---
license: gemma
language:
- en
- ko
- multilingual
library_name: transformers
pipeline_tag: text-generation
tags:
- darwin
- darwin-family
- darwin-duo
- duo
- ensemble
- mixture-of-models
- router
- korean
- reasoning
- finalbench
- vidraft
model-index:
- name: Darwin-60B-DUO
results:
- task:
type: question-answering
name: Question Answering
dataset:
name: GPQA Diamond
type: Idavidrein/gpqa
config: gpqa_diamond
metrics:
- type: accuracy
value: 88.38
name: Accuracy
---
> ### π± Run it on your phone or a GPU-less PC β **POCKET** Β· π **[Try it live (CPU chat)](https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU)**
> VIDRAFT's on-device family: a 35B model that runs on **iPhone** and on **CPU with no GPU** β stock `llama.cpp`, no fork.
>
> [](https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU) [](https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6) [](https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF) [](https://huggingface.co/FINAL-Bench/POCKET-KR-MLX) [](https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF)
>
# Darwin-60B-DUO
> The Darwin family's first **DUO** model β two domain-verified SOTA models served as a single OpenAI-compatible endpoint.
>
> ## π GPQA Diamond (full 198 questions): **88.38%**
> Darwin-28B-REASON single 69.70 % Β· AWAXIS-Think-31B single 77.27 % Β· first cascade 83.84 % Β· **Darwin-DELPHI DUO 88.38 %**
Darwin-60B-DUO unifies two specialist models from the Darwin family behind a single API:
- **Darwin-28B-REASON** β Hugging Face leaderboard **GPQA Diamond rank #3**, English graduate-level reasoning specialist.
- **AWAXIS-Think-31B** β National **K-AI Leaderboard rank #1** (operated by the Ministry of Science and ICT of the Republic of Korea), Korean specialist.
A Hybrid-A router automatically dispatches each request to the optimal strategy (single route / sequential collaboration / ensemble), so callers see one model and one endpoint while internally benefiting from both specialists.
---
## Model Description
Darwin-60B-DUO is a **gateway-orchestrated aggregate** of two constituent base models. The repository contains a FastAPI orchestrator, configuration, and Docker Compose recipe. The model weights themselves live in the constituent repositories and are loaded at runtime by two vLLM backends.
| Component | Source | Architecture | Parameters | Verified Rank |
|-----------|--------|--------------|-----------:|---------------|
| English reasoning specialist | [FINAL-Bench/Darwin-28B-REASON](https://huggingface.co/FINAL-Bench/Darwin-28B-REASON) | Qwen3.5 multimodal | 26.9 B | HF GPQA Diamond #3 |
| Korean specialist (Darwin family, Gemma-4 branch) | [Anserwise/AWAXIS-Think-31B](https://huggingface.co/Anserwise/AWAXIS-Think-31B) | Gemma-4 multimodal | 31.27 B | National K-AI Leaderboard #1 |
| Aggregate | This repository (`FINAL-Bench/Darwin-60B-DUO`) | DUO orchestrator | **58.17 B** | β |
> **Note on AWAXIS membership.** AWAXIS-Think-31B is also part of the Darwin family β it is the Korean specialist branch distilled by the Darwin team on top of Google's Gemma-4 base, complementing the original Qwen3.5-line Darwin lineage as the family's second axis.
### Hybrid-A Orchestration
The gateway analyzes each incoming request and selects one of five strategies. The default distribution observed on representative traffic is:
| Strategy | When it fires | Backends called | Cost vs. single 30 B | Share |
|----------|---------------|-----------------|---------------------:|------:|
| `route_awaxis` | Korean-dominant input | AWAXIS only | 1Γ | ~50 % |
| `route_darwin` | English-dominant input | Darwin only | 1Γ | ~20 % |
| `split_refine` | Korean output requiring rigorous English / STEM reasoning | Darwin (draft) β AWAXIS (polish) | 2Γ | ~15 % |
| `split_refine_reverse` | English output requiring Korean cultural / linguistic context | AWAXIS (draft) β Darwin (polish) | 2Γ | ~5 % |
| `ensemble_v1` | Short-answer / multiple-choice queries | Both backends with self-consistency + cross-verification tournament | 2Γ | ~10 % |
Average effective cost is approximately **1.3Γ a single 30 B model** β 70 % of traffic is served by a single backend; the remaining 30 % uses both.
---
## Intended Use
### Primary use cases
- **Bilingual Korean-English assistants** that require both Korean fluency and high-quality English reasoning.
- **Single-endpoint integration** where downstream tooling already targets the OpenAI Chat Completions API (LangChain, LlamaIndex, OpenAI SDK, Continue, Cursor, etc.).
- **Cost-conscious deployment** β most traffic is served by a single backend at 1Γ cost while difficult cross-domain queries automatically receive a 2Γ collaboration.
### Out-of-scope
- **Vision / video generation.** Both constituent models are text-mode only as deployed here (`--limit-mm-per-prompt {"image":0,"video":0}`).
- **Real-time streaming.** Initial gateway release does not stream token-by-token. Streaming is planned for v1.1.
- **Direct `AutoModel.from_pretrained()` loading.** This repository contains an orchestrator, not unified weights. Use the gateway or Docker Compose.
- **Safety-critical decision making** (medical diagnosis, legal advice, autonomous control). LLM hallucinations apply.
---
## How to Use
### Option A β Docker Compose (recommended)
```bash
git clone https://huggingface.co/FINAL-Bench/Darwin-60B-DUO
cd Darwin-60B-DUO
# HF token is needed only to download constituent weights on first launch
export HF_TOKEN=hf_xxx
docker compose -f docker/docker-compose.yml up -d
# Verify
curl http://localhost:8000/v1/models
# {"object":"list","data":[{"id":"darwin-60b-duo",...}]}
```
### Option B β Manual launch (2 Γ B200 / H100, FP8)
```bash
# Clone the repo (includes both constituent weights, ~120GB LFS)
git lfs install
git clone https://huggingface.co/FINAL-Bench/Darwin-60B-DUO
cd Darwin-60B-DUO
# 1. Darwin-28B-REASON on GPU 0 (from local subfolder)
CUDA_VISIBLE_DEVICES=0 VLLM_DP_MASTER_PORT=45011 \
vllm serve ./darwin-28r \
--port 8021 --served-model-name darwin-28r \
--quantization fp8 --enforce-eager \
--limit-mm-per-prompt '{"image":0,"video":0}' &
# 2. AWAXIS-Think-31B on GPU 1 (from local subfolder)
CUDA_VISIBLE_DEVICES=1 VLLM_DP_MASTER_PORT=45012 \
vllm serve ./awaxis-31b \
--port 8022 --served-model-name awaxis-31b \
--quantization fp8 --enforce-eager \
--limit-mm-per-prompt '{"image":0,"video":0}' &
# 3. Gateway
pip install -r gateway/requirements.txt
python gateway/server.py --port 8000 \
--darwin-url http://127.0.0.1:8021/v1 \
--awaxis-url http://127.0.0.1:8022/v1
```
> **No external HF download.** Both weights are inside this repo (`./darwin-28r`, `./awaxis-31b`), so vLLM serves directly from local LFS without re-downloading from upstream repos.
> **Single-GPU collocation.** With FP8 quantization the combined footprint is ~30 GB. Set both backends to `CUDA_VISIBLE_DEVICES=0` and `--gpu-memory-utilization 0.45` to colocate on a single 80 GB B200 / H100.
### OpenAI-compatible call
```python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="anything")
# The router picks split_refine (Korean output with English/STEM reasoning)
resp = client.chat.completions.create(
model="darwin-60b-duo",
messages=[{
"role": "user",
"content": "Explain the practical difference between GPT-5 and Claude's reasoning, in Korean.",
}],
)
print(resp.choices[0].message.content)
# Darwin produces the English reasoning; AWAXIS polishes it into natural Korean.
```
You can also force a specific strategy via the non-standard `duo_strategy` field:
```python
resp = client.chat.completions.create(
model="darwin-60b-duo",
messages=[{"role":"user","content":"Which is correct? (A) ... (B) ..."}],
extra_body={"duo_strategy": "ensemble_v1"}, # force MAJ@8 + cross-verify
)
```
Inspect the chosen strategy in the response under `_duo_route`:
```json
{
"choices": [{"message": {"role":"assistant", "content":"..."}}],
"_duo_route": {
"strategy": "split_refine",
"reason": "korean_output_with_english_reasoning",
"elapsed_s": 4.83,
"language_ratio": 0.42
}
}
```
---
## π οΈ Build Your Own DUO (the gateway is model-agnostic)
The `gateway/` in this repo is **not specific to Darwin and AWAXIS** β it is a general, Apache-2.0 orchestrator that fuses **any two OpenAI-compatible backends** into a single endpoint. We're open-sourcing it because the pattern is broadly useful, and we'd love to see what the community builds.
### The core idea in one sentence
> Two small specialists, automatically routed, beat one big generalist on cost **and** quality β and the user only ever sees a single model.
### Pair anything
You don't need our models. Point the gateway at *your* favourite two and you have a DUO in minutes:
| Pair | Why it works |
|------|--------------|
| reasoning model + multilingual model | rigorous logic, fluent output (what we did) |
| fast small model + slow accurate model | cheap on easy queries, escalate only hard ones |
| code model + general chat model | route by "is this code?" |
| English specialist + your-language specialist | localize without losing reasoning depth |
### Adapt in 2 steps
```bash
# Step 1 β point at your two backends (any vLLM / TGI / OpenAI-compatible server)
python gateway/server.py --port 8000 \
--darwin-url http://your-model-A:8000/v1 \
--awaxis-url http://your-model-B:8000/v1
```
```python
# Step 2 β teach the router when to use which model (gateway/router.py)
def select_strategy(text):
if is_code(text): return "route_model_a" # send code to the code model
if is_korean(text): return "route_model_b" # send Korean to the KO model
if is_mcq(text): return "ensemble_v1" # short-answer β both + pick
if needs_polish(text): return "split_refine" # draft with A, polish with B
return "route_model_b" # sensible default
```
That's it. The gateway handles the OpenAI API surface, parallel calls, response merging, and `_duo_route` transparency for you.
### What's included vs. what's proprietary
- β
**Open (Apache-2.0):** the gateway β router, single-route, sequential-refine, and a baseline ensemble. Enough to build a fully working DUO of your own.
- π **Proprietary (Darwin-DELPHI):** the test-time engine and prompt recipes that push *this* DUO to **88.38 % on GPQA Diamond** (above the naive oracle ceiling). The open gateway gives you the architecture; the leaderboard score is ours.
### Extend it
- **DUO β TRIO β N-way:** add a third backend and a third route; the merge logic generalizes.
- **Smarter routing:** swap the keyword heuristics in `router.py` for a tiny learned classifier.
- **New strategies:** contribute your own (debate, best-of-N, tool-use) via PR.
> **Fork it, point it at your models, and tell us what you built.** Open a discussion on the Community tab β good routing recipes and new backend pairings are exactly the kind of thing we want to feature.
---
## Darwin Family
Darwin-60B-DUO sits at the confluence of two complete lineages β the **Qwen3.5-based Darwin lineage** (English reasoning) and the **Gemma-4-based Darwin Korean-specialist branch (AWAXIS)**. The full family tree, with both constituent ancestries fully expanded:
```mermaid
graph TD
%% Lineage A β English reasoning
A1[Cohere Command A+ - 218B foundation]:::found --> A2[Darwin-28B-Opus - English reasoning base]:::parent
A2 --> A3[Darwin-28B-REASON - HF GPQA Diamond #3]:::spec
%% Other Darwin parents
A1 -.-> P1[Darwin-218B-Delphi - cascade flagship GPQA 90.91%]:::parent
A2 -.-> P2[Darwin-9B - omni-modal ko/en compact]:::parent
P2 -.-> P3[Darwin-31B-Opus - Korean multimodal base]:::parent
%% Lineage B β Korean specialist
B1[Google Gemma-4-31B-it - Korean/multilingual base]:::found --> B2[TeichAI gemma-4-31B-it-Claude-Opus-Distill-v2]:::parent
B2 --> B3[AWAXIS-Think-31B - National K-AI Leaderboard #1, Darwin family Korean specialist]:::spec
%% The DUO unification
A3 --> DUO[Darwin-60B-DUO - this model]:::duo
B3 --> DUO
classDef found fill:#e8f0fe,stroke:#1a73e8,color:#0a0a0a
classDef parent fill:#fff4e5,stroke:#f29900,color:#0a0a0a
classDef spec fill:#e6f4ea,stroke:#34a853,color:#0a0a0a
classDef duo fill:#fce8f3,stroke:#d81b60,color:#0a0a0a,stroke-width:3px
```
### Plain-text fallback
```
Darwin Family
Lineage A β English reasoning (Qwen3.5-line)
βββ Cohere Command A+ (218B foundation)
βββ Darwin-28B-Opus (English reasoning base)
βββ Darwin-28B-REASON π₯ β HF GPQA Diamond #3
(English reasoning specialist)
β
β Related Darwin parents in this lineage:
β β’ Darwin-218B-Delphi (cascade flagship, GPQA Diamond 90.91 %)
β β’ Darwin-9B (omni-modal ko/en compact)
β β’ Darwin-31B-Opus (Korean multimodal base)
Lineage B β Korean specialist (Gemma-4-line)
βββ Google Gemma-4-31B-it (Korean / multilingual base)
βββ TeichAI gemma-4-31B-it-Claude-Opus-Distill-v2
βββ AWAXIS-Think-31B π₯ β National K-AI Leaderboard #1
(Darwin family Korean specialist)
β β
βββββββββ DUO unification βββββ
β
β Darwin-60B-DUO β β THIS MODEL
"Two SOTAs, one OpenAI-compatible endpoint."
```
> The HF **Model tree** widget (right sidebar) automatically renders the upstream chain from each `base_model` declared in the YAML frontmatter, so the full ancestry β Cohere Command A+ and Google Gemma-4-31B-it at the roots β is browsable directly on this page.
---
## Operation Modes
### Mode 1 β Route (single backend, ~70 % of traffic)
The router selects one backend based on language ratio and lightweight keyword heuristics:
- `korean_ratio(prompt) > 0.3` β AWAXIS
- ASCII / code / math markers (`def `, `import`, `\boxed`, `prove`, β¦) β Darwin
- Mixed β AWAXIS (Korean-first default)
### Mode 2 β Split / Refine (sequential collaboration, ~20 % of traffic)
One model drafts, the other polishes. The polish instruction is language-adaptive:
```
User: "Explain entropy intuitively in Korean."
Step 1 β Darwin (rigorous English reasoning)
"Entropy quantifies the number of microstates compatible
with a given macrostate, representing disorder ..."
Step 2 β AWAXIS (natural Korean polish)
"μνΈλ‘νΌλ μ½κ² λ§νλ©΄ '무μ§μν¨μ μ λ'μ
λλ€.
κ°μ λͺ¨μ΅μΌλ‘ 보μ΄μ§λ§ μ¬μ€ κ·Έ μμ ..."
```
The reverse path (`AWAXIS draft β Darwin polish`) fires when the output language is English but the prompt requires Korean cultural or linguistic context.
### Mode 3 β Ensemble Vβ Tournament (~10 % of traffic)
For multiple-choice and short-answer queries, both backends produce **N = 8 samples** at temperature 0.7. Each backend's answer is its own majority vote (self-consistency). If the two majorities agree β return. If they disagree β each backend verifies the pair (cross-verification), and the tournament winner is selected. A confidence tiebreaker (own-vote count) resolves split verdicts.
---
## Repository Layout
```
Darwin-60B-DUO/
βββ README.md β this model card
βββ image.png β cover image
βββ config.json β DUO configuration & orchestration metadata
βββ tokenizer_info.json β constituent tokenizer references
βββ LICENSE β Gemma + Apache-2.0 dual notice
βββ darwin-28r/ β Darwin-28B-REASON full weights + tokenizer + config (~54GB)
β βββ config.json
β βββ tokenizer.json
β βββ chat_template.jinja
β βββ model.safetensors.index.json
β βββ model-00001-of-00012.safetensors ... model-00012-of-00012.safetensors
βββ awaxis-31b/ β AWAXIS-Think-31B full weights + tokenizer + config (~62.5GB)
β βββ config.json
β βββ tokenizer.json
β βββ chat_template.jinja
β βββ model.safetensors.index.json
β βββ model-00001-of-00002.safetensors ... model-00002-of-00002.safetensors
βββ gateway/
β βββ server.py β FastAPI OpenAI-compatible orchestrator
β βββ router.py β language / domain / MCQ classifier
β βββ refine.py β sequential refine (drafter β polisher)
β βββ ensemble.py β Vβ MAJ@N + cross-verification
β βββ requirements.txt
βββ docker/
β βββ docker-compose.yml β vLLM Γ2 + gateway integrated launcher
βββ benchmarks/
βββ README.md β full evaluation results
βββ cascade_v4_32k.json β cascade v4 raw data
βββ recovery_v5.json β recovery v5 raw data
βββ poe_v6.json β POE v6 raw data (latest)
```
> **Self-contained:** all model weights (both constituents) are included in this repository β `git clone` once and you have everything. Total LFS payload: **~120GB**.
---
## Evaluation
### Verified constituent scores
| Constituent | Benchmark | Rank |
|-------------|-----------|------|
| Darwin-28B-REASON | Hugging Face GPQA Diamond | **#3** |
| AWAXIS-Think-31B | National K-AI Leaderboard (Republic of Korea, MSIT) | **#1** |
### Darwin-60B-DUO aggregate
| Benchmark | Status |
|-----------|--------|
| GPQA Diamond (full 198 questions) | **88.38%** (175/198) β see `benchmarks/` |
| KMMLU | TBA |
| CLIcK (Korean cultural reasoning) | TBA |
| Helmet / Ruler (long context) | TBA |
| Needle-in-Haystack 32 K / 128 K | NIAH 32 K: 5/5 per backend (sanity, single model only) β full DUO numbers TBA |
> Aggregate DUO benchmark results will be published in `benchmarks/` after formal evaluation. The verified constituent ranks above are independent third-party measurements and are not aggregate DUO scores.
### Cost / latency notes
- **Route mode:** comparable to a single 30 B FP8 backend (1Γ latency, 1Γ compute).
- **Split mode:** ~2Γ latency (two sequential generations).
- **Ensemble Vβ:** ~2Γ compute (parallel) plus a short cross-verify round when majorities disagree.
---
## Bias, Risks, and Limitations
- **Hallucination.** Standard LLM caveats apply. Both backends can produce confident but incorrect outputs, especially on out-of-distribution queries.
- **Disagreement bias.** Empirically, the Vβ tournament occasionally selects a wrong answer that both backends collectively favor over a single backend's correct one. The gateway exposes the routing decision in `_duo_route` for auditability.
- **Language coverage.** Best performance on English and Korean. Other languages fall back to the closer-fit backend without explicit optimization.
- **Combined weights are not bundled.** The aggregator pulls each backend's weights from the constituent repositories. Network and disk for both is required.
- **Two-GPU baseline.** BF16 deployment requires two GPUs. FP8 quantization enables single-GPU operation on B200 / H100 (80 GB).
- **Training data cut-off.** Darwin-28B-REASON: ~ 2026-Q1. AWAXIS-Think-31B: ~ 2026-Q1.
---
## License
Darwin-60B-DUO inherits the **Gemma Terms of Use** as its effective combined license β the more restrictive of the two constituent licenses.
| Constituent | License |
|-------------|---------|
| Darwin-28B-REASON | Apache-2.0 |
| AWAXIS-Think-31B | Gemma (inherited from Gemma-4) |
| **Darwin-60B-DUO (aggregate)** | **Gemma** (combined-license inheritance) |
The orchestrator code (`gateway/`, `docker/`) is offered under Apache-2.0 to maximize developer flexibility; combined-license inheritance applies to served model behavior only.
Please review the [Gemma Terms of Use](https://ai.google.dev/gemma/terms) and the [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy) before commercial deployment.
---
## Acknowledgments
- **FINAL-Bench team** β Darwin family architecture and DUO concept
- **Anserwise Korean specialist team** β AWAXIS-Think-31B development
- **VIDRAFT** β orchestration framework and the Hybrid-A routing strategy
- **Google DeepMind** β Gemma-4 foundation
- **Cohere and Qwen teams** β Command A+ / Qwen3.5 foundation lineage
---
## Contact
- Hugging Face org: [FINAL-Bench](https://huggingface.co/FINAL-Bench)
- Sister orgs: [Anserwise](https://huggingface.co/Anserwise) Β· [VIDraft](https://huggingface.co/VIDraft)
- Issues and discussions: please open a thread on the **Community** tab of this repository.
---
## Citation
```bibtex
@misc{darwin60b-duo-2026,
title = {Darwin-60B-DUO: A single-endpoint DUO of an English-reasoning SOTA
and a Korean SOTA via Hybrid-A orchestration},
author = {FINAL-Bench Team and Anserwise Team and VIDRAFT},
year = {2026},
howpublished = {Hugging Face},
url = {https://huggingface.co/FINAL-Bench/Darwin-60B-DUO}
}
```