tags:
- about
David Aylward
☕ Support this work
Whittle is built by one person on a grocery budget and rented GPU hours. If this research is useful to you, or you want to see it finished: ko-fi.com/davida81328. Every hour of GPU time goes straight into the next checkpoint, and every checkpoint, table and log lands in these repos.
AI tinkerer in South Africa. I take language models apart on two 8GB gaming GPUs to see what's inside, then put smaller ones back together. Everything I break and everything I learn ships public: weights, measurements, mistakes and all.
Current
| Whittle-Qwen-3.8-35B-A3B | 35B-total, ~3B-active Qwen3.8-Flash-Next-format MoE with a 10B n-gram memory, distilled from Qwen3.8-27B; root = Phase-2 step 32010, published 30 Sep 2026. GGUFs on Whittle-Qwen-3.8-35B-A3B-GGUF. Parent: Whittle-Next-27B-A3B; base lineage: Qwen3.6-35B-A3B pruned 256→180 experts. |
Earlier work (Aug 2026)
| Qwen3.8-Whittle-16B | A 27B whittled to 16.8B with a logit lens and a pricing table, healed with one A100 evening. 36/39 on the field battery at 20 tok/s on consumer GPUs. Research preview, v2. |
| The un-repaired cut | The raw surgery artifact: full measurement history, every pricing run, every script. The damage profile is the science. |
How this works
No lab, no cluster. Measurements run on a Ryzen 2700X with an RTX 4060 and an RTX 3050. Training runs on rented A100 hours. Every experiment publishes its wins, its bugs, and its dead ends on the model cards themselves.
Support the tinkering
Every donation becomes A100 hours, and every A100 hour ends up as a public model or a public measurement.