Local Weights · curator journey · concept

Curating a managed profile

A curator selects models and runtime settings for a team. This document explores tools for comparing candidates, recording benchmark evidence, editing tiers, and publishing signed releases. The benchmark and Radicle COB workflows are proposals; the maturity tags distinguish current code from possible future work.

today implemented near possible next step speculative requires new infrastructure

A profile tier identifies the model, runtime artifact, arguments, environment, and context size. Benchmark results are useful only when they refer to that exact configuration. If the pinned runtime artifact also contains llama-bench, the manager could run the matching benchmark tool and record the result.

tested unit = (model sha256 · artifact build · argv · env · context)  →  signed benchmark result  →  evidence considered by the curator  →  signed profile release

Step 1 · compare candidates

Review results for the hardware you support.

The curator needs a model that fits several hardware classes. A comparison view can show throughput, peak memory, configuration, and the number of submitted runs. Results with limited evidence should be clearly marked.

implementation — proposed benchmark COBs speculative would live in the registry's Radicle repository and be signed by each submitter's DID. The dashboard could aggregate local copies of those records and show medians, sample counts, and raw submissions.

Curator (concept)
Candidates · code-assistillustrative data · apple-silicon · 32 GB
Tested unittok/speak memevidence
Qwen3.7-32B · IQ4_XSb10112 · metal · fa=on · kv q8_0 · 128k 44.224.9 GB 14 nodes · 31 runs fits
Qwen3.6-35B · IQ4_XSb10034 · metal · defaults · 128k 38.725.4 GB 9 nodes · 22 runs fits · current
Devstral-2-30B · Q4_K_Mb10112 · metal · fa=on · 96k 41.027.8 GB 2 nodes · 3 runs thin evidence
GLM-5-Air · MXFP4b10112 · metal · fa=on · 128k 47.531.2 GB 6 nodes · 11 runs no headroom @32 GB
Bench on my hardwareRequest fleet runsView raw COBs

Step 2 · run a benchmark

Test the exact configuration locally.

The curator acquires the candidate through the normal verified path, runs llama-bench from the pinned runtime artifact, reviews the result, and can choose to publish it. A separate request workflow could ask other machines to run the same test.

implementation — acquisition and verification today; running llama-bench from a pinned artifact near; measurement requests and unattended fleet scheduling speculative. A result would include the tested configuration and the hardware facts already collected by the tier matcher.

cob · xyz.ccweights.benchsigned · did:key:z6Mk…kobina
tested_unit:
  model: sha256:2fdd…58d3 (Qwen3.7-32B IQ4_XS)
  artifact: llama-b10112-macos-arm64-metal
  argv_sha256: 9c41…07aa · ctx: 131072
hardware: macos · arm64 · m4-pro · 48 GB
results:
  pp2048: 512.4 t/s · tg128: 44.7 t/s
  peak_rss: 24.9 GB · ttfb_64k_ctx: 8.1 s
tool: llama-bench (from pinned artifact) · harness: ccw-bench/1

Pinning the configuration makes the run repeatable. Results can still vary with hardware state, software outside the bundle, and benchmark conditions, so raw runs remain available for review.

Step 3 · authoring

Edit the tier beside its supporting results.

The editor shows the selected model, runtime, context, arguments, and environment next to available fit and performance results. It uses the same schema validation as the signed release tooling.

implementation — schema validation, reserved-flag checks, and memory-range checks today; estimated memory fit calibrated against measured peak RSS near; evidence sidebar speculative.

Curator (concept)
code-agent · tier: apple-silicon-32-to-63gbdraft v3
Tested unit
Qwen3.7-32B-IQ4_XS.gguf · sha256 2fdd…58d3
llama-b10112-macos-arm64-metal
131072
--jinja --flash-attn on
--cache-type-k q8_0 --cache-type-v q8_0
GGML_METAL_NO_RESIDENCY=1
Evidence · 14 nodes
fits: est. peak 24.9 GB vs 32 GB floor (7.1 GB headroom)
44 tok/s median on tier hardware (31 runs)
args valid for b10112 · no reserved flags
64k-context first-token: 8.1 s — note in release?
3 fleet volunteers reproduced within 4%
Attach evidence & save tierBench this exact unit

Step 4 · release

Review, sign, and publish the release.

The release view shows the changes by hardware tier, the evidence used, and the note developers will see. The curator reviews the diff before signing a new rollback-protected release version.

implementation — signing, versioning, and rollback protection today; a release-note field near; publishing through Radicle and linking evidence COBs speculative.

Curator (concept)
Release · code-agentv2 → v3
TierChangeEvidence
apple-silicon-64gb-plusQwen3.6 Q4_K_M → Qwen3.7 Q4_K_M · fa=on, kv q8_011 nodes
apple-silicon-32-to-63gbQwen3.6 IQ4_XS → Qwen3.7 IQ4_XS · +14% tok/s14 nodes
linux-strix-halo-64gbunchanged5 nodes
Qwen3.7 improves tool-call performance. At 64k context, first-token latency is about 8 seconds.
Sign & publish release 3Dry-run against fleet

Step 5 · review the release

Review adoption, reports, and new benchmark results.

After release, the curator can review update status, diagnostics shared by developers, and benchmark results for the released configuration. Differences by hardware class can become inputs to the next revision.

implementation — copyable diagnostics near; report COBs, adoption reporting, and a regression view speculative. Reports should omit prompts, source code, local paths, and hardware serial numbers.

Curator (concept)
code-agent · release 3day 6
SignalDetail
adoption9 of 11 fleet machines updated · 7 via LAN, 2 via HF
regressionm1-max · 64 GB: tg128 down 22% vs release evidence (2 nodes agree)view COBs
report1 × port conflict (foreign process) · repair succeeded after rebootview

The M1 Max result can be investigated and, if confirmed, added as a test case for the next release.

Independent publisher

Publishes a profile for hardware or workflows they know well.

  • Public distribution through Radicle seeds or HTTPS mirrors.
  • Trust on first use from an independently verified key; contribution history and repository delegates provide additional context.
  • Forks and patches allow others to change a tier or propose an update.
  • Broader hardware coverage can provide results for configurations that a small team does not own.

same schema · same signing · same verification · same app

Organization curator

Maintains approved profiles for managed machines.

  • Private Radicle seeds or HTTPS hosting. The registry signature remains the trust root.
  • Keys distributed through device management; benchmark submissions may be limited to approved DIDs.
  • Release records identify the approved profile, signer, version, and supporting evidence. Profile policy can restrict redistribution.
  • Structured diagnostics can be associated with the release while excluding prompts and source code.

same schema · same signing · same verification · same app

Possible community contributions

DataUseLimits and checks
Benchmark results (tok/s, peak memory, TTFB)usefulPin the tested configuration and tool, retain raw runs, show sample counts, and separate submitters. A signature identifies the submitter; it does not prove the result is accurate.
Hardware compatibility (for example, OOM on 32 GB)usefulRecord the exact configuration and hardware class. Confirm important results with more than one machine where possible.
Evaluation scoreslimitedComparable only when the harness, dataset, prompts, and scoring version are pinned. Results may be expensive to reproduce and easier to optimize selectively.
Argument and tuning proposalspatchesSubmit changes as repository patches with benchmark or compatibility results. The curator reviews and decides whether to include them.
Selection, signing, and release notescuratorCommunity results are inputs. The curator remains responsible for the selected profile and signs the release.

Capabilities this concept assumes

  1. A tested-unit digest near — a canonical hash over the model, artifact, arguments, environment, and context size.
  2. llama-bench integration near — run the benchmark tool from the pinned runtime artifact against an acquired model.
  3. Versioned COB schemas speculative — separate records for benchmark results, reports, and measurement requests.
  4. An explicit identity boundary speculative — registry release keys and Radicle DIDs remain separate. Repository delegates may manage distribution and key rotation without becoming the release trust root.
  5. Memory-fit estimates near — use measured peak RSS to improve estimates for each hardware tier.
  6. Profile discovery speculative — a catalog could show evidence volume, independent reproductions, publisher identity, and trust information.
  7. A privacy-limited diagnostics schema near — exclude prompts, source code, local paths, and serial numbers.