A smaller coding model needs its own runtime

  • NariQ shrinks Qwen3.8-27B for local coding work — Nari Labs targets coding quality on local hardware, with a custom runtime and additional memory requirements to check before downloading.
    • Its compressed models have weights stored at lower precision to fit in less memory. Three-, 2.5-, and two-bit files occupy about 10.1, 8.4, and 6.8 GB. Calibration uses traces from an agent, a model running a multi-step task with tools.
    • On DeepSWE, the paper reports 47 fully solved tasks for three-bit weights versus 44 for full precision, out of 113. Each task gets one attempt. The two-bit file solves 33, showing the smallest download’s quality cost.
    • Terminal-Bench 2.1 mean test reward is 63.8% versus 62.6% for full precision, across 82 tasks with three attempts each. A benchmark is a fixed set of tasks for comparing models. These small differences remain within reported uncertainty.
    • The GGUF files, a model-file format for local runtimes, require the Nari llama.cpp fork. Stock llama.cpp, Ollama and LM Studio currently lack support for its new tensor types. Apple Silicon requires building from source.
    • For three-bit weights at the full 262,144-token context, the model card estimates roughly 30 GB of GPU memory, including vision. Tokens are the pieces into which models divide text. File size alone therefore understates the memory needed for long sessions.
    • Caveat: The fork is tuned for RTX PRO 6000-class Blackwell GPUs, while other listed GPUs and Apple Silicon currently run slower than comparable files. (arXiv 2611.26157 · model card, full paper, runtime repository and release assets) · Paper notes

Keep screen-reading assistants out of private chats

  • Signal adds privacy controls against AI agents reading chats — Signal is restricting screen capture as operating systems give AI assistants more access to private conversations.
    • On iOS 27, Signal blocks screenshots to obstruct AI capture. Users can disable the protection, accepting the risk that assistants capture their own and their contacts’ messages.
    • On supported Android devices, Signal automatically enables “Screen security” when it detects an AI assistant attempting screen access. Detection differs across Android versions and vendors.
    • Signal already blocks Windows Recall screenshots by default. The new announcement extends its response to mobile operating-system assistants.
    • Signal says reliable screenshot blocking remains unavailable on macOS. Its recommendation for macOS 27 is adding Signal under System Settings → Siri → Excluded Apps, protecting against proactive Siri access.
    • Caveat: Screen blocking leaves other access routes, such as device filesystem contents, outside these protections. (Techmeme · Bluesky · Signal’s primary release and platform-specific limitations)

Watch several coding assistants from one browser

  • DreamCoder puts coding conversations, terminals and reviews in one workspace — Mickey Petersen’s locally running application helps users monitor parallel coding work without repeatedly switching terminal tabs.
    • Its browser interface tiles conversations, terminals and reviews. A workspace minimap shows which panes are running, waiting for attention or finished. An event graph links prompts, actions and results for inspection.
    • It supports hosted and local models, background workers, reusable instructions and configuration import from Pi, OpenCode or Codex. The author explicitly describes it as a standalone application, separate from Emacs.
    • Pricing is free for personal use and businesses with up to five users. Once a business exceeds five, every user needs a $20/month seat, covering up to five installations per person.
    • Downloads cover Apple Silicon macOS, Windows, Debian/Ubuntu Linux and Docker. The server runs on your own machine and is reached through a browser, including from a phone.
    • Caveat: A local application can still send prompts and project content to whichever hosted model provider you configure. (Lobsters · full launch account, product documentation, current pricing and downloads)

Remember enough to finish the task

  • CompAct-Bench tests whether agents finish after their conversation is shortened paper notes → — The test checks whether compressed conversation preserves the working state an agent needs to finish.

    • Its 200 tasks cover web research, repository coding, and a simulated document workspace. Each task family uses a fixed executor after 12 models compress earlier conversation to 10% of its original character length.
    • The strongest compactor achieved 68.5% completion accuracy, averaged across three independent continuations. Every source task had first been solved without compaction, making this a measure of lost continuation ability on a selected set.
    • Compression sometimes retained a fact but changed its meaning: an unverified lead became an established answer, or a ruled-out option became a candidate again.
    • Coding survived compression better when edits were already saved in the repository. Testing resumed tasks and preserving verification status can catch failures that a readable summary conceals.
    • Caveat: The aggressive 10% character budget and fixed executors limit how directly these losses apply to other production setups. (arXiv 2611.25673 · full methods, task screening, results and continuation-error analysis)
  • GrepMem combines compact summaries with searchable original history paper notes → — Retaining original interaction records lets an agent recover details that have fallen out of its active conversation.

    • After compaction, grep_history searches an append-only log of messages and tool results. Reinforcement learning, training from task rewards, improves task actions, summaries and recall together. The study uses Qwen3.5-9B with a 32,768-token context.
    • On BCP-Multi, which combines three research questions in one session, trained GrepMem scores 46.9% versus 38.1% for summary-only SUPO. This tests retaining findings while investigating other questions.
    • On ContextBench’s log-recall task, it scores 95.3% versus SUPO’s 6.6%. Even before training, GrepMem scores 75.0% there. Exact lookup gains therefore come from both searchable history and improved use of it.
    • The implementation and reproduction instructions are public. Extra retrieval can be costly: on AMA-Bench’s streamed trajectories, it uses 2.3× SUPO’s tokens for a small, statistically uncertain accuracy gain.
    • Caveat: The results cover one model and the repository still lists trained checkpoints as pending, so its base-model quickstart differs from the scored checkpoint. (arXiv 2611.25123 · full paper, implementation, training protocol and task-specific cost comparisons)

Check agent reports and serving capacity

  • FaithDetector checks coding agents’ reports against execution evidence paper notes → — Comparing final claims with recorded actions offers a way to catch statements such as tests passing without supporting evidence.

    • Researchers annotated 435 sampled work reports from SWE-chat. They found at least one unfaithful claim in 162 reports, or 37.24%, including claims that exceed what the recorded work establishes.
    • The study identifies eleven subtypes in four categories, including wrong observations, misplaced evidence, overstated success and faulty explanations. Handling challenged reports averaged 4.3 additional user turns and 12.71 minutes of recorded response and execution time.
    • Across 2,356 claims from 409 reports, FaithDetector reached 93.06% macro-F1 versus the baseline’s 80.42%. Macro-F1 averages the scores for faithful and unfaithful claims equally. Both use the same model and pre-report records; the detector adds structured, taxonomy-guided checks.
    • A replication package includes data, annotations, detector code and evaluation outputs. The detector adds a review aid alongside rerunning tests and inspecting the actual artifact.
    • Caveat: These public repository sessions may differ from private enterprise work, and the recorded handling time includes waits rather than measuring human labor. (arXiv 2611.25760 · full study, sampling limitations, detector evaluation and replication-package listing)
  • A serving study replays complete agent sessions to test capacity paper notes → — Replaying recorded workflows offers a repeatable way to size inference servers while keeping live agent testing as a separate quality check.

    • Researchers released 10,056 sessions, containing 241,674 model requests across six benchmarks, five models and five agent frameworks. Traces retain complete messages, request settings, timestamps and tool calls—the model asking a program to do something.
    • In two 40-task comparisons, replay’s 95th-percentile input-token load stayed within 1.18× the recording. Fresh live agent runs reached 4.98×. These figures describe token-volume drift, not elapsed time or a speedup.
    • In a separate four-model deployment, captured workloads reached about half the optimal concurrent-request count and 60% of the token-processing rate seen with public traces. Capacity therefore depended heavily on which workload was used.
    • The inference-perf guide documents replay that infers request dependencies and passes new model responses into subsequent requests. Production traces need privacy review before export because they contain messages and tool data.
    • Caveat: Replay reuses recorded tool outputs and can force tool selection, so its stable load partly follows from constraints absent in live task execution. (arXiv 2611.25905 · full experimental methods, trace dataset and replay documentation)

The research measurements above are author-reported preprint results.

All gathered items - what was cut and why (8)