Small local models get the harness treatment: Reika, Jeff-Code and d1 target on-device agents

On 5 October, two GitHub projects appeared with near-identical framing. Reika, a coding agent CLI from developer alexwkleung, was built around small local models first: 8B to 35B parameters, often at Q2 to Q4, on a 16 to 32k context window. Jeff-Code, a fork of Mario Zechner's Pi agent, puts a 0.8B decision model alongside Qwen 3.8-27B. It reports coding tasks 47% faster at the same pass rate.

That matters to anyone running agents on their own hardware. If the agent loop, not the model, decides whether a local run finishes or spirals, small models become usable for work that currently needs a hosted API. Neither project claims frontier results. What is still unknown is how either harness performs outside its authors' own evaluations.

Reika's README publishes the measurements behind its design choices, and they are specific. One mid-context rewrite re-processed 8,453 tokens of prompt, which the project puts at 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 context fold. That arm folded five times, re-read files it had already read and never answered. The arm asked for its findings first answered from the digest. Prompt wording, the project says, is the weakest lever. On a vague task at Q2 a model converges or spirals about 50/50, and the harness cannot move that rate, only what the failing half costs. Reika also says 14 to 35B models handled multi-file tasks most reliably in testing. 8 to 9B models managed simple, well-scoped ones, and below 7B rarely got far.

Jeff-Code's numbers come with intervals, which is rare for a project README. Across 1,242 paired tasks, pass rate was 62.4% with Jeff against 62.8% for Qwen 3.8-27B alone. That is a paired difference of minus 0.2 points, with a 95% interval of minus 2.6 to plus 2.1. Time per task averaged 0.68x, with a median of 0.70x. Total time across all tasks fell less, to 0.86x, because in about 5% of tasks Jeff-Code runs more than 30 minutes longer than Qwen alone. Turning thinking off on every turn was faster still but cost 7.6 points. A less cautious threshold of 0.7 cost 4.5 points.

Decision models as the other path

Liquid AI took a different approach on the same day. Its d1 model, announced on 5 October, answers questions about a situation with a probability per possible answer. It generates no tokens at all, returning results in 200 to 300 ms for a text decision. Liquid says d1 matches or beats GPT-6.1 Sol on four of six real applications it tested, from filtering support tickets to inspecting circuit boards. It does so at 19x to 200x lower cost and faster on every task. The company also reports 85 to 97% accuracy sorting good and defective parts on the public VisA dataset, a task it was never trained for.

The most relevant demo for agent builders is context compaction. Liquid says d1 reads each tool output in a coding agent's session and keeps, trims or drops it. It removes 52% of the tokens while keeping every output the task needs. That is the same problem Reika attacks with context folds, solved by a much smaller model asked a narrower question. Liquid describes d1 as its first decision model and as the first to extend these capabilities to images.

The harness thesis also shows up in a post by Sean Goedecke published on 5 October. Goedecke argues that models like Jev, which only output decisions, can be distilled into hand-built classifiers once a prompt has been validated. Collect the inputs and outputs, then train something smaller and faster for that one task. He notes that training a bespoke classifier is outside most engineering teams' skillset. That is the gap a general decision model fills first.

Not every small model is a decision model. Interfaze 1.1 Lite, published on Hugging Face on 5 October, is a mixture-of-architectures model. It runs on a single 80 GB GPU with no external services. It reports 83.8 on olmOCR-Bench, 85.9 on GPQA Diamond and 3.01 word error rate on VoxPopuli-Cleaned. It transcribes a 95-minute recording in about 90 seconds. That is on-device capability for document and speech work rather than for code.

The open-weight frontier moved in the other direction the same day. Reflection announced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active. It was pretrained on 23.8 trillion tokens and trained with over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks. Reflection says Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks. It says Kimi K3 remains ahead on raw capability. The weights, technical report and model card are promised later this month, and the model is still in red-teaming.

For now, the practical claim from the small-model side is narrower than the marketing around it. Reika is pre-1.0 and says commands and behaviour can change between releases. Jeff-Code's speed-up holds on SWE-bench Verified at 0.63x and Terminal-Bench Pro at 0.64x. It shows no clear gain on Terminal-Bench 2.0 at 0.96x or SkillsBench at 0.91x. The harness helps most where the task is routine, which is also where a decision model fits.