Reika and Jeff-Code show small local models work best with a harness built around them

Reika is a coding agent command line tool for local and hosted models. It was published on GitHub on 5 October. It is built around small models first: 8B to 35B parameters, often run at Q2 to Q4 on a 16 to 32k context window. The author, Alex Leung, says the design goal is not to make a small model smarter. It is to make working with one less frustrating.

That framing matters for anyone running a coding assistant on a laptop or a single GPU. If the harness, not the model, sets most of the failure cost, then the cheapest upgrade is a better loop. The project is pre-1.0, so commands and settings can still change.

The most useful part of the Reika release is a published record of measurements. One mid-context rewrite re-processed 8,453 tokens of prompt, costing 7.9 minutes of prefill. An append in the same session cost 25 tokens and 3.5 seconds. In an A/B test, one arm dropped its findings before a fold. It folded five times, re-read files it had already read and never answered. The arm asked for its findings first answered from the digest. Reika's own account says a grep returning "0 matches" for a call it could not serve sent a model rewording a correct pattern for five rounds.

Reika reports a split in model behaviour by size. In testing, 14B to 35B models handled multi-file tasks most reliably. 8B to 9B models managed simple well-scoped ones, and models below 7B rarely got far.

Prompt wording moved little. On a vague task at Q2, a model converges or spirals about 50/50. The harness cannot change that rate, only what the failing half costs.

A second harness, a different trick

Jeff-Code was also published on 5 October. It forks Pi, the coding agent by Mario Zechner and contributors, and puts a 0.8B decision model called Jeff inside the loop alongside Qwen 3.8-27B. Around every Qwen turn, Jeff makes two decisions of about 0.2 seconds each. The first is whether it can take the next step itself (reading a file, listing a folder, searching code, running tests). The second is whether Qwen needs full thinking on this turn. Writing and editing files always stay with Qwen.

The results are narrow but honest. Across 1,242 paired tasks, pass rate was 62.4% with Jeff-Code against 62.8% for Qwen alone. That is a paired difference of minus 0.2 points, with a 95% interval of minus 2.6 to plus 2.1. Average time per task fell to 0.68x, with a median of 0.70x. Total time across all tasks was 0.86x, because in about 5% of tasks Jeff-Code ran more than 30 minutes longer than Qwen alone.

Some benchmarks showed no clear gain: Terminal-Bench 2.0 at 0.96x and SkillsBench at 0.91x. Turning thinking off on every turn was faster still but cost 7.6 points. A less cautious Jeff at a 0.7 threshold cost 4.5 points. That is the trade the release makes explicit: speed comes from skipping thinking, not from removing it.

The same week brought a decision model pitched at the same problem. Liquid AI introduced d1 on 5 October, its first decision model, now with vision. It returns probabilities without generating tokens at 200 to 300 ms per text decision. Liquid says d1 matches or beats GPT-6.1 Sol on four of six tested applications and costs 19x to 200x less. In one demo it compacts a coding agent's context, removing 52% of tokens while keeping every output the task needs. Those are vendor figures and have not been independently checked.

Elsewhere the on-device case is being made with different numbers. Interfaze 1.1 Lite was published on Hugging Face on 5 October. It runs on a single 80 GB GPU with no external services and transcribes a 95-minute recording in about 90 seconds. Its benchmark table puts MMMU-Pro at 73.2 against 71.1 for Interfaze and 40.4 for GPT-5.4-Mini. It trails on GPQA Diamond at 85.9 against 92.4.

The open question is how much of this survives contact with real repositories. Neither Reika nor Jeff-Code reports results on a standard frontier coding benchmark at full settings. Both are early projects. What they do show is that context handling, not raw parameter count, is where small local models lose time.