Open Weight Models

last updated 2026-08-31

Physics / mechanism

Open weight models are models whose trained parameters are published for download, so that any party can serve, fine-tune, quantise or distil them on their own hardware rather than calling a hosted API. The distinguishing technical property is not the architecture but the point of control: once weights are downloadable, the serving stack and the post-training stack become separable from the lab that trained the model.

Recent frontier-scale open weight releases are extremely sparse mixture-of-experts systems, which shifts the cost structure of self-serving. Kimi K3 was published with 2.8T total parameters and “effectively activating 16 out of 896 experts” per token under a scheme Moonshot calls Stable LatentMoE, with 1M context and native vision ref. Sparsity of that degree means the whole model is never woken for a single token, but it also distributes the router’s 16 selected experts across 16 different chips, which is why Moonshot recommend a 64-plus accelerator supernode to serve it ref. Serving requirements, not licence terms, therefore set who can actually run a frontier open weight model.

The second parameter set is memory per token. K3 stacks DeepSeek-lineage Gated MLA, which compresses the KV cache, with Kimi Delta Attention, a linear attention variant with per-channel forgetting that replaces the KV cache with a fixed-size compressed running state, interleaved with regular attention layers; Attention Residuals selectively retrieve representations across depth rather than accumulating them uniformly ref. Moonshot claim 2.5x scaling efficiency from this chain, a vendor number ref. Pricing was set at $3 in / $15 out per million tokens, parity with Claude Sonnet rather than the historic Chinese discount, with the model positioned by its own vendor as trailing Fable 5 and GPT-5.6 Sol overall while winning on game-dev and UI tasks ref.

Downstream, open weights are mostly consumed as substrates for specialisation rather than as drop-in replacements. Fireworks reported that 95% of tokens it serves come from customer-specialised models, that is fine-tuned open weights, adapters and distillations, rather than downloadable third-party models served as-is ref.

Competitive landscape

The competitive question is whether openness commoditises the closed frontier. One reading holds that releasing frontier open weight models commoditises closed models, challenges the closed labs’ businesses and splinters their most valuable asset, their teams, since staff leave to build on open weights when the labs’ primacy looks uncertain; on that reading Chinese labs stay ahead and could switch to closed once the lead is large enough, with usage-driven dependence as the strategic payoff ref. The macro corollary is that AI provision looks like a competitive, mediocre-return industry, partly because of intense and cheap Chinese competition, which would force providers to invest less and hit the “shovel sellers” that led equities for three years.

Against a full commoditisation reading, serving vendors compete on the same fine-tune plus RL plus serve stack, so differentiation may collapse to execution rather than structure, and closed frontier models may stay far enough ahead that “own your intelligence on open weights” remains a minority of workloads ref. Vendor self-positioning is consistent with a residual gap: K3 was framed by Moonshot as trailing the leading closed models overall while priced at parity with a closed mid-tier model ref.

Evidence base

Frontier (open questions)

Synthesised 2026-08-31 from 6 KB sources by the resynth pipeline; citations are KB source slugs.

Recent mentions

Frontier questions