jrad123777/effect-qwen36-35b-gguf

Effect-v4 Qwen3.6-35B-A3B — Champion (GGUF, llama.cpp)

A local model fine-tuned to write idiomatic, compiling Effect v4 (effect@4.0.0-beta.80) TypeScript. Built $0-local on a single Apple M5 Max (48 GB): continued-pretraining → instruction SFT (LoRA), fused into the base, then converted to GGUF for portable CPU/GPU inference with llama.cpp. Same champion weights (v7s43_i200) as the MLX release — this repo is the portable GGUF build.

Why this exists: effect@4.0.0-beta.80 is a beta that postdates the pretraining of essentially every LLM — its exact API surface is absent from base models, so they hallucinate v3-isms. That sparsity is the whole point: this is a small, honest domain expert for a library the big models haven't seen.

Honest framing first. These are the fine-tuned weights. On a frozen, real-tsc --strict compile gate (24 held-out tasks) they are a genuine but limited expert: single-greedy + RAG ≈ 9.7/24 mean (best checkpoint 13/24). The headline ~23/24 number is the full serving pipeline (best-of-16 sampling + retrieval + a deterministic import-resolver + a tsc verifier), not the bare weights — see How to actually get 23/24. Treat this as a research artifact, strongest when paired with retrieval and a compiler-in-the-loop.

What it is

• Base: mlx-community/Qwen3.6-35B-A3B-4bit — the text tower of the Qwen3.6 hybrid GatedDeltaNet MoE (qwen3_5_moe, 35.9B total / ~3B active). Vision tower dropped (text code model).
• Fine-tune (champion v7s43_i200): CPT on a curated Effect-v4 source corpus (effect-smol, EffectPatterns, examples) → instruction SFT (rank-8 LoRA, 423 gate-validated instruction→code pairs, every target compiled under the exact tsc gate).
• This repo: the champion LoRA fused into the base, dequantized to bf16, converted to GGUF and quantized with llama.cpp (mainline). Verified to generate coherent Effect TypeScript on a raw greedy CPU smoke test before release.

Conversion note (GatedDeltaNet + MoE): converting this hybrid arch from MLX-origin weights required one non-obvious fix — mlx-lm bakes the +1 zero-centered-RMSNorm shift into its saved norm weights, and convert_hf_to_gguf.py adds +1 again, so the norms must be un-shifted before conversion or every layer is double-shifted into garbage. With that corrected, the GGUF matches the MLX model's behavior. (The earlier …-v3-gguf repo predates this fix and is broken — use this repo instead.)

Files / quants

file	quant	size	notes
effect-qwen36-35b-champion-q4_k_m.gguf	Q4_K_M	~20 GB	recommended default — small, fast, smoke-verified
effect-qwen36-35b-champion-q8_0.gguf	Q8_0	~36 GB	near-lossless, for max fidelity

Eval (real tsc --strict, frozen 24-task held-out benchmark)

Raw single-greedy + RAG — honest, same-harness, multi-seed flat mean (never a cherry-picked run). This is the bare-weights number; the ~23/24 headline is the serving pipeline below, not this table:

config	compile@24
base model (no fine-tune)	3 / 24
this model, single-greedy + RAG (flat mean)	9.67 / 24
this model, best checkpoint single point	13 / 24

The dominant residual failure is decoding discipline (the model knows the API — best-of-N reaches 22–24/24 — but greedy decoding sometimes omits a namespace import). This is closed by tooling, not by more training: every $0 in-weights lever (more data, self-distillation, external real-repo data, RAG-tuning, decode-time constraints) was tested and plateaus here. Pushing raw ≥15 needs RL-with-compiler-reward (out of $0-local scope).

How to actually get 23/24

The production pipeline (open-source in the training repo, serve/serve.py) wraps these weights with:

1. best-of-16 sampling (temp 0.8 / top-p 0.95) — tsc is a perfect verifier; keep any sample that compiles,
2. targeted RAG over tsc-gated Effect-v4 idiom snippets,
3. a deterministic import-resolver (fixes TS2307/TS2304 namespace imports),
4. an optional 1-pass tsc-feedback repair.

That stack reaches ~23/24 on the broad served product. This repo gives you the expert weights; add your own best-of-N + a compiler check for production use.

Usage (llama.cpp)

# build/download llama.cpp, then:

# one-shot raw completion
llama-completion -m effect-qwen36-35b-champion-q4_k_m.gguf -no-cnv -n 256 --temp 0 \
  -p 'import { Effect } from "effect"'

# chat (Qwen chat template is embedded in the GGUF)
llama-cli -m effect-qwen36-35b-champion-q4_k_m.gguf \
  -p 'Write Effect v4 code: a Schema.Struct for a User with a branded UserId.'

# OpenAI-compatible server
llama-server -m effect-qwen36-35b-champion-q4_k_m.gguf --port 8080

Tip: for best results, prepend a few real Effect-v4 example snippets (RAG) and sample N times keeping the first that compiles under tsc.

Limitations

• Single-greedy compile rate is ~⅓–½ of hard tasks; pair with RAG + best-of-N + a tsc gate.
• effect@4.0.0-beta.80 only; later betas may shift APIs.
• Reasoning/thinking is disabled — it's a direct code generator.
• Quantized (Q4_K_M / Q8_0). For the native-precision Apple-Silicon build see the MLX repo.

Supersedes jrad123777/effect-qwen36-35b-v3-gguf (an earlier, weaker checkpoint from a broken pipeline).

Built $0-local. Trained, evaluated against the installed .d.ts with tsc as the only arbiter, and documented honestly.