Uncensored version of Qwen/Qwen3.6-35B-A3B with refusal behavior removed via abliteration (norm-preserving orthogonalization). Zero refusals on harmful prompts. No false refusals on harmless prompts. Abliteration identifies the "refusal direction" in the model's residual stream — the linear direction that activates when the model decides to refuse — and surgically removes it from all output projection weights using norm-preserving orthogonalization. 1. Collect residual stream activations (last token position) for 512 harmful + 512 harmless prompts across all 40 layers 2. Compute mean difference vector per layer → this is the "refusal direction" candidate 3. Score layers by…
Open-weight model · Text generation
dolphin-2.9.1-yi-1.5-34b
by Dolphin dphn/dolphin-2.9.1-yi-1.5-34b
Curated and trained by Eric Hartford, Lucas Atkins, and Fernando Fernandes, and Cognitive Computations This is our most spectacular outcome ever. FFT, all parameters, 16bit. 77.4 MMLU on 34b. And it talks like a dream.
Parameters34.4B
Context8,192
Weights68.8 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads4.8M
Runs On
What it takes to serve dolphin-2.9.1-yi-1.5-34b (34.4B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 68.8 GB | 82.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x MI325X $2.00 · 1x MI355X $2.59 |
| 8-bit | 34.4 GB | 41.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 17.2 GB | 20.6 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
SAVRN's Notes on dolphin-2.9.1-yi-1.5-34b
Check the context window first on this one: 8,192 tokens, which decides whether your documents fit before memory comes up. The 34.4B parameters need 82.5 GB at 16-bit, 41.3 GB at 8-bit and 20.6 GB at 4-bit, and the cheapest setup the Index lists at every precision is one MI300X with 192 GB at $1.85 an hour on-demand. The weights ship in bfloat16 across 24 files totaling 68.8 GB, so plan the disk before the card.
Apache 2.0 on the listing allows commercial use, modification and redistribution with notices kept and changes stated, and the page records a derivation from 01-ai/Yi-1.5-34B, so read the terms on that base before deploying. The training sets show where the tune was pointed: dolphin-coder, CodeFeedback-Filtered-Instruction, orca-math-word-problems-200k and function-calling-chatml alongside OpenHermes-2.5 and samantha-data. No Index host prices this model by the token, so the $1.85 card is the only price to compare.
Model Card
By Dolphin, published under apache-2.0, revision 0141cba238d0.
Curated and trained by Eric Hartford, Lucas Atkins, and Fernando Fernandes, and Cognitive Computations
This is our most spectacular outcome ever. FFT, all parameters, 16bit. 77.4 MMLU on 34b. And it talks like a dream.
Although the max positional embeddings is 4k, we used rope theta of 1000000.0 and we trained with sequence length 8k. We plan to train on the upcoming 32k version as well.
Website: https://dphn.ai
Twitter: https://x.com/dphnAI
Web Chat: https://chat.dphn.ai
Telegram bot: https://t.me/DolphinAI_bot
Our appreciation for the sponsors of Dolphin 2.9.1: - Crusoe Cloud - provided excellent on-demand 8xH100 node - OnDemand - provided inference sponsorship
This model is based on Yi-1.5-34b, and is governed by apache 2.0 license.
The base model has 4k context, but we used rope theta of 1000000.0 and the full-weight fine-tuning was with 8k sequence length.
Dolphin 2.9.1 uses ChatML prompt template format.
example:
<|im_start|>system
You are Dolphin, a helpful AI assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
Configuration
- Architecture
- LlamaForCausalLM
- Context length (tokens)
- 8,192
- Layers
- 60
- Hidden size
- 7,168
- Feed-forward size
- 20,480
- Attention heads
- 56
- Key/value heads
- 8
- Vocabulary size
- 64,000
- RoPE base
- 5e+06
- Stored precision
- bfloat16
- Model type
- llama
Identity and Version
- Repository
- dphn/dolphin-2.9.1-yi-1.5-34b
- Publisher
- Dolphin
- Task
- Text generation
- Modality
- Text
- Library
- transformers
- Parameters
- 34.4B parameters
- Languages
- Not stated by the source
- Revision
- 0141cba238d0faad09bc240ea7af14c9ea5aec44
- First published
- 2024-05-18
- Last updated
- 2025-09-08
Files and Weights
24 files, 68.8 GB in total. The weights are 15 files totalling 68.8 GB in safetensors.
Weights15 files · 68.8 GB
Configuration4 files · 46.2 KB
Tokenizer3 files · 4.6 MB
Documentation1 file · 7.6 KB
Repository1 file · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00015.safetensors | Weights | 4.8 GB | 9a10cac63db6 |
| model-00002-of-00015.safetensors | Weights | 4.8 GB | cb759e4d3baf |
| model-00003-of-00015.safetensors | Weights | 5.0 GB | 3abc57298674 |
| model-00004-of-00015.safetensors | Weights | 4.8 GB | 9f36d3cbe5ec |
| model-00005-of-00015.safetensors | Weights | 4.8 GB | fea9b665977a |
| model-00006-of-00015.safetensors | Weights | 5.0 GB | 63dc988a58eb |
| model-00007-of-00015.safetensors | Weights | 4.8 GB | ef1f23365c32 |
| model-00008-of-00015.safetensors | Weights | 4.8 GB | 9cf2b792a677 |
| model-00009-of-00015.safetensors | Weights | 5.0 GB | 4cfec1f95efb |
| model-00010-of-00015.safetensors | Weights | 4.8 GB | 03ea73514626 |
| model-00011-of-00015.safetensors | Weights | 4.8 GB | 23ab68eb8de3 |
| model-00012-of-00015.safetensors | Weights | 5.0 GB | 08f6cf39245a |
| model-00013-of-00015.safetensors | Weights | 4.8 GB | 7e400169e476 |
| model-00014-of-00015.safetensors | Weights | 4.8 GB | 474880d9ef9c |
| model-00015-of-00015.safetensors | Weights | 1.2 GB | 48c76af5810f |
| config.json | Configuration | 705 B | — |
| generation_config.json | Configuration | 158 B | — |
| model.safetensors.index.json | Configuration | 44.8 KB | — |
| special_tokens_map.json | Configuration | 570 B | — |
| README.md | Documentation | 7.6 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
| tokenizer.json | Tokenizer | 3.6 MB | — |
| tokenizer.model | Tokenizer | 1.0 MB | 386c49cf943d |
| tokenizer_config.json | Tokenizer | 1.7 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 68.8 GB
Download from Dolphin
Released by Dolphin through its official repository on Hugging Face. Read the license.
Built From
- Derived from 01-ai/Yi-1.5-34B
- Trained on (disclosed) Locutusque/function-calling-chatml
- Trained on (disclosed) cognitivecomputations/Dolphin-2.9
- Trained on (disclosed) cognitivecomputations/dolphin-coder
- Trained on (disclosed) cognitivecomputations/samantha-data
- Trained on (disclosed) internlm/Agent-FLAN
- Trained on (disclosed) m-a-p/CodeFeedback-Filtered-Instruction
- Trained on (disclosed) microsoft/orca-math-word-problems-200k
- Trained on (disclosed) teknium/OpenHermes-2.5
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 68.8 GB |
| 16-bit | 68.8 GB |
| 8-bit | 34.4 GB |
| 4-bit | 17.2 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Compare dolphin-2.9.1-yi-1.5-34b
Questions About dolphin-2.9.1-yi-1.5-34b
How much GPU memory does dolphin-2.9.1-yi-1.5-34b need?
About 82.5 GB at 16-bit and 20.6 GB at 4-bit: the weights (34.4B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run dolphin-2.9.1-yi-1.5-34b on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use dolphin-2.9.1-yi-1.5-34b commercially?
Yes. dolphin-2.9.1-yi-1.5-34b is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
What is dolphin-2.9.1-yi-1.5-34b's context length?
8,192 tokens, from the maximum position embeddings in its published configuration.
Similar Models
Qwen2.5 is the latest series of Qwen large language models. For Qwen2.5, we release a number of base language models and instruction-tuned language models ranging from 0.5 to 72 billion parameters. Qwen2.5 brings the following improvements upon Qwen2: - Significantly more knowledge and has greatly improved capabilities in coding and mathematics, thanks to our specialized expert models in these domains. - Significant improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g, tables), and generating structured outputs especially JSON. More resilient to the diversity of system prompts, enhancing role-play implementation and…
Qwen2.5-Coder is the latest series of Code-Specific Qwen large language models (formerly known as CodeQwen). As of now, Qwen2.5-Coder has covered six mainstream model sizes, 0.5, 1.5, 3, 7, 14, 32 billion parameters, to meet the needs of different developers. Qwen2.5-Coder brings the following improvements upon CodeQwen1.5: - Significantly improvements in code generation, code reasoning and code fixing. Base on the strong Qwen2.5, we scale up the training tokens into 5.5 trillion including source code, text-code grounding, Synthetic data, etc. Qwen2.5-Coder-32B has become the current state-of-the-art open-source codeLLM, with its coding abilities matching those of GPT-4o. - A more…
Qwen2.5-Coder is the latest series of Code-Specific Qwen large language models (formerly known as CodeQwen). As of now, Qwen2.5-Coder has covered six mainstream model sizes, 0.5, 1.5, 3, 7, 14, 32 billion parameters, to meet the needs of different developers. Qwen2.5-Coder brings the following improvements upon CodeQwen1.5: - Significantly improvements in code generation, code reasoning and code fixing. Base on the strong Qwen2.5, we scale up the training tokens into 5.5 trillion including source code, text-code grounding, Synthetic data, etc. Qwen2.5-Coder-32B has become the current state-of-the-art open-source codeLLM, with its coding abilities matching those of GPT-4o. - A more…
Qwen2.5 is the latest series of Qwen large language models. For Qwen2.5, we release a number of base language models and instruction-tuned language models ranging from 0.5 to 72 billion parameters. Qwen2.5 brings the following improvements upon Qwen2: - Significantly more knowledge and has greatly improved capabilities in coding and mathematics, thanks to our specialized expert models in these domains. - Significant improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g, tables), and generating structured outputs especially JSON. More resilient to the diversity of system prompts, enhancing role-play implementation and…
A fast and efficient 32B model optimized for CPU inference. The model was refactored with BitNet features and an updated tokenizer that includes new Routing, Media, Vision, Sound, Tool call, and Robotics tags. Built on a DeepSeek R1-32B architecture with native ternary (BitNet-style) support and ready-to-run GGUF quantizations. - JiRack is a cloud-ready model that helps save money on cloud infrastructure. It can be used as an expert model in RAG deployments, with the ONNX JiRack Java server as an alternative. - Benefits high quality CPU inference TQ2 on Llama.cpp and Ollama via QAT - Robotcs, Routing, Coding, Multimedia, Advanced tool calling via CMSManhattan/JiRackPrecisionTokenizer…