FINAL-Bench/Darwin-4B-David
Text Generation โข Updated โข 403 โข 16
chohyunkyu
chohyun
AI & ML interests
None yet
Recent Activity
liked a model about 24 hours ago
FINAL-Bench/Darwin-4B-David liked a Space about 24 hours ago
FINAL-Bench/Darwin-4B-david reacted to SeaWolf-AI's post with ๐ about 24 hours ago
Why This Matters โ David Defeats Goliath
MODEL: https://huggingface.co/FINAL-Bench/Darwin-4B-David
SPACE: https://huggingface.co/spaces/FINAL-Bench/Darwin-4B-david
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 โ surpassing its 58.6% original ancestor and even gemma-4-31B (84.3%) โ 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 โ 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 โ 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% โ 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 โ 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.
Organizations
None yet
liked a
model about 24 hours ago
liked a
Space about 24 hours ago
reactedto SeaWolf-AI's post with ๐ about 24 hours ago
Post
1986
Why This Matters โ David Defeats Goliath
MODEL: FINAL-Bench/Darwin-4B-David
SPACE: FINAL-Bench/Darwin-4B-david
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 โ surpassing its 58.6% original ancestor and even gemma-4-31B (84.3%) โ 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 โ 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 โ 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% โ 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 โ 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.
MODEL: FINAL-Bench/Darwin-4B-David
SPACE: FINAL-Bench/Darwin-4B-david
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 โ surpassing its 58.6% original ancestor and even gemma-4-31B (84.3%) โ 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 โ 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 โ 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% โ 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 โ 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.
liked a
Space 13 days ago
liked a
dataset 13 days ago
liked a
Space 13 days ago
upvoted an article 13 days ago
Article
Introducing WM Bench: A Benchmark for Cognitive Intelligence in World Models
โข
13
reactedto SeaWolf-AI's post with ๐ฅ 13 days ago
Post
4668
๐ World Model Bench โ 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 โ 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 โ 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.
3 Pillars ยท 10 Categories ยท 100 Scenarios ยท 1,000-point scale
- ๐ P1 Perception (25%) โ Can it read the scene?
- ๐ง P2 Cognition (45%) โ Does it predict threats, escalate emotions, utilize memory?
- ๐ฅ P3 Embodiment (30%) โ Does the body respond with the right motion?
All evaluation is via simple JSON I/O โ no 3D engine, no special hardware. Any model with an API can participate.
We also built PROMETHEUS as a live reference implementation โ runs in your browser on a T4, no install needed. Combines FloodDiffusion motion generation with a LLM cognitive brain (Perceive โ Predict โ Decide โ Act). Scored 726/1000 (Grade B) on Track C โ the only directly verified model so far. Submissions from other teams very welcome.
---
๐ Dataset โ FINAL-Bench/World-Model
๐ Demo โ FINAL-Bench/World-Model
๐ Leaderboard โ FINAL-Bench/worldmodel-bench
๐ Article โ https://huggingface.co/blog/FINAL-Bench/world-model
Part of the FINAL Bench Family โ alongside FINAL Bench (Feb 2026). Feedback on rubrics and missing models always welcome!
FID measures realism. FVD measures smoothness. But neither tells you whether the model understood the scene.
We just released WM Bench โ 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 โ 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.
3 Pillars ยท 10 Categories ยท 100 Scenarios ยท 1,000-point scale
- ๐ P1 Perception (25%) โ Can it read the scene?
- ๐ง P2 Cognition (45%) โ Does it predict threats, escalate emotions, utilize memory?
- ๐ฅ P3 Embodiment (30%) โ Does the body respond with the right motion?
All evaluation is via simple JSON I/O โ no 3D engine, no special hardware. Any model with an API can participate.
We also built PROMETHEUS as a live reference implementation โ runs in your browser on a T4, no install needed. Combines FloodDiffusion motion generation with a LLM cognitive brain (Perceive โ Predict โ Decide โ Act). Scored 726/1000 (Grade B) on Track C โ the only directly verified model so far. Submissions from other teams very welcome.
---
๐ Dataset โ FINAL-Bench/World-Model
๐ Demo โ FINAL-Bench/World-Model
๐ Leaderboard โ FINAL-Bench/worldmodel-bench
๐ Article โ https://huggingface.co/blog/FINAL-Bench/world-model
Part of the FINAL Bench Family โ alongside FINAL Bench (Feb 2026). Feedback on rubrics and missing models always welcome!
liked a
Space 27 days ago
liked a
dataset about 1 month ago
upvoted an article about 1 month ago
Article
๐๏ธ Smol AI WorldCup: A 5-Axis Benchmark That Reveals What Small Language Models Can Really Do
โข
38
reactedto SeaWolf-AI's post with โค๏ธ about 1 month ago
Post
11122
๐๏ธ Smol AI WorldCup: A 4B Model Just Beat 8B โ 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.
Community Article: https://huggingface.co/blog/FINAL-Bench/smol-worldcup
Live Leaderboard: ginigen-ai/smol-worldcup
Dataset: ginigen-ai/smol-worldcup
What we found:
โ Gemma-3n-E4B (4B, 2GB RAM) outscores Qwen3-8B (8B, 5.5GB). Doubling parameters gained only 0.4 points. RAM cost: 2.75x more.
โ GPT-OSS-20B fits in 1.5GB yet matches Champions-league dense models requiring 8.5GB. MoE architecture is the edge AI game-changer.
โ Thinking models hurt structured output. DeepSeek-R1-7B scores 8.7 points below same-size Qwen3-8B and runs 2.7x slower.
โ 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.
โ 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?" โ 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.
Top 5 by WCS:
1. GPT-OSS-20B โ WCS 82.6 โ 1.5GB โ Raspberry Pi tier
2. Gemma-3n-E4B โ WCS 81.8 โ 2.0GB โ Smartphone tier
3. Llama-4-Scout โ WCS 79.3 โ 240 tok/s โ Fastest model
4. Qwen3-4B โ WCS 76.6 โ 2.8GB โ Smartphone tier
5. Qwen3-1.7B โ WCS 76.1 โ 1.2GB โ IoT tier
Built in collaboration with the FINAL Bench research team. Interoperable with ALL Bench Leaderboard for full small-to-large model comparison.
Dataset is open under Apache 2.0 (125 questions, 7 languages). We welcome new model submissions.
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.
Community Article: https://huggingface.co/blog/FINAL-Bench/smol-worldcup
Live Leaderboard: ginigen-ai/smol-worldcup
Dataset: ginigen-ai/smol-worldcup
What we found:
โ Gemma-3n-E4B (4B, 2GB RAM) outscores Qwen3-8B (8B, 5.5GB). Doubling parameters gained only 0.4 points. RAM cost: 2.75x more.
โ GPT-OSS-20B fits in 1.5GB yet matches Champions-league dense models requiring 8.5GB. MoE architecture is the edge AI game-changer.
โ Thinking models hurt structured output. DeepSeek-R1-7B scores 8.7 points below same-size Qwen3-8B and runs 2.7x slower.
โ 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.
โ 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?" โ 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.
Top 5 by WCS:
1. GPT-OSS-20B โ WCS 82.6 โ 1.5GB โ Raspberry Pi tier
2. Gemma-3n-E4B โ WCS 81.8 โ 2.0GB โ Smartphone tier
3. Llama-4-Scout โ WCS 79.3 โ 240 tok/s โ Fastest model
4. Qwen3-4B โ WCS 76.6 โ 2.8GB โ Smartphone tier
5. Qwen3-1.7B โ WCS 76.1 โ 1.2GB โ IoT tier
Built in collaboration with the FINAL Bench research team. Interoperable with ALL Bench Leaderboard for full small-to-large model comparison.
Dataset is open under Apache 2.0 (125 questions, 7 languages). We welcome new model submissions.
- 1 reply
liked a
Space about 1 month ago
liked a
Space about 2 months ago
liked 5
Spaces 9 months ago
Unfiltered AI NSFW Image generation
โก
357
Generate high-quality custom images from text prompts
VEO3 Directors
๐
92
Wan2.1-T2V-14B + Fast 4-step with NAG + Automatic Audio
VEO3 Free
๐
457
Wan2.1-T2V-14B + Fast 4-step with NAG + Automatic Audio
Ghibli Meme Studio
๐ผ
25
Create avatars and profile images, turning your memes
ThinkFlow
๐จ
21
thought that reasoning into LLMs without modification
liked a
model 9 months ago
