < 400 MB)...
  [ISOM-R2] Prefill 555,008 / 2,773,953 tokens (20.0%) | Active buffer: 2112 tokens | VRAM: 29.34 GB
  [ISOM-R2] Prefill 1,110,016 / 2,773,953 tokens (40.0%) | Active buffer: 2112 tokens | VRAM: 29.34 GB
  [ISOM-R2] Prefill 1,665,024 / 2,773,953 tokens (60.0%) | Active buffer: 2112 tokens | VRAM: 29.34 GB
  [ISOM-R2] Prefill 2,220,032 / 2,773,953 tokens (80.0%) | Active buffer: 2112 tokens | VRAM: 29.34 GB
  [ISOM-R2] Prefill 2,773,953 / 2,773,953 tokens (100.0%) | Active buffer: 2112 tokens | VRAM: 29.34 GB
  [ISOM-R2] Retrieved salient context pages: [339, 1020] | Active KV: 2112 tokens

============================================================================================
2.77M-TOKEN MULTI-FILE MOE SYNTHESIS BENCHMARK: Prannesshkva/Ael-Coder-16B-MoE
============================================================================================
Total Real Context Tokens : 2,774,027 (82 files | 1355 chunks)
Inter-File Hop Distance   : 1,394,629 tokens (Chunk 339 <-> Chunk 1020)
Total Time (Prefill+Gen)  : 12.18 s (227,686 tok/s)
Model Weights VRAM        : 29.28 GB
Peak Total GPU VRAM       : 30.51 GB (KV + Activation Overhead: +1.23 GB)
Active KV Cache Length    : 2112 tokens
Retrieved Chunk Indices   : [339, 1020]
Multi-Hop Recall          : Hop 1 Hit: True | Hop 2 Hit: True
SO(192) MLA Orthogonality : 4.47e-05 ||R^T R - I||_F  |  O(1) Rollback Error: 3.06e-06
--------------------------------------------------------------------------------------------
MODEL SYNTHESIZED MULTI-FILE CODE:
--------------------------------------------------------------------------------------------
class LoggedGELU(GELUActivation):
    def forward(self, x):
        with CaptureStdout(replay=False) as cs:
            out = super().forward(x)
            print(x.shape)
        return (out, cs.out)
--------------------------------------------------------------------------------------------
LIVE GPU EXECUTION VERIFICATION OF SYNTHESIZED CLASS:
  • Live Instantiation & Forward Pass : PASSED (Output shape: [1, 4])
  • Captured Stdout via CaptureStdout : 'torch.Size([1, 4])'
  • Numerical Parity vs GELUActivation: max_err = 0.00e+00 (PASSED)
============================================================================================

Ael-Pro-40B — 2,399,330-Token Enterprise Security & Audit Benchmark

122 Django modules • Hop 1: Signer (Chunk 304) • Hop 2: PBKDF2PasswordHasher (Chunk 876) • 1,171,652-token gap

Hugging Face Repo →

nvidia-a100-sxm4-40gb — Prannesshkva/Ael-Pro-40B

Loaded AelPro40BForCausalLM (40.0B Dense | 60 Layers | 128 Q / 8 KV Heads) | Weights VRAM: 21.62 GB
Tokenizing 2.3M+ Enterprise Security & ORM corpus (122 files | 8,973,788 chars)...
Hop 1 Target   : class Signer (HMAC-SHA256 Signing)         @ Token #623,966 -> Chunk 304
Hop 2 Target   : class PBKDF2PasswordHasher(BasePasswordHasher) @ Token #1,795,618 -> Chunk 876
Inter-Hop Gap  : 1,171,652 real tokens (572 chunks apart)

  [ISOM-R2 Engine] Streaming 2,399,185 codebase context tokens across 1172 chunks (active GPU buffer < 400 MB)...
  [ISOM-R2] Prefill 479,232 / 2,399,185 tokens (20.0%) | Active buffer: 2112 tokens | VRAM: 21.71 GB
  [ISOM-R2] Prefill 958,464 / 2,399,185 tokens (39.9%) | Active buffer: 2112 tokens | VRAM: 21.71 GB
  [ISOM-R2] Prefill 1,437,696 / 2,399,185 tokens (59.9%) | Active buffer: 2112 tokens | VRAM: 21.71 GB
  [ISOM-R2] Prefill 1,916,928 / 2,399,185 tokens (79.9%) | Active buffer: 2112 tokens | VRAM: 21.71 GB
  [ISOM-R2] Prefill 2,396,160 / 2,399,185 tokens (99.9%) | Active buffer: 2112 tokens | VRAM: 21.71 GB
  [ISOM-R2] Prefill 2,399,185 / 2,399,185 tokens (100.0%) | Active buffer: 2112 tokens | VRAM: 21.71 GB
  [ISOM-R2] Retrieved salient context pages: [304, 876] | Active KV: 2112 tokens

============================================================================================
2.3M+ TOKEN ENTERPRISE SECURITY & AUDIT BENCHMARK: Prannesshkva/Ael-Pro-40B
============================================================================================
Model Architecture Class  : AelPro40BForCausalLM (40.0B Dense | 4-Bit NF4 | 60 Layers)
Enterprise Corpus         : 122 real Django modules (8,973,788 chars)
Total Real Context Tokens : 2,399,330
Inter-Module Hop Distance : 1,171,652 tokens (572 chunks apart)
Total Time (Prefill+Gen)  : 13.94 s (172,072 tok/s)
Model Weights VRAM        : 21.62 GB
Peak Total GPU VRAM       : 24.73 GB ( Overhead: +3.11 GB )
Active KV Cache Length    : 2112 tokens
Ground-Truth Chunks       : Hop 1 = Chunk 304 | Hop 2 = Chunk 876
Retrieved Chunk Indices   : [304, 876] (Hop 1 Hit: True | Hop 2 Hit: True)
SO(64) Lie Orthogonality  : ||A^T A - I||_F = 3.98e-06 | O(1) Rollback Err = 5.18e-07
--------------------------------------------------------------------------------------------
MODEL SYNTHESIZED ENTERPRISE CRYPTOGRAPHIC AUDIT CLASS:
--------------------------------------------------------------------------------------------
class SignedPBKDF2Hasher(PBKDF2PasswordHasher):
    @staticmethod
    def issue_audit_token(password, salt, key):
        encoded = PBKDF2PasswordHasher().encode(password, salt)
        signed_token = Signer(key=key).sign(encoded)
        verified = PBKDF2PasswordHasher().verify(password, Signer(key=key).unsign(signed_token))
        return (signed_token, verified)
--------------------------------------------------------------------------------------------
LIVE ENTERPRISE SECURITY & COMPLIANCE VERIFICATION:
  • Inheritance Check         : issubclass(SignedPBKDF2Hasher, PBKDF2PasswordHasher) = True
  • Signed PBKDF2 Audit Token : pbkdf2_sha256$1800000$ael_salt_99$l4kux5SXVPIiDs...vnWtZFJzaZm19ZgFflq6inYo
  • Round-Trip HMAC + PBKDF2  : verified=True (PASSED)
  • 1-Byte Tamper Rejection   : BadSignature raised=True (PASSED)
============================================================================================

05 • Interactive Simulator

Multi-Million-Token Memory Scaling Calculator

Simulate KV-cache memory growth from 100K to 4,000,000 tokens across the four Ael models.

Select Ael Model

Sequence Length 2,400,000 Tokens

100K 1.0M 2.0M 3.0M 4.0M

Base Model Weights VRAM: 21.62 GB

Standard Attention KV Cache ($O(N)$): 137.33 GB

Ael ISOM-R2 Active KV (2,112 tok): +0.09 GB

Standard Transformer

Weights + Unbounded FP16 KV Cache

158.95 GB

Single A100-40GB: CUDA OOM (>40 GB)

Ael (ISOM-R2 Engine)

Weights + Bounded 2,112-Token Buffer

21.71 GB

KV Memory Saved: 99.93% Saved

01 • Paged Virtual SVD Cache

Constant 2,112-Token Active Buffer

Incoming repository streams are partitioned into 2,048-token chunks. Evicted pages are compressed into low-rank spectral bases while the active GPU KV cache remains strictly bounded at $64 + 2048 = 2112$ tokens.

02 • Cayley Lie-Group Transport

Exact $SO(d)$ Isometry & $O(1)$ Rollback

Recurrent state transitions evolve on the Special Orthogonal Lie manifold via skew-symmetric generators $A_t = -A_t^\top$: $$U_t = (I - \tfrac{1}{2}A_t)^{-1}(I + \tfrac{1}{2}A_t) \in SO(d)$$ preserving norm and enabling $O(1)$ multi-hop state inversion.

03 • Multi-Hop Synthesis

Entity-Balanced Page Recall

Each chunk is indexed with structural class/function signatures and per-entity inverse-document-frequency weights, achieving 100% recall across 1.39M-token inter-module distances.

Multi-Million-Token Intelligence. Bounded in 2,112 Active Tokens.

Four Specialized Architectures. One Bounded State Engine.

Ael-Coder-1.5B

Ael-Reasoning-1.5B

Ael-Coder-16B-MoE

Ael-Pro-40B

Memory Compression & Streaming Speed

Total GPU Memory at 2.15M–2.77M Tokens

Audited Telemetry Chart

Complete Audited Evaluation Table

Verbatim A100 Hardware Transcripts

Ael-Coder-1.5B — 2,147,447-Token Multi-File Code Synthesis

Ael-Reasoning-1.5B-Instruct — 2,311,513-Token Symbolic Math Benchmark

Ael-Coder-16B-MoE — 2,774,027-Token Repository-Scale MoE Synthesis

Ael-Pro-40B — 2,399,330-Token Enterprise Security & Audit Benchmark

Multi-Million-Token Memory Scaling Calculator

Constant 2,112-Token Active Buffer

Exact $SO(d)$ Isometry & $O(1)$ Rollback

Entity-Balanced Page Recall

Multi-Million-Token Intelligence.
Bounded in 2,112 Active Tokens.