euangelion.
Artificial intelligence

New downloadable models widen local choice, with very different commercial terms

Flint, MiniMax M3.1 Flash, Liquid’s LFM3 and NVIDIA’s Nemotron Bolt add self-hosting options, but hardware needs and licence conditions make them distinct purchasing decisions.

122 words · 1 min

Recursive’s Flint, released on October 21, is a 40-billion-parameter coding model that activates roughly 2.5 billion parameters at a time.

Its weights are available under Apache 2.0. The smaller active count helps explain its design for local coding agents; it does not mean the full model occupies only 2.5 billion parameters of memory. Recursive’s speed figures depend on specified hardware, quantization and runtime. [1][2]

MiniMax’s M3.1 Flash has also crossed from a hosted offering to downloadable files. Its model repository records an October 23 initial upload and contains the weight shards, with a displayed repository size of about 252 GB. That is a consequential release for teams able to host it, rather than a claim that the full model fits an ordinary laptop. [3][4]

Sources & details 13 sources
Discuss in your voice app

Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.

Open ChatGPT
Paste the brief into your voice app to discuss this reporting.

Reporting sources

  1. Recursive — Introducing Flint

    Architecture and vendor measurements; release date cross-checked with OpenHands.

  2. Recursive — Flint model repository

    Apache 2.0 model card and actual files inspected.

  3. MiniMax — M3.1 Flash model repository

    Model card and downloadable release.

  4. MiniMax — commit history and files

    Initial upload October 23; file tree separately inspected, approximately 252 GB.

  5. MiniMax — M3.1 Flash licence

    Commercial attribution and notice; prior authorization above $20m yearly revenue generated by covered products/services, not total company revenue. Reopened October28.

  6. Liquid — October 27 LFM3 announcement listing

    Dated release listing; full launch page exceeded the retrieval size limit.

  7. Liquid — LFM3-1.7B model card

    Text-model scope, footprint estimate and limitations.

  8. Liquid — LFM Open License

    Legal-entity scope includes affiliates. Summary says exceeds $10m while threshold definition says $10m or more; exact-boundary permission is not assumed. Reopened October28.

  9. NVIDIA — Nemotron Bolt release

    October 23 release and intended device footprint.

  10. NVIDIA — Nemotron Bolt model card

    Quantization, model provenance and OpenMDW 1.1 terms.

  11. OpenMDW — licence 1.1

    Underlying permission and notice obligations inspected.

  12. OpenHands — Flint availability

    October 21 integration report; a launch partner, not an independent performance replication.

  13. MiniMax weight files

    Repository file tree inspected for uploaded shards and approximately 252 GB total; storage is not a runtime-memory estimate.

Images & figures

  • Each licence measures its own conditions. MiniMax uses yearly revenue from covered products or services; Liquid uses the legal entity, including affiliates. Liquid’s exact $10m boundary is inconsistent in the source and needs confirmation. Other terms still apply.

    Euangelion / Codex · Source · Usage terms