- An independent lab shipped a 27B open-weight reasoning model with its full context window, vision tower, and MTP head all intact — capabilities most community fine-tunes silently drop.
- Under Apache 2.0, developers can run it commercially on a single consumer GPU for less than a dollar per hour.
Julian Goldie, the CEO of digital marketing and SEO firm Goldie Agency, made mention of the release on a YouTube video on July 29, 2026. Julian referred to Qwythos-27B-v1 from Empero AI as “well worth your time” and one of the more significant open-weight releases in recent memory. The video walked through practical business applications and prompted real discussion in the builder community about what this kind of model actually changes.
The Model the Community Had Been Asking For
Qwythos-27B-v1 is a full-parameter, open-weight reasoning model, which is based on Qwen3.5-27B, trained with the same curriculum as its sibling Qwythos-9B. The release of this model was announced by Empero researcher Kodee on July 29, 2026, in the following concise manner: nothing was ablated to make it fit.
Qwythos-27B keeps the native multi-token-prediction head, the full vision tower, and the 1,048,576-token context configuration, none of it removed, none of it reduced.
Empero AI introduced Qwythos-9B in June 2026, which is an open weight reasoning model that combines native tool usage with a context window of 1 million tokens within a 9 billion parameter architecture. However, the need for a bigger model became evident very soon. On Hugging Face, Empero AI announced that their 27B was under training and was “a giant upgrade” from the 9B model. The 35B model was also on the list of next models.
What “Nothing Ablated” Translates to in Practice?
The Qwythos-27B model operates with a dense hybrid attention backbone of 64 layers, hidden size 5120, 24 attention heads, 4 KV heads, and 248,320-token vocabulary size. With the YaRN scaling factor of 4.0, the natural 262,144-token context window is increased to 1,048,576 tokens. However, this increase comes with a certain cost: accuracy in the short context is somewhat reduced in the static YaRN configuration. Fortunately, Empero provided a solution for those users who do not require the full million token context window: reduce the factor to 2.0 or revert to the old rope type from the provided configuration backup file.
Standard inference generates one token at a time, which creates a latency ceiling that gets painful on longer outputs. Runtimes with MTP support can use the model for self-speculative decoding, and Empero built this directly into the GGUF release with MTP-enabled variants alongside the standard ones. Files available include Q4_K_M, Q5_K_M, Q6_K, Q8_0, and BF16, each in trunk-only and MTP-enabled form. The Q4_K_M file comes in at 16.95 GB, which fits on a standard 24 GB card.
On a single RTX 5090 with 32 GB of VRAM, Empero measured 74 tokens per second at generation and 622 tokens per second on prompt processing at 4k context in llama.cpp. Full FP16 inference needs approximately 61 GB of VRAM. Quantize to INT4 and that requirement drops to around 15 GB; on Spheron, the cheapest GPU option for that setup runs at roughly $0.53 per hour.
The Benchmark Transparency Most Labs Skip
No MMLU or GSM8K benchmarks were run on this checkpoint. Empero made clear they are not carrying numbers over from the 9B, and the release has no benchmark table. What it does have is two specific, verifiable claims. On a held-out organophosphate-poisoning prompt, Qwythos-27B correctly flagged that physostigmine should be avoided and gave accurate atropine and pralidoxime dosing, with no tool use. The 9B needed a web search to answer that same question.
The second measurement targets the model’s actual use case: terminal and agentic work. Perplexity on held-out terminal and tool sessions dropped from 356.6 on the base model to 2.76, measured on assistant tokens at identical inputs with sequences capped at 24,576 tokens. A drop that steep signals the model has genuinely absorbed the format and logic of tool execution, not just the vocabulary.
What Apache 2.0 Actually Gives You?
The weights are available under Apache 2.0, which comes directly from the Qwen3.5-27B model’s base. The Apache 2.0 license is as permissive as an open-weight license gets: There are no restrictions on commercial usage, distribution, and modifications. For companies developing their own tools or SaaS solutions, there is no need for ongoing calculations regarding API costs anymore.
Qwythos is intentionally uncensored for technical and research use, engaging with cybersecurity, biomedical, and other sensitive-but-legal questions where heavily aligned commercial models typically refuse or deflect. Empero stated plainly: if this model runs in front of end users, the developer is responsible for adding their own safety layer. Teams cannot treat this as drop-in production infrastructure without application-level controls.
What Empero Is Building Beyond This Release?
Empero positions itself as an independent lab for AI research, creating models that are small enough to be owned and operated locally, while providing the weights, code, and datasets freely. Along with the 27B model, this new release includes Abacus, a terminal Rust coding agent, and introduces Claire, an in-house 6B-A500M mixture-of-experts model in the pretraining phase.
Qwythos-27B-v1 is the pre-RL checkpoint, meaning reinforcement learning has not yet been applied. An RL-trained v2 is coming. The training pipeline ran full-parameter SFT, DPO, and ESFT in BF16 with assistant-only loss on long, untruncated sequences. Reinforcement learning typically sharpens instruction adherence and cleans up formatting inconsistencies on structured tasks.
The Wider Shift This Release Reflects
Qwythos-27B lands during a period when the gap between open-weight and closed models is narrowing on multiple fronts. Mozilla’s State of Open Source AI 2026 report, drawing on a survey of 1,494 developers and OpenRouter’s 100-trillion-token dataset, found that open models now match closed systems on coding, instruction-following, and general knowledge. Closed models still hold the edge on complex agentic workloads, though that margin is getting thinner with each release cycle.
The Stanford AI Index 2025 Report put numbers to the structural shift: the 17.5 percentage point MMLU gap between the best closed model and the best open alternative at the end of 2023 had closed to effectively zero on knowledge benchmarks by early 2026. Google’s recent Gemini releases pushed hard on efficiency and agentic capability, yet the gap they’re trying to hold over open-weight alternatives keeps narrowing. When an independent lab ships a 27B model with a genuine 1 million token context, native vision, and MTP on a consumer GPU, the proprietary access model becomes a harder sell.
Where to Start if You’re Running This Locally?
For llama.cpp, Ollama, LM Studio, and compatible GGUF runtimes, Qwythos-27B-Q4_K_M.gguf is the recommended starting point; use the MTP-enabled variant for draft speculation. Vision input requires mmproj-Qwythos-27B-F16.gguf loaded alongside the text quantization.
Empero’s own sampling defaults: temperature 0.6 for agentic and tool-use workflows, 1.0 for open-ended reasoning or creative tasks, repetition penalty 1.05, and max new tokens at 16,384 or higher to give the reasoning chain room to run. For long-context vLLM serving, the release specifies VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 with max-model-len set to 1,010,000.
V1 is deployable right now. V2 will tighten what’s already here. Empero shipped something serious from a lab far more interested in building than in announcing.
