papacr0w's picture
Bring 32B card to the 14B shipped register; real weight checksums
d665fe2 verified
|
Raw
History Blame Contribute Delete
3.11 kB

Spec Sheet

The model

Name Flywheel-Local-Coder-32B
Architecture qwen2 (transformer, decoder-only)
Parameters 32.5B
Context length 32,768 tokens
Capabilities chat, code completion, tool calling
Base model Qwen/Qwen2.5-Coder-32B-Instruct
License Apache-2.0 with Qwen attribution

The files

Model telos-coder-32b-cpt2019-q4_k_m.gguf, GGUF single file
Quantization Q4_K_M
Size 18.5 GB (19,851,336,480 bytes)
SHA-256 65e6133fbe4d12579a776047a71bebb98ab86f9e3d343ed821b51dac0ce312f4
Adapter telos-coder-32b-cpt2019-lora.gguf, LoRA GGUF, 268 MB
Adapter SHA-256 08e7d21cfde1af768c877ecc18ee6343c87711c0c38b8c5b16feb9890f94cbac
Verify sha256sum telos-coder-32b-cpt2019-q4_k_m.gguf and compare with checksums.sha256

Hardware guidance

  • GPU: the Q4_K_M weights are ~18.5 GB. A 24 GB card holds the weights but leaves little headroom for a long context; expect to offload some layers to CPU as context grows. Cards below 24 GB run with partial CPU offload and lower speed.
  • CPU only: works, but this is a 32B model. Budget roughly 20 GB of free RAM for weights plus context. Generation is noticeably slower than the 14B; usable for chat and completion when you are not latency-sensitive.
  • Disk: 18.5 GB for the model file, plus 268 MB if you also pull the adapter.

Runs anywhere llama.cpp or Ollama runs: Windows, Linux, macOS. If you want the same behavior at lower memory and higher speed, the 14B sibling is the lighter option.

How it was trained

Continued pretraining (QLoRA) of the base model on a 66.2-million-token corpus of 17,997 files from a real, working development ecosystem: production code, tests, documentation, and research notes. This is the same packed corpus used for the 14B; Qwen2.5-Coder 14B and 32B share a tokenizer, so one corpus trains both. Training ran to adapter checkpoint 2019 (2019 steps, a quarter epoch); the adapter was merged into the base weights and the merge was quantized to Q4_K_M.

Every layer of that build is hashed and recorded in provenance.json: the corpus content, the packed training shards, the adapter checkpoint, the LoRA GGUF, and the final quantized GGUF. Given the same inputs, an outside observer can re-derive and verify each link.

The training corpus is proprietary and is not distributed with the model.

What this model is for

A local-first coding companion at 32B scale: code completion, functions, refactors, test writing, and tool-calling workflows where your code must not leave your machine. It pairs naturally with a verification loop (propose, test, accept only what passes).

Known boundaries

  • No claimed capability uplift over the base model.
  • No benchmark scores yet; only a deterministic generation smoke is recorded. See BENCHMARKS.md.
  • This is a quarter-epoch continued-pretraining adaptation, a light domain pass, not a full retrain.
  • Knowledge cutoff and multilingual behavior follow the base model.