I built Wisp to investigate a specific problem: a small model that completes code at the cursor and learns to draft several tokens ahead. The work covered the tokenizer, training pipeline, model, export, and evaluation. I trained the weights from scratch using MLX on Apple Silicon.
Project: Wisp Coder 110M
Try it: Wisp Coder — Code at the Cursor. The browser demo lets you edit the code before and after the cursor and generate a suggestion. It runs the 100.7M-parameter standard decoder on CPU, without the separate MTP module, and displays generated code without executing it.
Evidence: July–August 2026 experiments, published with the model. This case study reviews those records; it is not a new training run or benchmark.
Wisp uses a Llama decoder with 108.2 million parameters in total, including its multi-token-prediction module. I trained a 32,768-token byte-level BPE tokenizer on 400,000 documents and trained the model on a five-billion-token schedule.
Two choices support cursor completion. First, 70% of tokenized chunks were transformed for fill-in-the-middle training, giving the model both the code before the cursor and the code after it. Second, a shared prediction module was trained from the first step to predict ahead recursively, with training depth two. This lets the project study how suffix context and speculative drafting interact. It does not establish a deployment speedup. Architecture and training record.
The ordinary decoder exports as a Llama model that Transformers and MLX-LM can load without custom model code. The extra prediction module ships separately, so a generic Llama runtime does not use it automatically. I also published an MLX reference decoder that checks package hashes and requires its prediction path to reproduce the target's greedy tokens. That decoder establishes correctness; it recomputes prefixes and is not optimized for production latency. Published package, reference decoder.
I checked the numerical export against Transformers. The August 3 receipt records matching top-token choices on its 24-token probe, along with a relative maximum logit difference of approximately 0.00138. This is a bounded compatibility check, not a claim of identical outputs for every prompt. The model card quotes a different relative logit delta; these figures come specifically from the linked export receipt. Export verification.
The evaluation records distinguish token correctness, draft acceptance, and completion quality:
| Question | Recorded evidence | Interpretation |
|---|---|---|
| Did the registered decoding runs preserve greedy choices? | 38,400 exact-argmax tokens across 600 policy/document rollouts; zero failed reproductions | Correctness on those frozen runs, not a general code-quality score |
| Did adaptive drafting improve on fixed drafting? | Zero difference on the registered accepted-drafts endpoint | No adaptive-policy advantage established |
| Did Wisp outperform the comparison models on short code holes? | 1,372 holes across five models, or 6,860 generations; no registered superiority test passed | No claim that Wisp beats the comparators |
The short-hole study used one shared CPU float32 harness and an oracle-supplied target character budget. Its absolute exact-match rates were low: Wisp's macro-average was about 1.02%, versus 5.68% for Qwen2.5-Coder-0.5B. I published the adverse result with the examples, generations, and per-row scores. Those conditions do not measure live editor latency or normal interactive completion quality. Comparison protocol and results.
The replay verification is a local attestation tied to registered source, inputs, and checkpoint hashes, rather than an external audit. The model card also records training-data attribution gaps, validation overlap, and a sampler reset; those limits belong with the results. Replay evidence, model card.
I can take an original model through training and release, connect it to existing runtimes, and build experiments that distinguish a working implementation from a measured improvement. Keeping the null and adverse results makes the next engineering decision easier: the current evidence supports further research on the model and drafting design, while completion quality and production latency still need work.