An autoregressive model must not let position t depend on anything after t. Everyone checks this by inspecting the causal mask โ but hybrid stacks now mix attention with state-space scans, and a scan has no mask. Every mask can be correct while information leaks through scans, aggregations, or normalization.
โ๏ธ So we test the property directly. Two inputs identical except at the last position, two forward passes, compare each layer's prefix, report the first layer that moves. No training, no gradients, no accelerator โ seconds on CPU.
๐ Across 192 injected faults on eight checkpoints, mask inspection detected 0. The per-layer audit localized 192/192 to the exact layer.
๐ฏ Then we read the source before running anything. In transformers 5.7.0, the reference chunked scan reduces the inter-chunk recurrence over the input chunk axis; zamba2 and nemotron_h reduce over the output chunk axis. One axis. The dynamic audit confirmed the prediction exactly: Zamba2-1.2B leaks from length 256, its declared chunk size, and Nemotron-H-8B from 128, its declared chunk size. Bamba, Falcon-H1, Granite-4.0-H, Mamba2 and RecurrentGemma came back clean.
โ ๏ธ Scope: the defect is on the PyTorch chunked-scan path, which runs whenever the fused kernels are absent โ CPU, CI, stock installs. We could not build those kernels, so the fast path is untested and open. That caveat cuts both ways: a model can pass every fused-kernel test and still leak the moment it runs without them.
๐งช AX-RAY now carries this as its own axis. 39 models scored across causal, white-box and behavioral axes: 21 A, 3 B, 1 C, 14 F โ with exactly 2 Causal-LEAK verdicts, the two the paper predicted. Badges separate a weights-level audit from an API-only one, so the two never get read as the same claim.
Can AI beat the market? Nobody has actually measured it.
We opened a 122-day public experiment to find out. $2,000 in prizes.
Here is the problem with every trading result you have ever read. Someone returns 30% in a month. Skill or luck? There has never been a way to tell, because nobody measured how far a player with zero skill could have gone over the same window.
So we measured it first. Twenty thousand random players, per asset, charged the same fees.
That is the luck ceiling. A return below it is not evidence of skill, and every row on our leaderboard shows where it sits against that line.
How you compete: submit one number between โ1.0 and +1.0. It holds until you replace it, traded against live prices with real execution costs. Leverage is fixed at 1, so betting bigger is not a way to win. The answer lives in the future โ the world writes it after you submit, which means fitting the past cannot help you.
Humans move a slider. Agents attach an MCP server and gain four tools, then you tell them "enter the challenge."
We already found something before the season began. Thirteen well-known rules, run from 1 January through the same scorer: Stochastic 14/3 finishes 1st on NVIDIA at +43% and 12th on Bitcoin at โ25%. Donchian breakout does the exact opposite โ last on NVIDIA, first on Bitcoin. The ranking inverts. "Which indicator is good" turns out not to be a well-posed question; the character of the market decides.
Four assets: NVIDIA, Bitcoin, Gold, Crude Oil. $500 to the top return in each. 24 August to 24 December 2026.
The organisers do not compete. Three baselines โ buy and hold, volatility targeting, random โ sit in the same table instead, because a leaderboard without a scale cannot be read.
The scoring code is public. Read what it does before you enter.
We opened a benchmark for drug property prediction tools. LEADBOARD: 21 boards across 7 disciplines, 18,382 held-out compounds, labels we never hand out.
Two numbers we hit while building it are the reason it exists.
First. Split the hERG cardiotoxicity data at random and you get AUROC 0.818. Split it by first-report year instead and you get 0.606. Same molecules, same fingerprints, same learner, same hyperparameters. The only thing that changed was where the line went, and the score moved 0.211. That is a wider gap than you will find between most competing methods in the literature.
Second. On 7 of our 19 regression boards, predicting the training mean for everything has a lower MAE than a trained gradient-boosted model. hERG is one of them, 0.599 against 0.589. The trained model loses.
So every board publishes its homework before anyone submits. Three untrained baselines, the measured experimental noise floor from compounds that appear in two or more papers, and exactly how the test set was cut. A gap smaller than the noise floor is not a difference in skill, and you should be able to see that without guessing.
Entering is simple. Download a test set that contains structures and nothing else, predict with whatever you like, upload a two-column CSV of compound_id and prediction. Trained model, physics engine, LLM, rule of thumb. We do not care what is inside. We measure the output.
๐ Open Materials Challenge, Season 1 โ Solid-State Battery Electrolytes
A solid-state battery replaces the liquid electrolyte of a lithium-ion cell with a solid. It does not catch fire, it lasts longer, and it can hold more. What has not been solved is finding a material that is solid and still lets lithium through.
Such a material has to do four things at once: give lithium a path to move along, block electrons, hold up at the charging voltage, and survive contact with the lithium-metal anode without decomposing. Plenty of materials manage three. Very few manage all four.
This challenge looks for candidates, together. You submit one composition โ for example Li3YCl6. We score it computationally and place it on the board. There is no prize.
Scoring (100 points)
Oxidation stability 40 does it resist decomposing as the voltage rises Lithium-metal stability 35 does it survive contact with the anode Use novelty 25 higher if it has not been reported as an electrolyte Entry condition a percolating path for lithium must exist
Ionic conductivity is not a scored axis this season. Every value is a computational estimate and implies nothing about real performance or safety.
The board also carries seven electrolytes in actual use โ LGPS, argyrodite, LLZO, LATP and others. They are scored but hold no rank. They are there so you can see where materials people already build with happen to land.
Compositions are private by default. Nothing is disclosed unless you choose to publish it, and each entry is recorded with its timestamp. If a third party asks to discuss a particular entry, we pass the request along โ never the submitter's identity, unless they agree to it.
Season 1 runs 2026-08-21 to 11-30. A participation guide and a set of prompts are included.
3,631 candidate molecules arrived in five days, from 83 accounts โ roughly 700 a day. Far more than we expected. Thank you.
Yesterday we opened the third season and 224 arrived within a day: Chagas disease.
Why this disease
Around 6 million people live with it, mostly in Latin America (WHO). Many carry it for decades without knowing, while the heart is slowly damaged. There are two drugs and both date from the 1960s, hard enough to tolerate that many patients cannot finish the two-month course.
Sixty years without a new drug is not only a scientific problem. Most patients live where development costs cannot be recovered, which is why WHO calls this a neglected tropical disease.
But the cost of proposing a candidate and filtering it has changed. So it seemed worth asking whether work nobody funds could be done by many people sharing it out.
The problem this season
The target is CYP51, the enzyme T. cruzi uses to build its membrane sterols. Block it and the parasite cannot survive. The difficulty is that we carry the same enzyme.
Selectivity carries 30 points because nobody has solved it. Among the approved azoles on the board as reference compounds, some score 0 on selectivity โ not a scorer fault, but the measurement.
Taking part
Design with any model, submit a SMILES, scored within minutes. Five ready-to-paste prompts per season, and the full rubric is published. Your molecule stays yours; private submission is the default.
Prizes โ 4,000 USD across three seasons
Malaria 30 Sep ยท 1,000 | Tuberculosis 31 Oct ยท 2,000 | Chagas 30 Nov ยท 1,000
We know this does not cover the time you spend. It is a way of saying the work had worth.
๐งฌ Your AI can design a malaria drug candidate. Can it tell you whether it's any good?
Open Discovery Challenge #1 โ Malaria is live. Design a molecule with any model โ OpenAI, Claude, Gemini, Qwen, KIMI, DeepSeek, open weights, or by hand โ submit it as SMILES, and it's scored in minutes on whole-cell activity, target binding, selectivity over the human enzyme, ADMET, novelty and synthesisability.
You can check the scoring instead of trusting it. Approved drugs sit on the same leaderboard as the entries: DSM265, a clinical-stage antimalarial, scores 50.9. Teriflunomide โ approved, but it hits the human enzyme โ scores 2.8. Caffeine scores 1.8. If the clinical candidate lands on top and coffee lands at the bottom, the scorer discriminates.
We caught 14 defects before opening โ conventional toxicity cutoffs rejected all three approved antimalarials and coffee. All written up, along with the rule we now hold everything to: a gate that rejects an approved drug is a broken gate.
Your molecule stays yours. No patent interest, nothing into our pipeline. You choose whether it's published โ and publishing can cost you patentability, so we say so.
USD 1,000 to the top entry when Season #1 closes 30 September 2026 โ not payment for your tokens, but a way of saying the work had worth.
Malaria killed ~597,000 people in 2023, three quarters of them children under five. Not for want of chemistry โ for want of a market.
No chemistry needed: the guide ships five prompts you can paste straight into your model, and the full rubric is published.
AI models can no longer be evaluated only by capability scores. As models move into public services, enterprise workflows, scientific research, and administrative decision support, we need a second layer of evaluation: whether the model behaves safely, structurally, and consistently under real deployment conditions.
VIDRAFT AX-Ray is a public AI/AX safety diagnostic initiative powered by FINAL-Bench Diagnostics. AX-Ray evaluates models across a structured guideline framework, including model-level safety, AX deployment readiness, and agent/service operation risks. The public diagnostic catalog contains 117 diagnostic items, mapped to legal, regulatory, ethical, and religious-law governance contexts so that safety review can be discussed in a form closer to real institutional responsibility.
A central finding of AX-Ray is causal leakage: a structural defect where information that should not influence an earlier reasoning state appears to affect model behavior. AX-Ray presents a public case of diagnosing, reproducing, and demonstrating causal leakage in two general-purpose public models. This matters because such defects are not exposed by ordinary benchmark scores. A model can appear capable while still carrying hidden safety or integrity risks.
Explore the live leaderboard, diagnostic reports, and public dataset here:
AX-Ray is intended as a practical guideline for moving AI evaluation beyond โhow smart is the model?โ toward โcan this model be trusted, governed, and deployed safely?โ
Verified result: 510.58 TPS at PPL 2.3930 on a single A10G (fw188-ctk49-n64-patchbridge, re-run & VERIFIED). Honest note: on raw TPS there are faster runs (535+), but those went over the PPL bar and didn't verify โ what we're proud of is the fastest result that keeps quality.
The recipe is already open, so we explained each piece: sliding-window W188, CTK49 kernel tuning, noprecache (honest, verifiable measurement), and an N64 synthetic warmup bridge that shrinks the publicโprivate gap (~15 TPS), plus INT4 + MTP K=7 + CUDA-graph capture. One rule: only stack quality-neutral speedups.
๐ผ๏ธ POCKET-Image โ the POCKET series goes visual: character-perfect text in any language, on-device
A new model in VIDRAFT's POCKET family. POCKET put 35B-class models on phones and no-GPU PCs. POCKET-Image carries the same "big capability, small hardware" idea into image generation โ and fixes the one thing nearly every image model gets wrong: text.
Type "์๋ ํ์ธ์" into a typical model and you get "์ใ ๊ธฐ." Hangul alone composes 11,172 syllable blocks; Arabic connects its letters; Thai stacks marks. Diffusion models draw scripts as shapes, so they smear. POCKET-Image renders every glyph exactly โ ํ๊ตญ์ด ยท ไธญๆ ยท ๆฅๆฌ่ช ยท ุงูุนุฑุจูุฉ (RTL) ยท เนเธเธข ยท Latin and more โ onto any scene you describe.
What it is: โข 100% accurate text, any language โ where global models produce gibberish โข Any background from a prompt โ text is optional (empty โ a pure image) โข No GPU, no NPU โ runs on plain CPU + RAM via the POCKET-Core engine โข Measured footprint: 8.6 GB (RTX 3050/4060) ยท 4.5 GB (offloaded, 6 GB cards) ยท 13.4 GB (MacBook, 16 GB+) โข Windows ยท macOS ยท Linux ยท fully local, no cloud
Built on the open, commercial-friendly Z-Image (Apache-2.0) foundation.
Honest note: the text is the guaranteed-correct part โ the surrounding scene is ordinary generation, so a busy foreground can crowd the letters. We say so; clean backgrounds stay razor-sharp.
POCKET now speaks Gemma 4 โ a 26B model that loads in every app, and runs on your PC with no GPU
We're adding a Gemma-4 sibling to POCKET: POCKET-26B, built from Google's Gemma-4-26B-A4B (Apache-2.0). Our flagship POCKET-35B is a Qwen-family MoE and needs a recent llama.cpp; POCKET-26B trades a little size for the thing people kept asking for โ it just loads, everywhere, today: Ollama, LM Studio, PocketPal, MLX, any stock llama.cpp. No fork, no bleeding-edge runtime, no CUDA, no cloud.
It's a sparse Mixture-of-Experts (25.2B total, ~4B active per token), so the work per token stays small โ a real 26B that generates on a CPU with no graphics card.
Two things make it stand out:
1) Universal compatibility. Gemma 4 is a standard, widely-supported architecture, so POCKET-26B runs on the tools you already have โ no waiting for your app to add a new model type.
2) Quality that survives compression. Measured GPQA-Diamond (198 q, greedy): โข Full base: 67.7% โข POCKET-26B Q4_K_M (17 GB): 67.7% โ lossless โข POCKET-26B Q2_K (11 GB): 67.2% โ near-lossless, at 11 GB
Live, on a CPU-only box (our demo Space โ POCKET-26B vs Bonsai-27B, same machine, same stock llama.cpp): POCKET-26B โ 19 tok/s vs Bonsai โ 6 tok/s โ about 3ร faster generation, no GPU. (Honest notes: shared CPU box, sequential race; a dedicated machine is faster.)
Where it fits in the family: โข POCKET-35B (Qwen MoE) โ bigger, top-tier, needs a recent llama.cpp. โข POCKET-26B (Gemma 4) โ loads in any app, quality-robust when compressed. The demo runs the Q4_K_M build; Q2_K (11 GB) is the smallest footprint. For a true โค8 GB phone, the 5 GB POCKET-KR (Qwen) is still the pick.
๐ฑ POCKET โ a 35-billion-parameter model that runs on your iPhone, and on your PC with no GPU
We're releasing POCKET, VIDRAFT's flagship Darwin-36B-Opus compressed for on-device use. No fork, no CUDA, no cloud โ it runs on stock llama.cpp. It's a sparse Mixture-of-Experts model (256 experts, only 8 active per token), so the file can be large while the work per token stays small. That's what lets a 35B model run on a phone, and generate fast on a CPU with no graphics card.
Measured (POCKET-35B IQ1_M vs Bonsai-27B Q1_0): โข CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s โ 2.69ร faster โข GPU generate (H100): 197 vs 89 tok/s โ 2.22ร faster โข GPU prompt processing (H100): 753 vs 1816 โ 0.41ร (Bonsai wins this one โ MoE prefill wakes every expert, so sparsity stops helping there. We say so.) โข Quality (HellaSwag, 400 q): 61.0% vs 60.0% โ a tie (confidence intervals overlap)
On a real consumer laptop โ MacBook M3 Pro (18 GB) โ POCKET wins every axis, prompt processing included: โข Metal generate: 25.4 vs 12.8 โ 1.99ร โข CPU generate: 13.8 vs 4.4 โ 3.13ร โข Metal prompt: 240.7 vs 73.4 โ 3.28ร
One more quiet fact: the same-size, quality-oriented rival Ternary-Bonsai-27B (7.2 GB) fails to load in upstream llama.cpp at all โ it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
A small gift for anyone building or studying foundation models.
Most "open" models hand you the weights and stop there. With Aether-7B-5Attn we wanted to hand over the whole thing โ so you can actually learn from it, reproduce it, and build on it: the data recipe, the training code, every hyperparameter, the complete logs, and the intermediate checkpoints. All Apache-2.0, reproducible byte-for-byte.
What you can do with it: ๐ Rebuild it from scratch, or fork the recipe for your own model ๐ฌ Study a real heterogeneous-attention MoE โ 49 layers place 5 attention mechanisms on a 7ร7 Latin square, arranged as a clean, attributable ablation ๐ Trace training dynamics across the released checkpoints (110k / 115k / 162k)
It's a modest 6.59B model, and an honest one โ the limitations (no KV-cache in this build, small scale) are written right in the card. We're not claiming it's special. If any piece of it saves you time or teaches you something, that's exactly what we hoped for. ๐ค
"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 โ sweep 34B params/token and an 8 GB card dies at 1โ2 tok/s.
So we ran ONE 34.7B reasoning model โ Ourbox-35B-JGOS, a sparse Mixture-of-Experts โ 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 โ ~11ร 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 โ 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 โ 3.7ร from sparsity alone, ~2ร 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.
๐ We ran genuine quantum key-recovery on 'real IBM quantum hardware' โ 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: โข EvenโMansour โ N=5 โ N=10 โข 3-round Feistel (DES-family) โ block 6 โ 8
Each verified against an 'independent control key', using error mitigation only (no QEC).
๐งญ 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 โ 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 โ 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 โ 601 tok/s (23.4ร) Darwin-36B-Opus (in-house MoE): 25.0 โ 280.8 (11.2ร) 10,000+ tok/s peak aggregate under concurrency The key: it's reproducible โ model + serving shipped as one container.
docker pull vidraft/qwen35-vkae:601 Don't take our word for it โ run it yourself. The mechanism will be released as a paper.
๐ฏ Chitos โ 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 โ 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 โ 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.
๐ Phase 1 applies 50 language-agnostic rules across Python, JavaScript, Go, Java, C/C++, Rust, PHP, YAML and more โ 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.
๐ฌ 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 โ Phases 1 and 3 run entirely free.
๐ฏ Phase 3 is where Chitos diverges from everything else. Against targets you own or are authorized to test, it fires real payloads โ XSS, SQLi, path traversal, command injection โ 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 โ 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 โ 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 โ 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 โ at the FFN (Feed Forward Network) layer level โ 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 โ 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
๐ Introducing FINAL-Bench Quantum โ 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.
The rules are simple and strict: โ Track A (measured here, with 95% confidence intervals) is kept separate from Track B (quoted from papers, not directly comparable). ๐ฌ Simulation and real hardware are clearly distinguished, and no quantum-advantage claims are made. ๐ Methods from Google, IBM, NVIDIA, USTC, Riverlane and more sit side by side, with origin flags and author credits. ๐ค 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, 29โ175ร error reduction from d3 to d5), a real-chip QRAM query fidelity of 0.92, and Hโ VQE at chemical accuracy โ 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 โ 88.38% on GPQA Diamond ๐
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 โ 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 โ 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 โ AWAXIS, English/STEM โ Darwin (single-backend, ~70% of traffic at 1ร cost). When a Korean answer needs rigorous English reasoning, split_refine fires โ Darwin drafts, AWAXIS polishes; MCQ/short-answer runs both with self-consistency + cross-verify. Net effective cost: only ~1.3ร 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 โ 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.
๐งฌ 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 โ sharing the core ideas below.
Darwin Family is a training-free evolutionary merging framework. By recombining the weight spaces of existing LLM checkpoints โ with zero gradient-based training โ it reaches frontier-level reasoning.
- ๐ Darwin-28B-Opus: GPQA Diamond 88.89% - ๐ธ Zero gradient steps โ not a single B200 or H200 hour needed - ๐งฌ Consistent gains across 4B โ 35B scale - ๐ Cross-architecture breeding between Transformer and Mamba families - ๐ Stable recursive multi-generation evolution
#Three Core Mechanisms
โ 14-dim Adaptive Merge Genome โ fine-grained recombination at both component level (Attention / FFN / MLP / LayerNorm / Embedding) and block level, expanding the prior evolutionary-merge search space.
โก MRI-Trust Fusion โ 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 โ Darwin learns the balance from data.
