Falcon256/EDOS-Engineering-CL-Hades-20261005

CL-Hades 27B (2026-10-05 release; 2026-10-01 checkpoint) — GGUF Q5KM

This is a frozen snapshot of CL, EDOS Engineering's experimental continual-learning chatbot, deployed under the persona "Hades." CL learns while it runs by editing the model's original quantized weight tensors in place. Its training is self-directed: the model selects its own training corpus, decides what is important, and is then trained on what it selected. It reads new AI research papers in full, judges which findings matter, writes the training text itself as attributed reviews, and marks which of them to learn and how strongly. It also learns from conversation. It uses no LoRA, adapter or side network. This export is the live model after 4,798 online learning updates, merged into one standard GGUF file you can run in llama.cpp.

This release adds 3,707 updates since the 2026-09-17 release. The release date identifies the package; the weights come from the latest complete checkpoint, saved on October 1. All 64 checkpoint tensor checksums were verified before merging. Every byte outside those tensors, including GGUF metadata and the chat template, was compared with the base file.

Provenance

1. 1.Qwen/Qwen3.8-27B, the original model, released by Qwen under Apache-2.0.
2. 2.Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16, in which Blackfrost modified the refusal behavior at the weight level with no fine-tuning or pruning.
3. 3.Blackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF, the Q5KM quantization. We used this file unchanged as our starting point (SHA-256 37e02edc312a35e73eca50ae26d72458f341e17427eeb1724d9927723256c49c).
4. 4.This repository, where EDOS Engineering's CL harness trained part of the MLP weights online, directly in their quantized form.

All other tensors are byte-identical to the Blackfrost Q5KM file. The GGUF metadata and chat template are also unchanged.

We are not affiliated with Qwen, Alibaba or Blackfrost.

Refusal behavior is modified. Blackfrost's note on the base model says: "Refusal behavior in this checkpoint has been deliberately modified at the weight level. It is not a safety-stock model and must not be represented as one." That still applies here. Online learning does not restore safety behavior. You are responsible for deployment controls, access management and isolating any tool credentials.

Usage

This file needs a llama.cpp build that supports the qwen35 architecture. The harness was developed against llama.cpp commit 427291b5b34cd914a31b3fd3b61a68f6184f4b9f.

llama-cli -m EDOS-Engineering-CL-Hades-27B-20261005-Q5_K_M.gguf -ngl 99 -c 8192 -fa on -cnv
llama-server -m EDOS-Engineering-CL-Hades-27B-20261005-Q5_K_M.gguf -ngl 99 -c 8192 -fa on

For reference, the base Q5KM file needed about 19.6 GiB of total device memory on an RTX 5090 with a 4096-token context. It measured about 2,390 prompt tokens/s and 63–66 decode tokens/s. This snapshot uses the same tensor types and layout, so its memory use and speed should be the same.

The weights do not learn when you run this file elsewhere. Online learning requires the CL harness. This file is a fixed checkpoint.

Training method

CL is a direct continual-learning system. While it runs, it learns by updating the model's original weights in place, on the same quantized weights it uses for inference. There is no adapter, no separate fine-tuning run and no full-precision copy of the model. Each change is written directly into the quantized weights, and the next response uses the updated weights.

Self-directed learning

The model selects its own training corpus, decides what is important, and is trained on that. Most of this snapshot's training comes from a loop in which the model:

1. 1.Reads new AI research papers in full, as they are published.
2. 2.Judges each one: what is new, whether the evidence holds up, and how much it matters.
3. 3.Writes its own attributed takeaway. That review is the training text in the automatic review loop. The newer harness also lets Hades request full-source training on a local paper and available Python captures from its linked repositories.
4. 4.Decides which takeaways are worth remembering and how strongly to weight them.
5. 5.Learns what it chose, which changes its own weights.

The harness supplies the pool of papers and runs the optimizer. The model makes the decisions about content: what matters, what the lesson is, and what becomes permanent. The same applies in conversation, where the model can research its paper library with tools and mark its own replies as worth remembering.

The review loop is a form of self-distillation guided by the model's own judgment of importance. Full-source training additionally learns directly from the selected paper text and captured Python files. Anything the model misread or overstated in a review can be learned as well.

Update mechanics (overview)

• —Where: a subset of each layer's MLP weights, cycled through the depth of the network.
• —How: gradient updates computed against the model's actual quantized outputs, written back into the existing quantization format with an unbiased rounding scheme. Small updates therefore accumulate instead of being rounded away.
• —Retention: a replay mechanism with interference protection limits how much new learning overwrites earlier learning.
• —Safety checks: each update is validated before it is committed and rolled back if it fails.

Training data for this snapshot

The checkpoint records 4,798 cumulative online updates, compared with 1,091 in the previous release. Training ran from September 11 through the latest complete save on October 1, 2026.

• —Self-review of research papers: the model's own attributed reviews and takeaways selected for learning.
• —Conversation: user messages and the model's replies, with replay and memory weighting.
• —Full-source training: selected full paper text and available related Python captures. The checkpoint records progress through full-source training job 2, cursor 407; the local training ledger records two completed passes, of 175 and 407 updates.

A complete breakdown of all 4,798 updates by source has not been reconstructed. The previous release's counts and source proportions should not be extrapolated to this snapshot.

This snapshot is a record of one deployment's learning. It is not a general-purpose fine-tune.

Evaluation and limitations

What has been measured, in controlled runs of the same harness:

• —Fact acquisition: the harness taught the model new synthetic facts, reducing target loss by roughly four orders of magnitude. The learned weights reloaded bit-for-bit in a fresh process.
• —Short-stream retention: with replay protection, learned facts survived held-out paraphrases and further training on unrelated text, and general control questions stayed correct. Runs without protection forgot earlier facts.

What has not been established:

• —Long-stream retention. It has not been validated at the scale of this deployment. How much of this deployment's conversation the snapshot actually retains has not been evaluated.
• —Effects on general capability. We have not run standard benchmarks on this snapshot against the base model. It may be better or worse than Blackfrost's Q5KM on any given task.
• —Accuracy of learned paper knowledge. The reviews were never fact-checked before training. Any misreading or overstatement in a review was learned along with it. Recall of specific papers has not been evaluated.
• —Memorization of training text. The model may reproduce fragments of its training conversation and of its own paper reviews, full paper text, or captured source code.
• —Behavior on its own. The model was trained under a system prompt that describes the live-learning harness. Without that harness, the model may still claim it can learn from the current conversation. It cannot.

Files

• —EDOS-Engineering-CL-Hades-27B-20261005-Q5_K_M.gguf: the model
• —EDOS-Engineering-CL-Hades-27B-20261005-Q5_K_M.gguf.sha256: checksum

License

Apache-2.0, the same as the upstream Qwen3.8-27B and the Blackfrost abliterated derivatives. Please credit Qwen and Blackfrost along with this work.