Open Source This Month: a Cheaper DeepSeek, an Open Code Corpus, and Audits That Check Themselves

September’s open-source releases were about cost and plumbing. DeepSeek’s V4.1-Flash is a 552-billion-parameter multimodal model under MIT that fits a million-token context into 890 bytes of cache per token, and OpenBMB released about 550 billion tokens of open code data.
artificial-intelligence
data-science
Author

Kabui, Charles

Published

2026-09-22

Keywords

open-weights, kv-cache-compression, code-pretraining-data, research-agents, security-audit

Download as PDF

Open-source releases this month were about cost and infrastructure, not capability jumps. DeepSeek’s V4.1-Flash is a 552-billion-parameter multimodal model that activates 16 billion per token and fits a one-million-token context into a key-value cache of 890 bytes per token, a quarter of its predecessor’s. Artificial Analysis rates it 39 on its Intelligence Index at 226 tokens per second and $0.27 per task. OpenBMB released UltraData, which archives 192 million public GitHub repositories and ships about 400 billion tokens of selected code plus 150 billion tokens of generated exercises. OpenResearch turns Claude Code, Codex, Cursor and OpenCode into research agents with git-tracked experiments, and Cloudflare’s security-audit-skill runs six-phase code audits whose findings are independently checked.

The effect is a lower floor for building on open models. A team can fine-tune on a curated code mixture instead of scraping GitHub, or run a fast multimodal model locally. DeepSeek trails the leader on Artificial Analysis’ index by ten points but costs a fraction per task, because its design saves memory rather than chasing accuracy. OpenBMB’s own numbers show selection helps by itself: continued training on the selected code tier beat the unselected tier by 7.8% on EvalPlus.

Openness still needs qualifiers. UltraData’s licence is Apache-2.0, but its content comes from public repositories, and the Cloudflare skill earns trust by separating confirmed findings from ones needing validation. A survey of AI in games (arXiv 2609.16679) makes a related point: results rarely carry between engines, so evidence has to be rebuilt for each setting.

Read More: the previous open-source radar, on forecasting models and torrent-distributed weights

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Open {Source} {This} {Month:} A {Cheaper} {DeepSeek,} an
    {Open} {Code} {Corpus,} and {Audits} {That} {Check} {Themselves}},
  date = {2026-09-22},
  url = {https://toknow.ai/posts/open-source-radar-deepseek-v41-flash-ultradata/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Open Source This Month: A Cheaper DeepSeek, an Open Code Corpus, and Audits That Check Themselves.” https://toknow.ai/posts/open-source-radar-deepseek-v41-flash-ultradata/.