Qwythos-27B-v1¶

Development line: project:qwythos-27b-v1 · thread qwythos-27b-v1-release
Last event: 2026-08-03 · 1 dated since 2026-08-03 · Researched: 2026-09-05 · confidence: medium

What it is¶

Qwythos-27B-v1 is an open-weight bf16 checkpoint on Qwen3.5-27B for developers who need a local model with function calling and image input.

• Context window: provides 1,048,576 tokens through YaRN 4.0.
• Vision tower and native multi-token prediction: preserved from the base architecture.
• Post-training pipeline: full SFT → DPO → ESFT as v1 before RL.

The native window is 262,144 tokens; the 1M configuration degrades short-context quality.

The checkpoint fits self-hosted agent and tool workflows, but the vision tower was not fine-tuned and sensitive deployments require external review.

Development line¶

• 2026-08-03 — Qwythos-27B-v1 linked on Hugging Face. Third parties uploaded FancieF/Qwythos-27B-v1-MLX-VLM-4bit, FancieF/Qwythos-27B-v1-MLX-VLM-bf16 and a separate MTP draft; this is not a new official checkpoint.

What changed¶

2026-07-29 — Empero released Qwythos-27B-v1 and an official GGUF set: a 27B full-parameter model on Qwen3.5-27B under Apache-2.0, with 1M context, vision and MTP. 2026-08-03 — Third-party MLX ports of the base Qwythos-27B-v1 appeared: FancieF/Qwythos-27B-v1-MLX-VLM-4bit, FancieF/Qwythos-27B-v1-MLX-VLM-bf16 and a separate MTP draft; this is not a new official checkpoint.

How to use this¶

From 2026-08-03, practitioners should use the linked Hugging Face model page as the recorded reference for Qwythos-27B-v1 and verify its model-card details and availability there before use.

1. For local deployment on a 24 GB GPU, start with official GGUF Q4_K_M; add mmproj-Qwythos-27B-F16.gguf for image input. — https://empero.org/writing/qwythos-27b-the-long-awaited-27b
2. For Transformers, load empero-ai/Qwythos-27B-v1 via AutoModelForImageTextToText and use the built-in chat template. — https://huggingface.co/empero-ai/Qwythos-27B-v1
3. For OpenAI-compatible serving, run vLLM or SGLang; near 1M context, vLLM requires VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and explicit --max-model-len. — https://huggingface.co/empero-ai/Qwythos-27B-v1
4. For function calling, pass definitions in tools=[...] to the built-in chat template and handle standard tool call blocks. — https://huggingface.co/empero-ai/Qwythos-27B-v1

Best practices¶

• For agentic and tool use, start with temperature 0.6, top_p 0.95, top_k 20, repetition_penalty 1.05 and a budget of at least 16,384 new tokens. — https://huggingface.co/empero-ai/Qwythos-27B-v1
• If 1M context is not needed, lower the YaRN factor to 2.0 for about 512k or revert to native 262k configuration for better short-context quality. — https://huggingface.co/empero-ai/Qwythos-27B-v1
• For K-quants, pick official GGUF: it keeps the sensitive Gated-DeltaNet state path at Q8_0 or higher instead of quantizing it with the remaining weights. — https://empero.org/writing/qwythos-27b-the-long-awaited-27b
• Do not serve the uncensored checkpoint to users without application-level policy, review and access controls. — https://huggingface.co/empero-ai/Qwythos-27B-v1

Superseded by this¶

• 2026-08-03: treating this date as the release of base Qwythos-27B-v1 is obsolete; the primary announcement was 2026-07-29, while 3 August refers to third-party MLX conversions.

Still unknown¶

• The official model card describes v1 and plans for RL-v2, but primary sources do not confirm that Qwythos-27B-v2 was released by 2026-09-05.
• The 3 August list confirms third-party MLX conversions, but gives no independent evaluation of quality, compatibility or preserved capabilities of the original model.

Sources¶

Agent brief¶

• Subject: project:qwythos-27b-v1, thread qwythos-27b-v1-release, 1 dated events 2026-08-03 → 2026-08-03.
• Practical note: From 2026-08-03, practitioners should use the linked Hugging Face model page as the recorded reference for Qwythos-27B-v1 and verify its model-card details and availability there before use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.