HobbyLM-1B GGUF

GGUF files of HobbyLM-1B Instruct and the HobbyLM-1B base model. A patched runtime is required.
HobbyLM by Fuel Labs

Windows App + Patched Runtimes | Instruct Model | MLX (Apple) | Base Model | Online Demo | GitHub

Easiest way to try it on Windows: the double-click chat bundle in the HobbyLM 1.0.0 release. Download HobbyLM-1B-Chat-Windows-x64-1.0.0.zip, extract it and double-click Start HobbyLM.bat; the chat opens in your browser and runs offline. It already contains the Q4_K_M file below and the patched runtime.

[!IMPORTANT] These files need a patched runtime.

• Tested and fails: stock Ollama v0.35.1 (attn_qkv.weight shape error).
• Not run: unpatched llama.cpp. Its code expects the same tensor shape that Ollama rejects, so it needs our one-line patch.
• Not validated: LM Studio, Docker Model Runner and other GGUF tools.

Use the patched Windows runtimes from the release page, or build llama.cpp 63bef272 with bailingmoe2_decoupled_head_dim.patch.

Why do the files say bailingmoe2? bailingmoe2 is the name of an architecture already built into llama.cpp, added for inclusionAI's Ling 2.0 models (BailingMoeV2). HobbyLM is not based on, derived from or affiliated with those models: it was designed and trained from scratch (architecture by Harish; trained by Fuel Labs). The GGUF files reuse this llama.cpp code path only because its layout matches HobbyLM's: a sigmoid-routed mixture of experts with expert-bias selection, a shared expert and a leading dense layer, plus grouped-query attention with QK-norm. The one difference, HobbyLM's 2,048-wide attention against a 1,024 hidden size, is what bailingmoe2_decoupled_head_dim.patch fixes.

Files

• Same files as released: byte-identical to the files first published on 2026-10-07 in the base repository's gguf/ folder (harims95/hobbylm-1b-hf, now hobbylm-1B, revision dc11aab4e06fd2000313c8821557a86e8368bf58), where older pinned download commands still find them. Checksums: SHA256SUMS.txt.
• Conversion: from the released model.safetensors with to_gguf_1b.py, which reproduces the Instruct F32 file bit for bit. Q8_0 and Q4_K_M were made with llama-quantize and keep the router and expert-bias tensors in F32.
• Chat template: embedded in the instruct files; run llama.cpp with --jinja.

Download one file (pinned revision):

huggingface-cli download harims95/hobbylm-1B-gguf hobbylm-1b-sft-3450-Q4_K_M.gguf --revision f8bc034403026b2706f5275702d3801753b87c34 --local-dir .
sha256sum hobbylm-1b-sft-3450-Q4_K_M.gguf   # compare with SHA256SUMS.txt

Replace the file name for Q8_0, F32 or the base F32 file.

Checks against the original model (fp32 Transformers outputs)

• Non-identical outputs: the one non-identical case in each 6/7 row is the same long rewrite prompt, which stays on task.
• *Perplexity was measured on the model's own evaluation prompts and outputs, so it shows relative quantisation loss only.
• No benchmarks were run on the GGUF files. The release evaluation ran on the fp32 Transformers models; see the instruct and base model cards. These checks do not establish general quality parity.

Tested platform

Windows 11 x86-64, CPU backend, on one machine (AMD Ryzen 9 5900HX, AVX2 "haswell" code path). Not tested: other x86-64 CPU variants included in the runtimes, Windows on ARM, Linux, macOS and any GPU backend.

Run with the patched llama.cpp

Unzip hobbylm-llama.cpp-63bef27-patched-windows-x64-cpu-1.0.0.zip from the release page. In that folder, run:

start-chat-server.bat "C:\path\to\hobbylm-1b-sft-3450-Q4_K_M.gguf"

Then open the address it prints (http://127.0.0.1:8080/, this computer only). For the base model:

run-base-completion.bat "C:\path\to\hobbylm-1b-base-F32.gguf" "text to continue"

Run with the patched Ollama

Unzip hobbylm-ollama-0.35.1-patched-windows-x64-cpu-1.0.0.zip from the release page and follow its README. It runs separately from any installed Ollama (port 127.0.0.1:11435, its own data folder). The included Modelfile.hobbylm-1b-sft-q4 sets the chat template, the <|endoftext|> stop token and greedy decoding.

All runtime packages are unofficial builds for Windows x86-64 CPU, not produced, reviewed or endorsed by the llama.cpp/ggml or Ollama projects; the rebuilt binaries are not code-signed. Patches and build steps: runtimes/ on GitHub.

License

Apache License 2.0 (LICENSE): the model weights in these files and the HobbyLM architecture code they derive from (by Harish, released with its author's permission). Not covered: third-party training data (its terms and the open training-data question are in NOTICE and the base model card) and third-party runtime software (its own licences). Tokenizer: GPT-2 BPE (MIT, OpenAI).

Credits

Built by Hariharan and Prabhurajhan at Fuel Labs. Architecture based on HobbyLM by Harish.