SShortSingh.
Back to feed

POCKET-27B-Pi: 27B coding agent packed into a 16.5 GB file for a single 24 GB GPU

0
·3 views

VIDRAFT has released POCKET-27B-Pi, a coding agent distributed as a single 16.5 GB GGUF file in Q4_K_M quantization. On the partner's own Terminal-Bench 2.1 run under the pi agent, it solved 28 of 48 tasks (58.3%), compared with 19 of 48 (39.6%) for the base Qwen3.8-27B model. The model keeps key-value cache in only 16 of its 64 layers, so a 64K-token context needs about 4.0 GiB in f16 and less when quantized. By the article's arithmetic, weights plus a 64K cache total roughly 19.5 GiB, which fits on one 24 GB GPU. The benchmark figures are VIDRAFT's own measurements, and the memory calculations were made by the article's author, who is affiliated with the partner.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Russell's Silverstone Pole Lap Analysed Using Aligned Telemetry Streams

A personal educational project replays historical 2024 British Grand Prix qualifying telemetry from the public OpenF1 API as a stream, rather than connecting to a live car feed. George Russell took pole at Silverstone in 85.819 s, 0.1710 s ahead of Lewis Hamilton, but a finish-line gap alone does not show where on the 5.89 km circuit that advantage was gained or lost. Because car telemetry arrives asynchronously by timestamp, comparing drivers at equal times places them at different track positions, so the data must be remapped onto a common distance grid. The pipeline uses a producer, Redpanda messaging, and a stateful consumer that projects samples onto a 1.0 m grid to build a continuous delta curve. The author also describes validation checks that initially produced suspiciously clean results and implausible segment speeds, which prompted a stricter re-examination of the data.

0
ProgrammingDEV Community ·

Running Several Claude Code Sessions in Parallel Without Conflicts

The article advises giving each Claude Code session a single task, its own set of files, and a separate git worktree so parallel work does not collide. It recommends writing short task descriptions that specify the changes, the files involved, and what completion looks like. Sessions should leave hand-over notes in a shared folder so progress can be reviewed in one place rather than through transcripts. Finished branches should be merged and checked together on a schedule, since integration problems tend to surface where pieces meet. The author also cautions against starting more sessions than can be reviewed, because unreviewed work accumulates and risks being merged unchecked.

0
ProgrammingDEV Community ·

How to Raise a SaaS Price Gradually Without Losing Sales

The article advises SaaS founders to raise prices in small increments, changing only one variable per week. It is judged over a full seven-day period, using revenue rather than sales count as the main measure. The authors describe their product Off Grid AI, where cutting the price lifted daily sales from about 3 to about 14 on similar traffic. They then raised the price weekly from $29.99 to $34.99 to $39.99, keeping each step only if revenue rose. Large jumps are cautioned against because they obscure the point where buyers begin to resist.

0
ProgrammingDEV Community ·

Author Reviews His Supermarket Sales Dashboard and Lists Improvements

A data professional revisited a Looker Studio dashboard built on 817,741 supermarket transactions and critiqued its design. He argues that pie charts should be replaced with sorted bar charts, since people struggle to compare slice sizes. He also recommends showing profit margins by district and day rather than sales alone, because high sales can mask thin margins. Unexplained near-zero sales in late December should be annotated as likely missing data, and chart titles should state insights instead of descriptions. The core lesson is that a dashboard should be designed around the decisions it supports.