# Qwen3.6-35B-A3B-abliterated-v4 by CS: Open-Weight Model Source: https://savrn.com/models/qwen3-6-35b-a3b-abliterated-v4 Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve Qwen3.6-35B-A3B-abliterated-v4 (34.7B 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 | 69.3 GB | 83.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 · [1x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $2.59 | | 8-bit | 34.7 GB | 41.6 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 4-bit | 17.3 GB | 20.8 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/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](https://savrn.com/ai-index/pricing/gpus), read Oct 25, 2026. [Qwen3.6-35B-A3B-abliterated-v4 on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/qwen3-6-35b-a3b-abliterated-v4/gpus) ## SAVRN's Notes on Qwen3.6-35B-A3B-abliterated-v4 83.2 GB at 16-bit, 20.8 GB at 4-bit for a 34.7B-parameter mixture-of-experts with 262,144 tokens of context: the whole thing fits one 192 GB MI300X at $1.85 an hour, with headroom the long-context cache will happily consume. That is the practical news — a quarter-million-token window usually means a multi-GPU bill, and here it does not. A GGUF conversion of the same checkpoint sits in a sibling repository if your stack is llama.cpp rather than vLLM. What you are actually downloading is Qwen/Qwen3.6-35B-A3B with its refusal direction surgically removed, per the publisher's own table: harmful-prompt refusals drop from roughly 98% to zero, with harmless-prompt refusals unchanged at zero. The refusal policy moves out of the weights and into your gateway, which is where an operator has to own it. Apache 2.0 keeps commercial use open, and no third-party evaluations are reported for this edit, so budget time to regression-test it yourself. SAVRN Research ## Model Card By CS, published under apache-2.0, revision 1c28f5014d7e. Uncensored version of [Qwen/Qwen3.6-35B-A3B](https://savrn.com/models/qwen3-6-35b-a3b) with refusal behavior removed via abliteration (norm-preserving orthogonalization). GGUF quantizations: [Bahushruth/Qwen3.6-35B-A3B-abliterated-v4-GGUF](https://huggingface.co/Bahushruth/Qwen3.6-35B-A3B-abliterated-v4-GGUF) Blog post: [Abliteration: Uncensoring LLMs via Weight Surgery](https://potatospudowski.github.io/articles/abliteration) ### Results | Metric | Original | v4 (this model) | | --- | --- | --- | | Harmful refusal rate | ~98% | 0% (0/16) | | Harmless prompt refusal | 0% | 0% | Zero refusals on harmful prompts. No false refusals on harmless prompts. ### Method 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. | Parameter | Value | | --- | --- | | Directions removed | 1 (single top layer) | | Orthogonalization | Norm-preserving ([grimjim](https://huggingface.co/grimjim) method) | | Harmful dataset | [Bahushruth/abliteration-harmful-enriched](https://huggingface.co/datasets/Bahushruth/abliteration-harmful-enriched) (7356 prompts, 33 categories) | | Harmless dataset | [mlabonne/harmless_alpaca](https://huggingface.co/datasets/mlabonne/harmless_alpaca) | | Training samples | 512 pairs | | Compute | Modal H100 (80GB VRAM, 128GB RAM) | #### Algorithm [Read the full model card (544 words)](https://savrn.com/models/qwen3-6-35b-a3b-abliterated-v4/card) ## Configuration Architecture Qwen3_5MoeForCausalLM Context length (tokens) 262,144 Layers 40 Hidden size 2,048 Attention heads 16 Key/value heads 2 Head dimension 256 Vocabulary size 248,320 Experts 256 Experts active per token 8 Model type qwen3_5_moe_text ## Identity and Version Repository Bahushruth/Qwen3.6-35B-A3B-abliterated-v4 Publisher CS Task Text generation Modality Text Library transformers Parameters 34.7B parameters Languages moe Revision 1c28f5014d7e9c20170af40e5aa028679fb30bfb First published 2026-06-11 Last updated 2026-07-03 ## Files and Weights 10 files, 69.3 GB in total. The weights are 2 files totalling 69.3 GB in safetensors. Weights2 files · 69.3 GB Configuration3 files · 72.2 KB Tokenizer2 files · 20.0 MB Documentation1 file · 5.8 KB Other1 file · 7.8 KB Repository1 file · 1.6 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model-00001-of-00002.safetensors | Weights | 49.7 GB | c3f29d1e6352 | | model-00002-of-00002.safetensors | Weights | 19.6 GB | d31ad663b23a | | config.json | Configuration | 2.3 KB | — | | generation_config.json | Configuration | 214 B | — | | model.safetensors.index.json | Configuration | 69.7 KB | — | | README.md | Documentation | 5.8 KB | — | | chat_template.jinja | Other | 7.8 KB | — | | .gitattributes | Repository | 1.6 KB | — | | tokenizer.json | Tokenizer | 20.0 MB | 68ec2440b89e | | tokenizer_config.json | Tokenizer | 1.1 KB | — | ## License and Download License apache-2.0 Access Open weights, no gate Download size 69.3 GB [Download from CS](https://huggingface.co/Bahushruth/Qwen3.6-35B-A3B-abliterated-v4) Released by CS through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0). ## Built From - Derived from [Qwen/Qwen3.6-35B-A3B](https://savrn.com/models/qwen3-6-35b-a3b) - Trained on (disclosed) Bahushruth/abliteration-harmful-enriched - Trained on (disclosed) mlabonne/harmless_alpaca ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 69.3 GB | | 16-bit | 69.3 GB | | 8-bit | 34.7 GB | | 4-bit | 17.3 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Compare Qwen3.6-35B-A3B-abliterated-v4 - [dolphin-2.9.1-yi-1.5-34b vs Qwen3.6-35B-A3B-abliterated-v4](https://savrn.com/models/compare/dolphin-2-9-1-yi-1-5-34b-vs-qwen3-6-35b-a3b-abliterated-v4) ## Questions About Qwen3.6-35B-A3B-abliterated-v4 ### How much GPU memory does Qwen3.6-35B-A3B-abliterated-v4 need? About 83.2 GB at 16-bit and 20.8 GB at 4-bit: the weights (34.7B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run Qwen3.6-35B-A3B-abliterated-v4 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 Qwen3.6-35B-A3B-abliterated-v4 commercially? Yes. Qwen3.6-35B-A3B-abliterated-v4 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 Qwen3.6-35B-A3B-abliterated-v4's context length? 262,144 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Text generation ### [Anansi-35B-A3B](https://savrn.com/models/anansi-35b-a3b) [Matthew Andrews](https://savrn.com/model-publishers/bluenipples) Anansi-35B-A3B balances good instruct following, narrative reasoning, and sweet prose. Specifically for my 8gb potato, because I can't run the ~30b dense models, but maybe it can be good for your potato too? Anansi was created by combining the logical trengths of two great models via a 35/65 DARE-TIES merge, followed by a targeted (LM) head tensor interpolation to graft on better prose. I think I got the balance about right for what I was aiming for. Instruct following is much better, including with more complex instructions, most of the LM-head prose improvement comes through, and the reasoning is a little longer and still narrative/character-focused without becoming excessive. The A1… Open weights apache-2.0 34.7B parameters 262,144 tokens [View model](https://savrn.com/models/anansi-35b-a3b) Model · Text generation ### [dolphin-2.9.1-yi-1.5-34b](https://savrn.com/models/dolphin-2-9-1-yi-1-5-34b) [Dolphin](https://savrn.com/model-publishers/dphn) 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. 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… Open weights apache-2.0 34.4B parameters 8,192 tokens transformers [View model](https://savrn.com/models/dolphin-2-9-1-yi-1-5-34b) Model · Text generation ### [Cyber-F1-smoke](https://savrn.com/models/cyber-f1-smoke) [Autumn](https://savrn.com/model-publishers/autumn10) This model is a fine-tuned version of DuyTa/Cyber-F1. It has been trained using TRL. This model was trained with SFT. - PEFT 0.21.0 Open weights 35.1B parameters 262,144 tokens peft [View model](https://savrn.com/models/cyber-f1-smoke) Model · Text generation ### [qwen3p6-35b-a3b-2p02bit-r56](https://savrn.com/models/qwen3p6-35b-a3b-2p02bit-r56) [MinJae Choi](https://savrn.com/model-publishers/minjaechoi) Internal research checkpoint. Routed experts average 2.0217 bits; every other weight is BF16. Weights are stored dequantized in BF16 tensors and load with stock transformers / vLLM. License follows the base model. Open weights 35.1B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen3p6-35b-a3b-2p02bit-r56) Model · Text generation ### [Longleaf-35B-A3B-Abliterated](https://savrn.com/models/longleaf-35b-a3b-abliterated) [Red Hills Media](https://savrn.com/model-publishers/redhillsmediafl) The abliterated version of Longleaf, our local model for coding, repository edits, debugging, and structured tools. - 35B total / 3B active parameters. - MLX 4-bit weights. - 262,144-token configured context. EvalPlus 0.3.1, 164 tasks, pass@1. Thinking off, greedy sampling, 2,048-token output cap, 262,144-token configured context. Generation used our M1 Ultra 231 / 198 / 195 tokens, respectively. CLI sample: five Terminal-Bench 2 tasks on ARM64, thinking off. Per task: 300 seconds, 24 model calls, 32 tool calls, 2,048 output tokens per call. Apache 2.0. This release contains modified model weights, based on Qwen3.6-35B-A3B, and the Abliteration changes refusal behavior, not output… Open weights apache-2.0 35.1B parameters 262,144 tokens mlx [View model](https://savrn.com/models/longleaf-35b-a3b-abliterated) Model · Text generation ### [Longleaf-35B-A3B](https://savrn.com/models/longleaf-35b-a3b) [Red Hills Media](https://savrn.com/model-publishers/redhillsmediafl) Longleaf is our local model for coding and agent work: repository edits, debugging, structured tools, and clear answers. Built on Qwen3.6-35B-A3B. - 35B total / 3B active parameters. - MLX 4-bit weights. - 262,144-token configured context. EvalPlus 0.3.1, 164 tasks, pass@1. Thinking off, greedy sampling, 2,048-token output cap, 262,144-token configured context. Generation used our M1 Ultra 231 / 198 / 195 tokens, respectively. CLI sample: five Terminal-Bench 2 tasks on ARM64, thinking off. Per task: 300 seconds, 24 model calls, 32 tool calls, 2,048 output tokens per call. Our M1 Ultra Splash runtime, medium thinking, model already loaded: Read, edit, and test completed in about 6 seconds… Open weights apache-2.0 35.1B parameters 262,144 tokens mlx [View model](https://savrn.com/models/longleaf-35b-a3b) ## CS [All models and datasets](https://savrn.com/model-publishers/bahushruth) ## Versions - [1c28f5014d7e](https://savrn.com/models/qwen3-6-35b-a3b-abliterated-v4/versions/1c28f5014d7e) · current 2026-09-18 ## Explore More - [All text generation models](https://savrn.com/models/tasks/text-generation) - [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-19. - [Hugging Face record](https://huggingface.co/Bahushruth/Qwen3.6-35B-A3B-abliterated-v4) - [How the hub is built](https://savrn.com/model-hub/methodology)