Osaurus released Diplodocus preview 4 as llama.cpp GGUF files, a 9B finance and spreadsheet assistant built on K2-Horizon-7B with tool use.
- Osaurus released Diplodocus preview 4 GGUF , a 9B finance and documents assistant for llama.cpp.
- Built on IFM's K2-Horizon-7B, Apache 2.0, with native tool use and toggleable reasoning.
- Three quant sizes: Q4_K_M at 6.33 GB, Q6_K at 7.76 GB (recommended), Q8_0 at 9.57 GB.
- Requires llama.cpp with K2 Horizon support merged on 2026-10-06; older builds will not load it.
- Workbook and document task scores moved from 59/51 to 70/70 versus preview 3.
- Known limits: live formulas sometimes come out as text, signed PDF edits are unreliable.
Osaurus has released Diplodocus preview 4 as three GGUF builds for llama.cpp. The 7B model targets local workflows involving PDFs, spreadsheets, document edits, and calculator-backed financial arithmetic.
The model fine-tunes IFM’s K2-Horizon-7B on Osaurus’s workspace tools, including in-place workbook and document editing, a calculator, and optional reasoning. File operations still require a host application that implements those tools and executes the model’s calls. The release uses the Apache 2.0 license and supports llama.cpp-based runtimes such as Ollama, LM Studio, Jan, and Docker Model Runner once they include the required architecture support.
Three quantizations, one hard dependency #
All three builds use an importance matrix, which applies calibration data during compression to preserve influential weights.
| Build | File size | Intended use |
|---|---|---|
Q4_K_M |
6.33 GB | 8 GB GPUs and memory-constrained systems |
Q6_K |
7.76 GB | Osaurus’s recommended balance of size and fidelity |
Q8_0 |
9.57 GB | Behavior closest to the original BF16 weights |
Actual memory use exceeds the file size because the runtime also allocates memory for the context window, KV cache, and other buffers. Longer contexts can therefore require partial CPU off even when the model file fits in VRAM.
any build requires llama.cpp with K2 Horizon support from PR #29535, merged on October 6, 2026. Older llama.cpp builds will reject the k2-horizon architecture, including applications that bundle an outdated engine. Mac users running the Osaurus application can instead use the separate JANG_6M MLX package containing the same weights.
This story is for Pro members #
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.