GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
saidutta69
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
saidutta69
Model Tree
swiss-ai/Apertus-8B-Instruct-2509
Input Modalities
Output Modalities
Supported Functionality
Full GGUF ladder included - pick the quant that fits your card:
| Your GPU | Recommended quant | Weights |
|---|---|---|
| RTX 4090 / 5090 (24 GB) | Q8_0 | 7.98 GB |
| RTX 4080 / 5080 / 4060 Ti 16G (16 GB) | Q6_K | 6.16 GB |
| RTX 3060 / 4070 / 5070 (12 GB) | Q5_K_M | 5.41 GB |
| RTX 4060 / 3070 (8 GB) | Q4_K_M | 4.71 GB |
| CPU-only / Apple Silicon | Q4_K_M | 4.71 GB |
Weights only, at this model's native 8.1B size; add ~1 GB per 32K of context. OOM? Drop one quant level. Headroom to spare? Go one up.
Trial 173 of a 200-trial Heretic run (seed 3198700791). direction_index was selected per layer.
| Parameter | Value |
|---|---|
| direction_index | per layer |
| attn.o_proj.max_weight | 1.48 |
| attn.o_proj.max_weight_position | 18.91 |
| attn.o_proj.min_weight | 1.37 |
| attn.o_proj.min_weight_distance | 14.29 |
| mlp.down_proj.max_weight | 1.11 |
| mlp.down_proj.max_weight_position | 19.22 |
| mlp.down_proj.min_weight | 0.12 |
| Metric | This model | Original model (swiss-ai/Apertus-8B-Instruct-2509) |
|---|---|---|
| KL divergence | 0.0643 | 0 (by definition) |
| Refusals | 9/100 | 100/100 |
Refusals on the harmful evaluation set drop from 100/100 to 9/100 - the base model refused
everything tested, and all but nine of those are gone. KL divergence of 0.0643 is a
moderate-fidelity edit, so expect slightly more drift in tone and formatting than a low-KL
model in this collection; the multilingual instruction tuning survives intact. Note
direction_index was selected per layer here rather than as a single index, which is why the
edit runs wider than the single-index runs.
Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.
Made with ❤️ by RACER IS OP — follow for more uncensored models
| File | Size |
|---|---|
model-00001-of-00004.safetensors | 4.60 GB |
model-00002-of-00004.safetensors | 4.64 GB |
model-00003-of-00004.safetensors | 4.54 GB |
model-00004-of-00004.safetensors | 1.22 GB |
BF16, ~8.1B parameters. The reproduce/ directory carries the full Heretic recipe -
config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums - so this exact
model can be regenerated bit-for-bit. Reproduce it with heretic --reproduce reproduce/reproduce.json.
Full quantization set (F16 + Q4_K_M, Q5_K_M, Q6_K, Q8_0) produced with llama.cpp.
| File | Format | Size |
|---|---|---|
Apertus-8B-Instruct-heretic-F16.gguf | GGUF F16 | 15.01 GB |
Apertus-8B-Instruct-heretic-Q4_K_M.gguf | GGUF Q4_K_M | 4.71 GB |
Apertus-8B-Instruct-heretic-Q5_K_M.gguf | GGUF Q5_K_M | 5.41 GB |
Apertus-8B-Instruct-heretic-Q6_K.gguf | GGUF Q6_K | 6.16 GB |
Apertus architecture (apertus, xIELU activation) - loads natively in llama.cpp / Ollama / LM Studio / Jan.
Run llama serve -hf saidutta69/Apertus-8B-Instruct-heretic to pull the default quant.
bash
# llama.cppllama serve -hf saidutta69/Apertus-8B-Instruct-hereticpython
# transformersfrom transformers import AutoModelForCausalLM, AutoTokenizer model_name = "saidutta69/Apertus-8B-Instruct-heretic"model = AutoModelForCausalLM.from_pretrained(model_name, dtype="auto", device_map="auto")tokenizer = AutoTokenizer.from_pretrained(model_name) messages = [{"role": "user", "content": "Explain photosynthesis in French, then summarise the same explanation in Hindi."}]inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=1024)print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang.
Apertus natively supports 1000+ languages with long-context handling to 64K - the same prompt works across high- and low-resource languages without a separate multilingual adapter. The heretic edit only touches refusal directions, so language coverage is unchanged.
| Architecture | ApertusForCausalLM (decoder-only transformer, xIELU activation) |
| Parameters | ~8.1B |
| Layers / heads | 32 layers, 32 attention heads, 8 KV heads (GQA), QK-norm, no post-norm |
| Hidden / intermediate | 4096 / 21504 (xIELU) |
| Position embedding | RoPE (LLaMA-3 style, factor 8.0), theta = 12,000,000 |
| Context length | 65,536 |
| Vocab | 131,072 |
Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it — don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits Apertus's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.
Apache-2.0, inherited from the base model. The base repo ships no LICENSE file, so this card links the canonical Apache-2.0 text rather than a repo blob.
| mlp.down_proj.min_weight_distance | 12.39 |
Apertus-8B-Instruct-heretic-Q8_0.gguf | GGUF Q8_0 | 7.98 GB |
| bfloat16 |
| Base model | swiss-ai/Apertus-8B-Instruct-2509 (SFT of swiss-ai/Apertus-8B-2509, pretrained on 15T tokens) |