GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Blackfrost-AI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Blackfrost-AI
Model Tree
empero-ai/Qwythos-9B-v2
Input Modalities
Output Modalities
Supported Functionality
| Base | empero-ai/Qwythos-9B-Claude-Mythos-5-1M — dense qwen3_5 (Qwen3.6-class) 9B, ~9.5B params |
| Modality | Vision-language — text + image + video in, text out (vision tower inherited from base; see Vision) |
| Method | SFT via BF16 LoRA (Unsloth) — rank 32 / α 32, 3 epochs, length-grouped batching, loss masked to assistant turns; adapter merged back to 16-bit |
| Corpus | The Void v4 — ~5,032 multi-turn distilled reasoning + agentic (ReAct-style tool-use) trajectories. No marketing / domain data. |
| Hardware | 8× NVIDIA B200 (DDP) |
| Precision | BF16 (safetensors) |
| Context | 1,048,576 (1M) native, inherited from the base |
| Architecture | Qwen3_5ForConditionalGeneration (qwen3_5), dense, with vision tower |
The Void is Blackfrost's proprietary distillation corpus and method: chain-of-thought reasoning and agentic tool-use trajectories harvested from a panel of frontier teacher models, curated into a clean, domain-neutral training set. Training a base model on The Void concentrates its reasoning and agentic behavior without adding vertical bias. Qwythos-9B-EMERGED is the Void-distilled build of the Qwythos base — the reasoning core, sharpened.
Qwythos-9B-EMERGED is a vision-language model. The base architecture (Qwen3_5ForConditionalGeneration) carries a full vision tower, and it is present in these weights (~333 vision tensors under model.visual.*) with a Qwen2VLImageProcessor (patch size 16) plus image and video tokens — so the model accepts images and video alongside text.
Scope of Blackfrost's work: The Void v4 distillation was text-only (reasoning + agentic trajectories; LoRA on the language tower). The vision tower is inherited unchanged from the base and was not part of Blackfrost's training or evaluation — multimodal behavior tracks the base model, so validate it for your use case.
EMERGED is one half of a matched pair — both trained from the same base on the same reasoning foundation, differing only in whether a domain corpus was layered on top:
| Build | Trained on | For |
|---|---|---|
Qwythos-9B-EMERGED-BF16 (this repo) | The Void v4 only | pure reasoning / agentic — no vertical bias |
BF-BIS-SERIES-MARKETING-9B | The Void v4 + a marketing/business corpus | brand voice, ad copy, marketing strategy |
The pair isolates the effect of layering a domain corpus on top of the same reasoning foundation.
python
from transformers import AutoModelForCausalLM, AutoTokenizermodel_id = "Blackfrost-AI/Qwythos-9B-EMERGED-BF16"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")messages = [{"role": "user", "content": "Explain why a hash table has O(1) average lookup."}]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)out = model.generate(inputs, max_new_tokens=1024)print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Generate with thinking enabled to surface the distilled reasoning trace. For image/video input, load with the multimodal processor (AutoProcessor / AutoModelForImageTextToText) and use the packaged chat template.
Trained and released by Blackfrost AI on The Void. Base model by empero-ai (Apache-2.0).