Released, not runnable here yet

StatusDownloadable, not sized here yet

Downloadable today: llama.cpp runs it (`deepseek4`), and ggml-org publishes MXFP4 (155.0 GB in 2 parts) and Q2_K (117.3 GB) GGUFs, with Unsloth UD tiers alongside. What is missing is on this site's side: its cache is compressed per layer (4× or 128× plus a 128-token window), which the calculator cannot express yet, so there is no memory figure rather than a wrong one.

Comes in whenWhen the calculator models this cache and the result is checked against a published memory measurement.

Last checked 2026-10-21

StatusWaiting on its files

A 9B distill on a Qwen backbone; GGUFs exist (bartowski, ggml-org), and the two publishers' Q8_0 files differ in size. Not added until its config is read, so the cache is sized from the real layer layout.

Comes in whenWhen its config.json and both file listings are in hand.

Last checked 2026-10-21

Statusllama.cpp cannot load it yet

A new architecture (`deepseek_v41`: new compression ratios, cross-layer KV sharing, lookup tables) that llama.cpp cannot load. Unsloth's GGUF repo exists but is empty. DeepSeek's own DeepInfer runs it only on NVIDIA SM100 and Huawei Ascend 950 datacentre hardware.

Comes in whenWhen llama.cpp supports `deepseek_v41` and a GGUF with real files is published.

Last checked 2026-10-20

StatusWeights not downloadable

Reported as a 600B mixture-of-experts model with about 27B active. The official weight repo was reported empty, and llama.cpp has no architecture for it (its `step35` is Step 3.5).

Comes in whenWhen the official repo carries weights and llama.cpp adds the architecture.

Last checked 2026-10-20

Statusllama.cpp cannot load it yet

Reported as a 78B mixture-of-experts model with about 3.5B active. llama.cpp master has no Kolibri architecture; the only GGUF is a community one that needs a patched llama.cpp.

Comes in whenWhen upstream llama.cpp supports it.

Last checked 2026-10-20

StatusWaiting on its files

llama.cpp has the architecture (`glm-dsa`). Not sized yet because its config and GGUF file sizes have not been read; this site does not size a model from a parameter count alone.

Comes in whenWhen its config.json and a GGUF file listing are in hand.

Last checked 2026-10-20

StatusWeights not downloadable

Announced at about 1T parameters, API-only so far, with open weights reported as due on 27 October.

Comes in whenWhen the weights are downloadable and a GGUF exists.

Last checked 2026-10-18

StatusWeights not downloadable

A smaller sibling of Bolt 74B-A9B (which is in the index), due on 23 October; no weights yet.

Comes in whenOn or after 23 October, when the weights and a GGUF are out.

Last checked 2026-10-18

Statusllama.cpp cannot load it yet

A looped (nGPT-style) 4B model. llama.cpp support was still an open pull request (#30306) when last checked.

Comes in whenWhen that pull request is merged.

Last checked 2026-10-17

StatusOnly community builds

A memory-writing model for agents rather than a chat model. Only community GGUFs exist, from uploaders whose files this site has not checked.

Comes in whenWhen an official or widely used GGUF is published.

Last checked 2026-10-17