`text-to-code`

HF group: Multimodal (no dedicated HF tag; functionally a fine-tune of text-generation) · Status: 🟡 partial

What it is

LLM specialized on code — completion, refactor, explain. Architecturally identical to text-to-text but with code-specific fine-tuning that meaningfully changes quality, particularly on multi-file edits and long contexts.

Open-weight models

Model	Params	Released	License	Quality	Notes
Qwen2.5-Coder-1.5B	1.5 B	2024-11	Apache-2.0	Best small coder	Solid for autocomplete.
Qwen2.5-Coder-7B-Instruct	7 B	2024-11	Apache-2.0	Aider 50 %+	Good chat-with-codebase pick.
Qwen2.5-Coder-32B-Instruct	32 B	2024-11	Apache-2.0	Aider 73.7	Top open coder under 40 B.
Qwen3-Coder-Next	80 B / 3 B active	2026	Apache-2.0	70.6 % SWE-bench Verified	Hybrid attention + MoE; agentic-tuned.
DeepSeek-Coder-V2-16B-A2.4B	16 B / 2.4 B active	2024	DeepSeek	Strong	MoE; 300+ languages.
Yi-Coder-9B-Chat	9 B	2024	Apache-2.0	85 % HumanEval	128k context.
StarCoder2-3B	3 B	2024	OpenRAIL-M	Lightweight	Good for laptop autocomplete.
Devstral Small 2	24 B	2026	Mistral Research	68 % SWE-bench	Built for agentic coding, not chat.

Infrastructure required

Inference

✅ Reuses locara-llama (text-to-text inference path).
❌ Tree-sitter integration for code-aware sampling constraints (nice-to-have, not required).

Input

Plain text (same as text-to-text). Apps that want to feed a codebase do their own chunking.

Output

✅ Same streaming token Channel as text-to-text.

Storage

✅ Weights via locara-models::Cache.
App-side: code-edit history typically held in app SQLite; not framework-level.

Interaction (IPC + SDK)

✅ Reuses llm.chat / llm.chat_stream. No code-specific IPC needed.

Capabilities (manifest)

✅ capabilities.models[] lists a coder model.
❌ No text-to-code modality entry yet in spec/04-modalities.md — apps declare text-to-text and pick a coder model in the picker.

Gaps

Add text-to-code modality entry to the spec so the picker can route to coder models specifically.
Curated coder model list in the future models.json.
No reference app exists yet (existing reference apps are voice / transcribe / chat / docvault).

text-to-code