"Frontier models need a datacenter GPU" rests on a hidden assumption: that the model reads ALL its parameters every token. Decode is memory-bandwidth bound — sweep 34B params/token and an 8 GB card dies at 1–2 tok/s.
So we ran ONE 34.7B reasoning model — Ourbox-35B-JGOS, a sparse Mixture-of-Experts — as the identical weights across the whole hardware spectrum. All measured:
Why it works: Ourbox holds 34.7B params but only ~3B are active per token (256 experts, top-8). Since decode is bandwidth-bound, a dense 34B moves ~16.7 GB/token while Ourbox moves ~1.45 GB — ~11× less traffic. Put the experts in system RAM, keep attention/router/shared on the GPU, and a 34.7B reasoner runs on an 8 GB laptop — or no GPU at all.
Sparsity alone, proven (same laptop, same quant, ~same footprint): Ourbox-35B (A3B) 20.01 tok/s vs Qwen2.5-32B (dense) 5.36 → 3.7× from sparsity alone, ~2× the best dense-32B on any 8 GB machine. Not a toy: GPQA Diamond 86.4% (maj@8).
Try it live (same prompt, GPU vs GPU-less CPU, live tok/s). Honest scope: one machine's measurements; the CPU path proves it RUNS without a GPU, not that it beats one.
🔓 We ran genuine quantum key-recovery on 'real IBM quantum hardware' — and pushed the frontier well past the largest hardware demos we're aware of (which sat at N=4).
Using Simon's algorithm on ibm_kingston, we recovered the secret key of two symmetric-cipher structures: • Even–Mansour — N=5 → N=10 • 3-round Feistel (DES-family) — block 6 → 8
Each verified against an 'independent control key', using error mitigation only (no QEC).
🧭 Honest scope: this is not a quantum speedup (the effective difficulty tracks the classical birthday bound ~2^{n/2}), not a break of real AES/RSA, and not 16-round DES (ours is 3-round). The recovery method is reserved for a forthcoming paper; formal record status is pending peer review.
AI is usually framed as "how smart is the model / how many GPUs did you buy." The real bottleneck is elsewhere — how efficiently you use the GPUs you already have.
Training happens once; inference runs the entire time users use your product. So a service's economics come down to cost per token. Inference acceleration uses software to pull several times more out of the same GPU — the effect of plugging in one more "virtual GPU."
VIDRAFT's VKAE, measured (B200, same-harness, no quality loss):
Qwen3.5-35B-A3B (MoE): 25.7 → 601 tok/s (23.4×) Darwin-36B-Opus (in-house MoE): 25.0 → 280.8 (11.2×) 10,000+ tok/s peak aggregate under concurrency The key: it's reproducible — model + serving shipped as one container.
docker pull vidraft/qwen35-vkae:601 Don't take our word for it — run it yourself. The mechanism will be released as a paper.
🐯 Chitos — The Security Scanner That Actually Proves It
Most security scanners hand you a suspect list and walk away. That gap between detection and proof is where attackers live — and it's exactly the gap that Chitos was built to close.
Chitos is the successor to Mythos, a static analyzer built for quick code health checks. Mythos was good at pattern matching — spotting dangerous sinks, mapping CWEs, producing readable reports. But static analysis has a structural ceiling. A rule that sees eval(user_input) can tell you that looks dangerous. It cannot tell you whether the input is reachable, whether sanitization three layers up covers this path, or whether there's a live exploit chain for your exact framework version. Chitos was built to answer those questions.
🔍 Phase 1 applies 50 language-agnostic rules across Python, JavaScript, Go, Java, C/C++, Rust, PHP, YAML and more — covering injection sinks, deserialization gadgets, credential leakage, broken crypto, and prototype pollution. Every candidate is re-verified before reaching the report. Findings that can't be substantiated are excluded, not handed to you as noise.
🔬 Phase 2 dispatches an autonomous web-search agent to hunt live CVE databases, exploit advisories, and public PoC repositories. It formulates hypotheses, verifies them, and synthesizes a structured threat narrative. This phase needs a user-supplied Claude API key — Phases 1 and 3 run entirely free.
🎯 Phase 3 is where Chitos diverges from everything else. Against targets you own or are authorized to test, it fires real payloads — XSS, SQLi, path traversal, command injection — mutates on block, captures hard evidence, and connects every proven finding into a kill-chain showing which vulnerabilities to remediate first.
No installation. No account. No code sent to third-party APIs.
Darwin V9 — GPQA Diamond 90.9%, #1 on the leaderboard, with pure greedy decoding Darwin-398B-JGOS reaches 90.9% (180/198) on GPQA Diamond, the PhD-level scientific reasoning benchmark, ranking #1 on the Hugging Face GPQA Diamond leaderboard. No self-consistency, no test-time compute scaling — this was achieved with a single greedy decode (temperature 0, single sample, max 16,384 tokens). The full eval config is published in the model card, so anyone can reproduce it. Raw reasoning, no score inflation. The result comes from Darwin V9, a patented evolutionary model-development platform. Its core idea: it never trains a model from scratch. Why Darwin V9 beats training from scratch
Cost & speed: no trillion-token pretraining run, no months of compute — a purpose-built, high-performance model is produced in a fraction of the time. Reuse of proven intelligence: instead of re-learning every capability from a blank slate, it selects and combines only the strengths of already-trained, already-validated models, so results are stable and predictable. Surgical transplantation: it identifies which neural region of which model holds which capability — at the FFN (Feed Forward Network) layer level — and grafts in only the segments that contribute to the target skill.
How it works: a large model (Qwen 3.5 397B) serves as the mother model (the substrate); several father models specialized in reasoning, coding, and language are analyzed layer-by-layer across their FFN regions; the segments that contribute to the target performance are extracted and transplanted into the mother model to produce a new child model. The result is a ~400B MoE that activates only ~17B parameters per token at inference — large-model capacity with efficient inference. If training from scratch means rebuilding everything from a blank page, Darwin V9 means precisely recombining intelligence that has already been proven. GPQA Diamond #1 is the proof. Model: FINAL-Bench/Darwin-398B-JGOS
🚀 Introducing FINAL-Bench Quantum — an open, neutral benchmark that finally puts quantum-computing methods on one fair yardstick.
Quantum results are notoriously hard to compare. The same "logical error rate" or "query fidelity" means very different things depending on the code, noise model, hardware, and shot count. FINAL-Bench Quantum fixes that: five events judged under identical, published protocols, where every number is labeled as either measured here or quoted from a source.
Five events: ① QEC Decoder ② Optimization (Max-Cut) ③ VQE ④ QRAM ⑤ Quantum Simulation
The rules are simple and strict: ✅ Track A (measured here, with 95% confidence intervals) is kept separate from Track B (quoted from papers, not directly comparable). 🔬 Simulation and real hardware are clearly distinguished, and no quantum-advantage claims are made. 🌍 Methods from Google, IBM, NVIDIA, USTC, Riverlane and more sit side by side, with origin flags and author credits. 📤 Anyone can submit their own method via the Submit tab for review and listing.
Already on the board: real IBM Heron r2 measurements (repetition-code distance boundary, 29–175× error reduction from d3 to d5), a real-chip QRAM query fidelity of 0.92, and H₂ VQE at chemical accuracy — always labeled honestly as simulation vs hardware.
A leaderboard is only useful if you can trust it, so neutrality is the whole point: strong competitors stay in even when they beat the host, sources are quoted faithfully, and a simulation is never rounded up into a hardware claim.
Darwin-60B-DUO: Two SOTAs, One Endpoint — 88.38% on GPQA Diamond 🚀
We're excited to release Darwin-60B-DUO, the Darwin family's first DUO model. Take two domain-verified specialists, hide them behind a single OpenAI-compatible endpoint, and let a router decide which one (or both) answers. You see one model, one API — but get the best of both.
The number that matters: on the full 198-question GPQA Diamond, Darwin-60B-DUO hits 88.38%. The constituents alone land at 69.70% (Darwin-28B-REASON) and 77.27% (AWAXIS-Think-31B); a naive cascade only reaches 83.84%. The DUO clears them all. Two small specialists, intelligently routed, beat one big generalist on cost and quality. Both are independently verified — Darwin-28B-REASON is #3 on the HF GPQA Diamond leaderboard, AWAXIS-Think-31B is #1 on Korea's national K-AI Leaderboard (MSIT).
The brains is a Hybrid-A router picking one of five strategies on the fly. Korean → AWAXIS, English/STEM → Darwin (single-backend, ~70% of traffic at 1× cost). When a Korean answer needs rigorous English reasoning, split_refine fires — Darwin drafts, AWAXIS polishes; MCQ/short-answer runs both with self-consistency + cross-verify. Net effective cost: only ~1.3× a single 30B model.
The part the community will care about: the gateway is model-agnostic and Apache-2.0. Point it at any two OpenAI-compatible backends and you've got a DUO in minutes — teach router.py when to use which, and parallel calls, response merging, and routing transparency via _duo_route are handled for you. Fork it and tell us what you built.
Painless deploy: docker compose up for both vLLM backends + gateway; FP8 ~30GB colocates on a single B200/H100. One git clone (~120GB). Text-only for now, streaming in v1.1. Two SOTAs, one endpoint. Come build your own on the Community tab.
🧬 Darwin Family: Zero Gradient Steps, GPQA Diamond 88.89%
How far can we push LLM reasoning *without* training?
Our team at VIDRAFT submitted this paper to Daily Papers yesterday, and it's currently #3. Huge thanks to everyone who upvoted — sharing the core ideas below.
Darwin Family is a training-free evolutionary merging framework. By recombining the weight spaces of existing LLM checkpoints — with zero gradient-based training — it reaches frontier-level reasoning.
- 🏆 Darwin-28B-Opus: GPQA Diamond 88.89% - 💸 Zero gradient steps — not a single B200 or H200 hour needed - 🧬 Consistent gains across 4B → 35B scale - 🔀 Cross-architecture breeding between Transformer and Mamba families - 🔁 Stable recursive multi-generation evolution
#Three Core Mechanisms
① 14-dim Adaptive Merge Genome — fine-grained recombination at both component level (Attention / FFN / MLP / LayerNorm / Embedding) and block level, expanding the prior evolutionary-merge search space.
② MRI-Trust Fusion — we diagnose each layer's reasoning contribution via an **MRI (Model Reasoning Importance)** signal and fuse it with evolutionary search through a **learnable trust parameter**. Trust the diagnostic too much and search collapses; ignore it and search becomes inefficient — Darwin learns the balance from data.
③ Architecture Mapper — weight-space breeding across heterogeneous families. Attention × SSM crossover actually works.
Why It Matters > Diagnose latent capabilities already encoded in open checkpoints, > and recombine them — no gradients required.
This Space is a fork of the brilliant Eliahu/Model-Atlas, the official demo of "Charting and Navigating Hugging Face's Model Atlas" (Horwitz et al., arXiv 2503.10633). Their pre-computed HF model graph is the foundation of every node and edge you see, and we are deeply grateful for its open release.
The original atlas is a static snapshot of early 2025. Model Galaxy turns it into a living, multimodal map. We injected the 2026 trending originals that did not exist when the atlas was frozen — DeepSeek-V4, Hy3-preview, GLM-5.1, Kimi-K2, gpt-oss, Nemotron-3 Super / Nano / Omni, Hermes-4.3, Qwen3-Coder-Next, Llama-3.3, Granite-4.1, plus the latest multimodal releases (FLUX.2, ERNIE-Image, HunyuanImage / Video, LTX-2.3, Wan2.2, Kokoro-82M, VoxCPM2, Voxtral-TTS, whisper-v3-turbo, Gemma-4, Qwen3-Omni, Phi-4-mm) — each with proper base_model lineage edges.
We also added the complete VIDRAFT Darwin family ontology: 120 nodes covering Darwin Core, AETHER, every brand variant (Rogue, AWAXIS, TenOS, Warecube), NOESIS-Darwin multimodal extensions, and 40+ community quantizations — the most complete Darwin lineage view anywhere.
The name "Galaxy" is now literal: our three injected clusters are re-laid out as logarithmic spiral galaxies, with bigger models near the bright cores and quantizations scattering to the outer arms — just like real star mass distribution. A top-right toggle switches between Galaxy mode (deep-space gradient with 220 animated stars) and Atlas mode (clean white panels for reports). A 15-second progress bar narrates the render, and per-modality / per-company colors make every cluster legible at a glance.
Final scale: 22,480 nodes in the default Modalities atlas, 137,324 in the Large NLP atlas, and a 277-node compact Darwin + Trending view for instant exploration. Feedback and PRs welcome.
We're thrilled to release Darwin-9B-NEG, a 9B-parameter reasoning model that embeds an architecturally-internalised sense of self-confidence directly into the transformer — our proprietary Native Entropy Gating (NEG) technology.
With only 9 billion parameters and 1× inference cost, Pure NEG jumps +12.63 %p over the same model without NEG. Going all-in with ensemble refinement pushes it to 84.34 % — surpassing the published Qwen3.5-9B leaderboard score (81.7 %) by +2.64 %p.
🔬 What makes NEG different from Multi-Turn Iteration (MTI)?
Classical MTI needs 3-8× extra inference passes. NEG instead lives INSIDE the single decoding loop. Two tiny modules ride with the transformer: NEG-Head predicts per-token entropy from the last hidden state, and NEG-Gate conditionally restricts the top-k choice when confidence is low. The gate activates in only 4.36 % of tokens — essentially free at inference time.
✨ Key differentiators • Architecturally internalised — model file *is* the feature • 1× inference cost (vs. 3-8× for MTI) • Drop-in with vLLM / SGLang / TGI / transformers — no extra engine • +12.63 %p reasoning at zero latency overhead • Single-file deployment, Apache 2.0 licensed
Darwin-TTS: 3% of an LLM's Brain Makes TTS Speak with Emotion — Zero Training
We blended 3% of Qwen3-1.7B (LLM) FFN weights into Qwen3-TTS-1.7B's talker module. The result: emotionally enhanced speech synthesis — with zero training, zero data, and zero GPU hours.
Qwen3-1.7B (LLM) and Qwen3-TTS-1.7B's talker share 100% identical architecture — same hidden_size (2048), same layers (28), same heads (16). This enabled pure 1:1 weight blending across 84 FFN tensors with a single lerp operation. At 3% blend, emotion appears. At 5%, emotion intensifies. At 10%, the model breaks — producing 655-second outputs for a 3-second sentence, because the LLM's "keep generating" pattern overwhelms the TTS stop signal.
To our knowledge, this is the first training-free cross-modal weight transfer between an LLM and a TTS model. Prior work either requires adapter training (SmolTolk, 2025), fine-tuning (CSLM, 2025), or massive end-to-end compute (GPT-4o). Darwin-TTS achieves cross-modal capability transfer in under 2 minutes on CPU.
The key insight: TTS models with LLM backbones already "think" in language. We're just restoring 3% of the original LLM's language understanding patterns — particularly those related to emotional semantics and prosody planning. The code is three lines: load the model, load the LLM FFN, call p.lerp_(llm_weight, 0.03).
creators of the Darwin Evolutionary Merge Framework. Darwin LLM V7 achieved GPQA Diamond 86.9% (HF Benchmark #3) through CMA-ES optimized FFN crossbreeding. Darwin-TTS extends this principle from LLM-to-LLM merging into cross-modal LLM-to-TTS transfer. Apache 2.0.
🧬 Darwin-27B-Opus: 86.9% on GPQA Diamond — World #5, Zero Training We are excited to share Darwin-27B-Opus, a 27B model that achieved 86.9% on GPQA Diamond — ranking #5 globally on the HuggingFace leaderboard — without a single gradient update.
How? Darwin breeds pretrained models through evolutionary FFN crossbreeding. The father (Qwen3.5-27B) provides the reasoning architecture; the mother (Claude 4.6 Opus Reasoning Distilled) contributes structured chain-of-thought knowledge. CMA-ES automatically discovers optimal per-layer blending ratios — no human tuning required.
The result surpasses the original Qwen3.5-27B (85.5%), GLM-5.1 (744B, 86.2%), and Qwen3.5-122B (86.6%). A 27B model outperforming 744B — with zero training, zero data, one GPU, ~2 hours.
We also confirmed hybrid vigor on Korean benchmarks: Darwin-27B-KR (2nd generation offspring) surpassed both parents on CLIcK, winning 7 out of 11 categories. The evolutionary optimizer independently assigned 93% of FFN from the Korean-specialized mother while preserving 93% of attention from the reasoning-specialized father — autonomously validating our core principle: FFN carries knowledge, Attention carries reasoning.
📊 Public release: 10 days → 300+ community derivatives, 120K+ downloads.
We're releasing Darwin-4B-David, the first second-generation model in the Darwin Opus family. By evolving an already-evolved model, it achieves 85.0% on GPQA Diamond — surpassing its 58.6% original ancestor and even gemma-4-31B (84.3%) — with just 4.5B parameters.
Second-Generation Evolution Most merges start from a base model and produce a single offspring. Darwin-4B-David breaks this pattern. The Father (Darwin-4B-Opus) was already evolved from gemma-4-E4B-it with Claude Opus reasoning distillation — a Gen-1 model. The Mother (DavidAU's DECKARD-Expresso-Universe) brings Unsloth deep tuning across 5 in-house datasets with thinking mode by default. Crossbreeding these two produced the first Gen-2 Darwin model.
Darwin V6's Model MRI scanned both parents across all 42 layers, assigning independent optimal ratios per layer. The Mother's creativity and Korean language hotspot (Layer 22-25, weight 0.95) was maximally absorbed, while the Father's reasoning core (Layer 30-40, weight 0.48) was preserved. This is "Merge = Evolve" applied recursively — evolution of evolution.
Benchmarks Darwin-4B-David scores 85.0% on GPQA Diamond (+26.4%p over original 58.6%), evaluated generatively with maj@8 (8 generations per question, majority vote), Epoch AI prompt format, thinking mode enabled, 50 sampled questions. On ARC-Challenge (25-shot, loglikelihood), both score 64.93% — expected, as loglikelihood doesn't capture thinking-mode reasoning differences.
Why This Matters gemma-4-31B (30.7B) scores 84.3%. Darwin-4B-David surpasses it at 1/7th the size — no training, no RL, just 45 minutes of MRI-guided DARE-TIES on one H100. The name "David" honors Mother creator DavidAU and evokes David vs. Goliath.
🧬 Darwin V6: Diagnostic-Guided Evolutionary Model Merging
We are releasing Darwin-31B-Opus — a reasoning-enhanced model merging Google's Gemma-4-31B-it and TeichAI's Claude Opus Distill using the Darwin V6 engine.
Conventional merging tools (mergekit, etc.) apply a single ratio to all tensors. Set ratio=0.5 and all 1,188 tensors blend identically, with no distinction between which tensors matter for reasoning versus coding.
Darwin V6 diagnoses both parents at the tensor level before merging. It measures Shannon entropy, standard deviation, and L2 norm for every tensor, then passes 5 diagnostic probes (REASONING, CODE, MATH, KNOWLEDGE, LANGUAGE) through the model to determine layer-wise functional importance. Each of the 1,188 tensors receives an independent optimal ratio.
combined = static(entropy/std/norm) x 0.4 + probe(cosine_distance) x 0.6 final_ratio = mri_ratio x mri_trust + genome_ratio x (1 - mri_trust)
When one parent is overwhelmingly superior for a tensor (ratio < 0.15 or > 0.85), Darwin transplants it directly without interpolation. The mri_trust parameter itself is optimized by CMA-ES evolutionary search, so optimal transplant intensity is determined automatically. After merging, a Health Check compares the child against both parents layer-by-layer to detect interference or function loss.
👉 Try it now: FINAL-Bench/Gemma-4-Multi Two Models, One Space Switch between both Gemma 4 variants in a single interface:
⚡ Gemma 4 26B-A4B — MoE with 128 experts, only 3.8B active params. 95% of the 31B's quality at ~8x faster inference. AIME 88.3%, GPQA 82.3%. 🏆 Gemma 4 31B — Dense 30.7B. Best quality among Gemma 4 family. AIME 89.2%, GPQA 84.3%, Codeforces 2150. Arena open-model top 3.
Features
Vision — Upload images for analysis, OCR, chart reading, document parsing Thinking Mode — Toggle chain-of-thought reasoning with Gemma 4's native <|channel> thinking tokens System Prompts — 6 presets (General, Code, Math, Creative, Translate, Research) or write your own Streaming — Real-time token-by-token response via ZeroGPU Apache 2.0 — Fully open, no restrictions
Technical Details Built with the dev build of transformers (5.5.0.dev0) for full Gemma 4 support including multimodal apply_chat_template, variable-resolution image processing, and native thinking mode. Runs on HF ZeroGPU with @spaces.GPU — no dedicated GPU needed. Both models support 256K context window and 140+ languages out of the box.
🧬 Darwin-35B-A3B-Opus — The Child That Surpassed Both Parents
What if a merged model could beat both its parents? We proved it can. Darwin-35B-A3B-Opus is a 35B MoE model (3B active) built with our Darwin V5 engine — the first evolution system that CT-scans parent models before merging them. 🤗 Model: FINAL-Bench/Darwin-35B-A3B-Opus
The result speaks for itself: GPQA Diamond 90.0%, versus Father (Qwen3.5-35B-A3B) at 84.2% and Mother (Claude 4.6 Opus Distilled) at 85.0%. That's +6.9% over Father and +5.9% over Mother. Not a tradeoff — a genuine leap. Meanwhile, MMMLU sits at 85.0% (Father: 85.2%), multimodal is fully intact, and all 201 languages are preserved.
How? Model MRI changed everything. Traditional merging is guesswork. Darwin V4 added evolution. Darwin V5 added X-ray vision. Model MRI scans each parent layer by layer and discovers: Mother's L34–L38 is the reasoning engine (peak cosine distance), 50–65% of Mother's experts are dead (killed by text-only distillation), and Father is a healthy generalist with every expert alive. The prescription: transplant Mother's reasoning brain at L38 (90% weight), replace her dead experts with Father's living ones, and let Father's router handle the output layer. Reasoning went up. Versatility stayed intact. No tradeoff — just evolution.
35B total, 3B active (MoE) · GPQA Diamond 90.0% · MMMLU 85.0% (201 languages) · Multimodal Image & Video · 262K native context · 147.8 tok/s on H100 · Runs on a single RTX 4090 (Q4) · Apache 2.0 Darwin V5's full algorithm and technical details will be released alongside an upcoming paper.
🌍 World Model Bench — does your world model actually think?
FID measures realism. FVD measures smoothness. But neither tells you whether the model understood the scene.
We just released WM Bench — the first benchmark for cognitive intelligence in world models. The core question: when a beast charges from 3 meters away, does the model know to sprint — not walk? Does it respond differently to a human vs an animal? Does it remember the left corridor was blocked two steps ago?
Those are cognitive questions. No existing benchmark asks them. So we built one.
- 👁 P1 Perception (25%) — Can it read the scene? - 🧠 P2 Cognition (45%) — Does it predict threats, escalate emotions, utilize memory? - 🔥 P3 Embodiment (30%) — Does the body respond with the right motion?
All evaluation is via simple JSON I/O — no 3D engine, no special hardware. Any model with an API can participate.
We also built PROMETHEUS as a live reference implementation — runs in your browser on a T4, no install needed. Combines FloodDiffusion motion generation with a LLM cognitive brain (Perceive → Predict → Decide → Act). Scored 726/1000 (Grade B) on Track C — the only directly verified model so far. Submissions from other teams very welcome.
🏟️ Smol AI WorldCup: A 4B Model Just Beat 8B — Here's the Data
We evaluated 18 small language models from 12 makers on 125 questions across 7 languages. The results challenge the assumption that bigger is always better.
→ A 1.3B model fabricates confident fake content 80% of the time when prompted with nonexistent entities. Qwen3 family hits 100% trap detection across all sizes.
→ Qwen3-1.7B (1.2GB) outscores Mistral-7B, Llama-3.1-8B, and DeepSeek-R1-14B. Latest architecture at 1.7B beats older architecture at 14B.
What makes this benchmark different?
Most benchmarks ask "how smart?" — we measure five axes simultaneously: Size, Honesty, Intelligence, Fast, Thrift (SHIFT). Our ranking metric WCS = sqrt(SHIFT x PIR_norm) rewards models that are both high-quality AND efficient. Smart but massive? Low rank. Tiny but poor? Also low.
We evaluated 9 SOTA models (GPT-5.2, Claude Opus 4.6, Gemini 3 Pro, etc.) across 1,800 assessments in FINAL Bench and found a 39.2%p gap between "recognizing potential errors (MA=0.694)" and "actually finding and fixing them (ER=0.302)."
MARL (Model-Agnostic Runtime Middleware for LLMs) was built to close this metacognitive gap. It decomposes a single LLM call into a 5-stage expert pipeline (Hypothesis → Solver → Auditor → Adversarial Verifier → Synthesizer), transforming "answer in one shot" into "think, doubt, correct, and rewrite."
No weight modification — works instantly with GPT-5.4, Claude, Gemini, Llama, or any OpenAI API-compatible LLM by changing one line: base_url. Ships with 9 domain-specific emergence engines (invention, pharma, genomics, chemistry, ecology, law, and more — 5,538 expert data items) activated by a simple tag like model="gpt-5.4::pharma".
pip install marl-middleware
MARL is also officially registered on ClawHub, the skill marketplace of OpenClaw — an AI agent platform with 260K+ developers and 3,200+ skills. It's the first middleware in the Reasoning Enhancement category. One command — clawhub install marl-middleware — gives your AI agent a metacognition upgrade.
The AI benchmark ecosystem has three structural problems. Major benchmarks like MMLU have surpassed 90%, losing discriminative power. Most leaderboards publish unverified self-reported scores — our cross-verification found Claude Opus 4.6's ARC-AGI-2 listed as 37.6% (actual: 68.8%), Gemini 3.1 Pro as 88.1% (actual: 77.1%). OpenAI's own audit confirmed 59.4% of SWE-bench Verified tasks are defective, yet it remains widely used.
ALL Bench addresses this by comparing 91 models across 6 modalities (LLM · VLM · Agent · Image · Video · Music) with 3-tier confidence badges (✓✓ cross-verified · ✓ single-source · ~ self-reported). Composite scoring uses a 5-Axis Framework and replaces SWE-Verified with contamination-resistant LiveCodeBench.
Key finding: metacognition is the largest blind spot. FINAL Bench shows Error Recovery explains 94.8% of self-correction variance, yet only 9 of 42 models are even measured. The 9.2-point spread (Kimi K2.5: 68.71 → rank 9: 59.5) is 3× the GPQA top-model spread, suggesting metacognition may be the single biggest differentiator among frontier models today.
VLM cross-verification revealed rank reversals — Claude Opus 4.6 leads MMMU-Pro (85.1%) while Gemini 3 Flash leads MMMU (87.6%), producing contradictory rankings between the two benchmarks.
