K2 Horizon: A connected fleet of six open models

Original Article ↗ Hacker News Discussion ↗

A new family of “radically open” K2 Horizon language models is drawing attention both for publishing weights and large portions of training data, and for aiming to provide a fully transparent training pipeline. Commenters welcome another serious open-stack entrant—especially a 7B model that benchmarks competitively for coding on modest hardware—but note that larger variants still trail top open models like Qwen and that some repos and recipes remain incomplete. The release also revives broader debates over whether fully open models can or should include training data, how copyright and fair use apply, and whether the rapid flood of new LLMs is leading to “model fatigue.”

Model Fatigue, Hype, and Analogy to Past Tech Cycles

  • Several commenters express “model fatigue” from constant new releases.
  • Comparisons are made to earlier eras of CPU and smartphone hype: early launches were huge events, now only specialists track every new chip/phone.
  • Expectation that LLMs will follow a similar path: most users will just pick “good enough” models.
  • Some note that “you don’t have to pay attention”; what you build with models matters more than chasing frontier releases.

Openness, Training Data, and “Radically Open” Claims

  • Strong support for fully open stacks: weights, code, training data, and recipes, to reduce black-box manipulation.
  • Skepticism toward “radically open” marketing until training data and full recipes actually ship.
  • It’s noted that K2’s training datasets are already on Hugging Face, but pre/post-training repos are still placeholders.
  • Prior work from the same ecosystem (e.g., earlier K2 models) makes some expect full recipes eventually, but they reserve judgment.

Performance and Benchmarks

  • Dense 32B is reported (by their own charts) to trail stronger open competitors like Qwen 3.8 27B; Gemma 4 31B isn’t in the comparison set.
  • Some argue the 32B is labeled “stage 1” and unfinished, so early release is mainly about showcasing the open pipeline.
  • The 7B model is repeatedly called out as impressive, potentially “best under 10B” and competitive on coding benchmarks (e.g., SWE-bench-verified scores comparable to much larger models).
  • Others highlight that comparisons omit newer strong open models and that quantization classes/variants aren’t consistently matched.

Coding Capabilities and Small Models

  • A few users test smaller models (3.7B, 7B) on interview-style coding problems.
  • Reports: 3.7B fails basic tasks, hallucinates APIs, and gets stuck; 7B gives mixed or partially correct answers.
  • Debate over whether sub-10B models are suitable for serious coding vs just autocomplete/summarization.
  • Some compare to other small coding models that already handle such tasks well.

Tooling, Deployment, and Practicalities

  • Interest in running K2 locally; users mention vLLM support and recent dflash2 support for other models.
  • Frustration that new releases often support vLLM quickly but lag on llama.cpp, which is key for older/low-end hardware.
  • Observations that the hosted K2 demo is extremely fast when it works, though initially some models were unavailable.

Broader IP, Copyright, and Synthetic Data Debates

  • Long subthread on whether fully open models are even possible under current copyright law.
  • Ideas discussed:
    • Using LLMs to generate synthetic training data from non-permissive sources.
    • Fully synthetic or public-domain-only corpora.
    • Decentralized training and storage.
  • Controversial proposals to treat all publicly reachable internet data as public domain are pushed and strongly criticized as impractical and dystopian.
  • Some point to existing projects (e.g., OLMo/Dolma and similar) as partial proof-of-concept for transparent pipelines, even if underlying scraped content isn’t fully redistributable.

Miscellaneous Notes

  • Some complain about missing/locked blog charts and small, hard-to-read visuals.
  • Confusion over naming (“K2” already used elsewhere).
  • Appreciation that this launch coincides with major outages of closed LLMs, reinforcing the perceived value of open models.