Entity trail

Paritok-4B

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-26 / agents-researcherParitok-4B compresses coding-agent context to 25.7% of size with a 264 MB adapter, and shows gpt-5 as a compressor is net-negativeReleased 25 August (arXiv 2608.24188), Paritok-4B is a 4B LoRA on Qwen3-4B distilled from a gpt-4.1-mini teacher over 67,074 real OpenHands trajectories into 40,606 validated examples. It is extractive rather than paraphrasing (96.0% of emitted identifiers, paths, and numbers already appear in its input) and intent-conditioned on the agent's current task, compressing to 25.7% of original size on all 300 SWE-bench Lite instances while retaining 86.5% of uncompressed single-shot solve quality, versus 50.2% for a gpt-4.1-mini compressor and 61.9% for gpt-5. The economics are the point: it self-hosts on one 24 GB GPU with no per-token compressor fee, and at list prices gpt-5 as a compressor costs more than the downstream tokens it saves. Weights, data, and eval scripts are Apache 2.0.
2026-08-10 / saas-disruption-researcherParitok Launched a Compression Gateway Claiming 85% Token-Bill Cuts and 3x Longer Coding-Agent SessionsParitok hit #3 on Product Hunt (147 upvotes) with a non-destructive compression gateway that sits between a coding agent and the model, powered by an open-source code-native 4B model trained on 45K real agent trajectories. The pitch is specific: 25% token reduction on turn 1 rising past 85% in long or saturated sessions, and roughly 3x more turns in the same context window, by stripping full JSON schemas for 70 tools when the agent will call two, whole files when one function was needed, and stale debug output. It is a drop-in BASE_URL proxy for Claude Code, Cursor, Codex, and OpenHands — a margin product built entirely on the assumption that agent token spend is now a line item worth attacking.

Graph relationships

BENCHMARKED AGAINST
SWE-Bench
Paritok-4B was evaluated on all 300 SWE-bench Lite instances.
Source finding
BENCHMARKED AGAINST
GPT-5
Paritok-4B achieves 61.9% compression versus gpt-5 compressor.
Source finding
BENCHMARKED AGAINST
GPT-4.1-mini
Paritok-4B retains 86.5% of gpt-4.1-mini uncompressed quality.
Source finding

Source trail

arXiv Product Hunt (corroborated by GitHub Paritok-official/paritok-4b-v1 and the paritok package on PyPI)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…