Hello, I'm Tokenbender
Hello!
I am Abhishek Harshvardhan Mishra, aka Tokenbender.
I am the founder of Deus Experiments, a lab specialising in post-training for enterprise use cases and supersparse inference. Before this, I was a technical lead at Intel, working on server firmware. Then I fell in love with language models, threw away everything I had done before, and became part of open-source LLM research.
I am good at making models do whatever is needed, as long as the use case is clearly specified.
The route so far
-
Before 2023
Intel Technical lead working on server firmware.
-
CodeCherryPop
The first friendly local coder of the Llama 2 era.
Jul 2023 · Independent receipts #1 Llama 2 coder.
Beat models 10× its size.
-
EvolvedSeeker
68.29% HumanEval from a 1.3B coding model.
Nov 2023 · Release-window receipts #1 coder in the 1B class.
Beat models 12× its size.
“1.3B with 68.29% HumanEval lol, don't behead me.” -
Navarna
Hindi chat and retrieval-augmented generation.
Feb 2024 · Open pipeline First Hindi chat model with retrieval built in.
One model for conversation and retrieval.
“Code/Data/process OSS.” -
avataRL
Pretraining language models from random weights with pure reinforcement learning.
May–Aug 2025 · Crown jewel Cross-entropy pretraining is optional.
I trained language models from random weights with pure reinforcement learning.
“Random weights. No cross-entropy pretraining.”- 01 The result Coherent language from random-weight models across the 30M–250M range.
- 02 University of Chicago Cited in a PhD dissertation on self-play reinforcement learning for language models.
- 03 The field noticed A 39-author Tsinghua and Shanghai AI Lab survey named it a concrete path to RL-from-scratch training.
- 04 The work escaped Rebuilt for Lightning, SAC, MoE and Qwen3 by independent researchers.
-
infinite
Rubric-driven prioritized replay for continual learning.
Aug 2025 · Continual learning The model’s failures should choose what it learns next.
Infinite turns continual learning into spaced repetition: train what is slipping, refresh what is stable, and keep the model in the game.
- Low-rank circuit conditioning 91.33% recovery from 5.05% of a model's MLP channels.