--- license: apache-2.0 base_model: Qwen/Qwen3.6-35B-A3B tags: [bob, gguf, mixture-of-experts, tool-use] --- # 28Bob Riker A Bob cut from Qwen3.6-35B-A3B for Bob's own jobs: tools, coding, exact formats and everyday chat. > **Warning: unintentionally poisoned, kept for study only.** This model was trained on Bob's old lesson sets (v1 to v3), which > we later found were unintentionally poisoned. When Claude Opus 5.5 read every tool lesson in full on 15 Nov 2026, it would keep > only 3% of the main tool set as written (189 of 5,788). 40% of the tool lessons it read had made-up tool results, 28% carried > private details picked up from the prompt, and 23% had a wrong tool step. The lessons also taught habits we have since removed, > such as handing jobs to `ask_specialist` and writing files through `run_command`. How much of this the model carries is > unproven in our lab, as we have yet to test. Bob no longer offers it. The full analysis and the new clean lessons (v4, made > from real jobs) are on [MouseWurX/bob-lessons](https://huggingface.co/datasets/MouseWurX/bob-lessons). Made by [Bob's Forge](https://github.com/MouseWurX/Bob/blob/main/app/docs/FORGE-JOURNAL.md) from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) (all experts kept per layer, about 29 B parameters, 3 B active per word). Method, numbers and logs are in the Forge journal. ## Files | File | Size | Needs | |---|---|---| | `28Bob-Q3_K_M.gguf` | 13.9 GB | about 15 GB of card memory | | `28Bob-Q4_K_M.gguf` | 17.5 GB | about 19 GB of card memory | Each GGUF has a `.bob.json` card beside it that Bob's catalogue reads when you add the file as your own brain. ## Home benchmarks (2026-11-08) | Test | 28Bob | Qwen3.6 35B-A3B | |---|---|---| | Coding | 69 | 92 | | Quick code | 100 | 100 | | Tools | 65 | 67 | | Facts | 90 | 100 | | Routing | 96 | | | ARC-Challenge | 92.9 | 95.8 | | ARC-Easy | 97.6 | 98.9 | | ARC-Challenge (thinking) | 93.9 | | | ARC-Easy (thinking) | 98.3 | | The project's own tests on one machine, for comparing Bobs with each other and with the model they came from, not for leaderboards. ## Part of Bob This is part of **[Bob](https://github.com/MouseWurX/Bob)**, a free, open-source AI agent that runs on your own computer (Windows, Mac, Linux, Raspberry Pi, Android). Nothing goes online unless you turn it on. **What Bob does** - **A real agent:** researches with sources, writes, runs and fixes code, reads and edits documents and files, works other apps, and makes pictures, clips, music and a voice. - **His own brains:** models cut and retrained to run Bob on everyday graphics cards, with every score, lesson set and dollar published. - **Bob-Net:** your Bobs share graphics cards and memory and check each other's work, on your home network. - **Gaming friendly:** the brain unloads when idle, and a Gaming session gives the graphics card back to your game. - **Private:** no account, no telemetry. Everything optional starts switched off. **The Computer Science Lab** ([guide](https://github.com/MouseWurX/Bob/blob/main/app/docs/USER-GUIDE.md#make-your-own-brain-the-lab), [code](https://github.com/MouseWurX/Bob/tree/main/app/model-lab)), where this was made and measured. It's a workbench inside Bob, with no command line: - **Make a brain:** shrink it, cut experts, or take out sight, hearing or the refusal habit. A brain is usable only after it has been tested. - **Forge:** every way to make a new brain, side by side. It writes the plan for jobs that need a rented card. - **Test:** scored challenges, one brain in six sizes, long-context and speed tests, and blind taste tests. - **Teach:** teach a brain a lesson set, or have a bigger brain teach a smaller one. - **Look inside:** see which experts your work uses. **fNRI** scans a brain while it answers and plays it back in 3D. **Training protocols** ([Forge guide](https://github.com/MouseWurX/Bob/blob/main/app/model-lab/forge/README.md), [Forge journal](https://github.com/MouseWurX/Bob/blob/main/app/docs/FORGE-JOURNAL.md), [Making smaller models](https://github.com/MouseWurX/Bob/blob/main/app/docs/MAKING-SMALLER-MODELS.md)). Every model goes through the same steps, in order: baseline → clean first → analyse → train → cut in steps → heal → check every target → learn from its own answers → blend → live dials → publish. Each step is benchmarked. Why we chose this order: - **Baseline first,** so every change is measured against the untouched model, not against a guess. - **Clean before teaching, cut after training,** so a cut keeps what training taught and no fault gets baked in. - **Measure before cutting (REAP),** so the experts that do the least work are the ones removed, per job, from real traces. - **Cut in steps and heal from the full model's odds** (distillation), because small steps show where quality breaks, and copying the teacher's odds recovers more than plain answers can. - **No target may drop,** so a gain in tools can't quietly cost coding, facts or safety. - **Learn from its own answers** (on-policy, reward for tool calls that really work), because a model fixes the mistakes it actually makes, not the ones a teacher imagines. - **Honest data:** test items are never taught, surgery and evil-check data never train, and everything is scrubbed of private data first. **The plan:** what comes next, and why, in the [roadmap](https://github.com/MouseWurX/Bob/pull/230). **Costs:** every rental, its card, its hours and what it bought, on the [costs page](https://github.com/MouseWurX/Bob/blob/main/app/docs/COSTS.md). ## Licence and credit Apache 2.0, as the base model (Alibaba's Qwen team). The lessons it learned are Bob's own lesson sets (CC BY-SA 4.0) plus openly licensed traces from other models, credited in the Forge journal. The Bob names are a tribute to Dennis E. Taylor's novels; the project is not affiliated with the author. ## Changelog - 2026-11-16: Marked as unintentionally poisoned (trained on the old lesson sets v1 to v3); see the warning above and the analysis on [bob-lessons](https://huggingface.co/datasets/MouseWurX/bob-lessons). The files are unchanged.