# Darwin-27B-RSI

> Qwen3.5-27B family · 27B dense · Thinking mode · BF16 · Apache 2.0 No human-written answers. The model generated its own learning signal — and got measurably better.

- **Organisation:** FINAL-Bench
- **Category:** Chat & Text Generation, Reasoning Models
- **Parameters:** 26.9B
- **Context length:** 262,144 tokens
- **Licence:** Apache 2.0 (Commercial use OK)
- **Released:** 2026-09-27
- **Hugging Face:** 388 downloads (30 days), 50 likes

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q2_K | 11.15 |
| Q3_K_M | 15.35 |
| Q4_K_M | 18.45 |
| Q5_K_M | 21.21 |
| Q6_K | 24.17 |
| Q8_0 | 30.45 |
| F16 | 55.1 |

## Runs with

- [llama.cpp](https://selfhost.directory/project/llamacpp)
- [LM Studio](https://lmstudio.ai)
- [SGLang](https://selfhost.directory/project/sglang)
- [Transformers](https://github.com/huggingface/transformers)
- [vLLM](https://selfhost.directory/project/vllm)
- [Jan](https://jan.ai) (community)
- [KoboldCpp](https://selfhost.directory/project/koboldcpp) (community)
- [LocalAI](https://selfhost.directory/project/localai) (community)
- [Ollama](https://selfhost.directory/project/ollama) (community)

## About

Qwen3.5-27B family · 27B dense · Thinking mode · BF16 · Apache 2.0 No human-written answers. The model generated its own learning signal — and got measurably better.

---

Darwin-27B-RSI is Darwin-27B-Opus after Recursive Self-Improvement (RSI): the model was improved using only signal it produced itself. During self-improvement, the model itself (its weights) improves by learning only from its own solutions. No human-written solutions or reasoning traces are used; correctness is checked automatically (agreement across its own samples and code execution).

Under an identical evaluation protocol, Darwin-27B-RSI improves over its parent on graduate-level science reasoning — +5.24 points on GPQA Diamond (single sample) and +3.79 points with majority voting — with every gain statistically significant in paired tests.

As the reasoning engine of Darwin-27B-JEV on the Decision Index, it lifts the hardest reasoning decisions: GPQA Diamond skill 0.31 → 0.71, GSM8K 0.61 → 0.97, MMLU-Pro 0.60 → 0.82.

---

Darwin-27B-RSI is Model-level RSI: the model itself (its weights) improves by learning only from its own solutions. No human-written solutions or reasoning traces are used; correctness is checked automatically. You download a new model file, and it is smarter on its own.

## Alternatives

- [nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16](https://selfhost.directory/models/nvidia/nvidia-nemotron-3.5-lightning-30b-a3b-bf16) (31.6B (3B active), openmdw-1.1, ≈19.63 GB at Q4): NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.
- [farbodtavakkoli/OTel-2.0-LLM-31B-IT](https://selfhost.directory/models/farbodtavakkoli/otel-2.0-llm-31b-it) (31.3B, Apache 2.0, ≈26.54 GB at Q4): Versioning notice: This card describes the current OTel 2.0 OSFT checkpoint. The model weights may be updated over time. For reproducible evaluation or production deployment, pin a specific model revision, checkpoint hash, or release tag rather than relying on the floating latest version.
- [ornith-ai/Ornith-1.0-35B](https://selfhost.directory/models/ornith-ai/ornith-1.0-35b) (35B (3.8B active), MIT, ≈21.87 GB at Q4): Aloha! 🌺 Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding.
- [zai-org/GLM-4.7-Flash](https://selfhost.directory/models/zai-org/glm-4.7-flash) (31.2B (4.6B active), MIT, ≈26.35 GB at Q4): 👋 Join our Discord community. 📖 Check out the GLM-4.7 technical blog , technical report(GLM-4.5) . 📍 Use GLM-4.7-Flash API services on Z.ai API Platform. 👉 One click to GLM-4.7 .
- [poolside/Laguna-XS-2.1](https://selfhost.directory/models/poolside/laguna-xs-2.1) (33.4B (2.2B active), Custom licence, ≈21.58 GB at Q4): Laguna XS 2.1 is a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token designed for agentic coding and long-horizon work on a local machine.
- [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://selfhost.directory/models/nvidia/nvidia-nemotron-3-nano-30b-a3b-bf16) (31.6B (3B active), nvidia-nemotron-open-model-license, ≈19.63 GB at Q4): Nemotron-3-Nano-30B-A3B-BF16 is a large language model (LLM) trained from scratch by NVIDIA, and designed as a unified model for both reasoning and non-reasoning tasks. It responds to user queries and tasks by first generating a reasoning trace and then concluding with a final response.

More: https://selfhost.directory/models/final-bench/darwin-27b-rsi/alternatives

## Runs well on

- [Tesla P40](https://selfhost.directory/models/final-bench/darwin-27b-rsi/on/tesla-p40): Q5_K_M, est. 10 tok/s
- [Radeon RX 7900 XTX](https://selfhost.directory/models/final-bench/darwin-27b-rsi/on/rx-7900-xtx): Q5_K_M, est. 28 tok/s
- [Instinct MI50 32GB](https://selfhost.directory/models/final-bench/darwin-27b-rsi/on/instinct-mi50): Q6_K, est. 25 tok/s
- [GeForce RTX 5090](https://selfhost.directory/models/final-bench/darwin-27b-rsi/on/rtx-5090): Q6_K, est. 44 tok/s
- [Apple M4 Pro (48 GB)](https://selfhost.directory/models/final-bench/darwin-27b-rsi/on/apple-m4-pro-48gb): Q8_0, est. 5.3 tok/s
- [RTX A6000](https://selfhost.directory/models/final-bench/darwin-27b-rsi/on/rtx-a6000): Q8_0, est. 15 tok/s

## Similar size

- [openai/gpt-oss-20b](https://selfhost.directory/models/openai/gpt-oss-20b) (20.9B)
- [zai-org/GLM-4.7-Flash](https://selfhost.directory/models/zai-org/glm-4.7-flash) (31.2B)
- [Qwen/Qwen3-32B](https://selfhost.directory/models/qwen/qwen3-32b) (32.8B)
- [Qwen/Qwen3-30B-A3B](https://selfhost.directory/models/qwen/qwen3-30b-a3b) (30.5B)
- [LiquidAI/LFM2.5-24B-A2B](https://selfhost.directory/models/liquidai/lfm2.5-24b-a2b) (24B)
- [Qwen/Qwen2.5-32B-Instruct](https://selfhost.directory/models/qwen/qwen2.5-32b-instruct) (32.8B)

## FAQ

**How much VRAM does Darwin-27B-RSI need?**
About 18 GB at Q4_K_M with an 8K context. Q8_0 needs about 30 GB and full precision about 55 GB. Longer contexts need more memory for the KV cache.

**Can Darwin-27B-RSI run on the Tesla P40?**
Yes. It fits at Q5_K_M (about 23 GB), at an estimated 10 tokens/s. It fits comfortably on the Radeon RX 7900 XTX, Instinct MI50 32GB, RTX 5090.

**Can I use Darwin-27B-RSI commercially?**
Yes. It is released under Apache 2.0. Permissive: use, modify and sell freely; keep the notice and licence text.

**What is the context length of Darwin-27B-RSI?**
262,144 tokens (256K). The full window needs far more memory than an 8K context.

**How do I run Darwin-27B-RSI with Ollama or llama.cpp?**
With Ollama: ollama run hf.co/FINAL-Bench/Darwin-27B-RSI-GGUF:Q4_K_M. With llama.cpp: llama-server -hf FINAL-Bench/Darwin-27B-RSI-GGUF:Q4_K_M, which starts an OpenAI-compatible server.

## Links

- huggingface: https://huggingface.co/FINAL-Bench/Darwin-27B-RSI
- demo: https://huggingface.co/spaces/multimodalart/jev-decision-index

---
Source: https://selfhost.directory/models/final-bench/darwin-27b-rsi