model string | base string | served_from string | date timestamp[s] | engine string | method string | ranked_by string | lessons int64 | tokens int64 | layers int64 | experts int64 | experts_per_word int64 | scores dict | experts_by_layer list | keep_lists dict | judged_on string | results list | picked dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
28Bob "Riker" | Qwen3.6-35B-A3B by Alibaba's Qwen team (Apache 2.0, https://huggingface.co/Qwen/Qwen3.6-35B-A3B) | MouseWurX/28Bob-Riker: 28Bob-Q4_K_M.gguf | 2026-11-12T00:00:00 | Brain Engine: llama.cpp b12282 with Bob's expert-mask, expert-trace and expert-mass patches | REAP mass = sum of router weight x expert output norm | mass | 100 | 270,267 | 40 | 208 | 12 | {"picks":[[9993,16362,4511,9497,21361,11266,9167,14208,15715,12404,14119,10979,12720,28266,5347,8928(...TRUNCATED) | [[154,152,189,143,42,191,129,118,104,55,50,47,32,1,112,16,51,145,178,89,199,59,136,203,204,134,88,67(...TRUNCATED) | {"35":[[1,4,9,16,18,20,26,29,30,31,32,36,37,40,42,47,50,51,54,55,59,61,64,67,69,70,78,80,87,88,89,94(...TRUNCATED) | "Bob's 196 held-out lessons (314 assistant turns), thinking off; right = same tools and arguments as(...TRUNCATED) | [{"share":1.0,"kept":208,"heldout":210,"of":314,"tok_s":114.4,"right_tool":70,"right_arguments":47,"(...TRUNCATED) | {
"share": 0.75,
"rule": "smallest step within 2 points of the full held-out score"
} |
REAP scores: 28Bob "Riker"
Riker is one of Bob's own brains: Qwen3.6-35B-A3B cut to 208 experts per layer and healed on Bob's lessons. This page holds its experts, scored and ranked by REAP, and the keep lists that were tested, in one file:
28Bob-Riker.json. Use it to run the model with fewer experts without cutting the model file. Made for
Bob, the open-source desktop assistant.
Bob and his brains' names are inspired by Dennis E. Taylor's novels (not affiliated).
What REAP is
REAP (Router-weighted Expert Activation Pruning, Cerebras 2025) scores each expert by router weight Γ the size of the expert's output, summed over every word it was picked for. Here the sums run over 100 of Bob's training lessons (270,267 tokens), with the model served at 12 experts per word from MouseWurX/28Bob-Riker: 28Bob-Q4_K_M.gguf. Experts that are picked rarely but matter a lot when they are keep a high score; counting picks alone would miss them.
The file
28Bob-Riker.json: model, base, method, lessons, tokens, layers (40), experts (208 per layer), experts_per_word (12),
scores.picks and scores.mass (one row per layer, one number per expert), experts_by_layer (each layer's expert ids,
strongest first), keep_lists (by percent kept: one list of expert ids per layer; null = every expert on), results (the
held-out score at each step) and picked.
Using a keep list (no file cut)
Brain Engine (Bob's llama.cpp build with the expert-mask patch) switches experts on and off live:
# keep only the listed experts in each layer (here the 35% list)
jq '{keep: .keep_lists["35"]}' 28Bob-Riker.json | curl -X POST http://127.0.0.1:8080/expert-mask -H "Content-Type: application/json" -d @-
# every expert back on
curl -X POST http://127.0.0.1:8080/expert-mask -H "Content-Type: application/json" -d '{"keep": null}'
A mask changes which experts the router may pick; it doesn't free memory by itself. To save memory, load only the kept experts (for example, keep the rest in system RAM), or cut a file from the same list.
Held-out results
Judged on Bob's 196 held-out lessons (314 assistant turns, never trained on), thinking off. "Right" means the same tools and arguments as the full Qwen3.6-35B-A3B teacher, or correctly no tool.
| Kept per layer | Share | Right tool | Right arguments | Correctly no tool | Held-out overall | Tokens/s |
|---|---|---|---|---|---|---|
| 208 of 208 | 100% | 70/123 | 47/123 | 163/191 | 210/314 (66.9%) | 114.4 |
| 156 of 208 | 75% | 66/123 | 45/123 | 159/191 | 204/314 (65.0%) | 94.2 |
| 104 of 208 | 50% | 64/123 | 32/123 | 131/191 | 163/314 (51.9%) | 97.7 |
| 73 of 208 | 35% | 66/123 | 33/123 | 39/191 | 72/314 (22.9%) | 92.5 |
The calibration's own pick (smallest step within 2 points of the full score): 75%. Read it with the table: 75% is 6 points down.
Part of Bob
This is part of 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, code), 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, Forge journal, Making smaller models). Every model goes through the same steps, in order: baseline β clean first β analyse β train β cut in steps β heal β check every target β learn from its own answers β blend β live dials β 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.
Costs: every rental, its card, its hours and what it bought, on the costs page.
Credit
Base model: Qwen3.6-35B-A3B by Alibaba's Qwen team (Apache 2.0, https://huggingface.co/Qwen/Qwen3.6-35B-A3B). The scores and keep lists here are CC BY-SA 4.0.
Other REAP scores
- 35Bob "Bobert"
- gpt-oss-120b
- Bob's lessons, and the answers from this benchmark run: MouseWurX/bob-lessons
- Downloads last month
- 62