Saved generation responses and a completed historical task battery. Some quality components remained on hold.
Decision / why kept or replaced
Retained historically as the fallback after the faster EXL3 setup failed in production. It did not clear every quality check.
Limits
vLLM. Exact earlier runtime build is not fully bound. The recorded Mia NVFP4 repository was not publicly accessible at this review; the upstream link is a family reference, not a substitute artifact. This is historical use, not proof of today’s routing.
The October 19 qualification passed 12/12 tasks in each of three repeats. Two GB10 nodes are documented; an October 20 read-only capture reports this served model.
Decision / why kept or replaced
Adopted on October 19 after meeting our quality and serving gates. The recorded comparison found a quality tie with the prior configuration and passed the speed gates.
Limits
Not a claim of global superiority. The October 20 serving settings differ from the adoption settings; they must not be combined into one recipe.
Three October 17–18 v3 runs each scored 11/12. Later v4 decision records report 12/12. These are different packs, not directly comparable scores.
Decision / why kept or replaced
Our former production route. The October 19 decision replaced it with DeepSeek after the separate quality and serving checks.
Limits
Repeated prompt-contract tests measure that intervention, not superiority over other models. The exact historical weight revision is not bound to every run.
The current MiniCPM deployment is verified on an Apple M3 Ultra with 96 GiB unified memory. Older Ornith and Gemma runs identify a Studio host class but do not establish each historical run’s chip and RAM. No M3 Ultra 96GB-specific benchmark or fit claim is made for those older runs.
MiniCPM-V 4.7 35B-A3B
Q4_K_M · llama.cpp · M3 Ultra / 96 GiB
Deployment + artifact verifiedCurrent Studio choice · recorded Deployment and artifact verified October 20, 2026
Test / record date: Deployment and artifact verified October 20, 2026
Read-only checks confirmed the M3 Ultra, 96 GiB RAM, loaded model identity and selection by the current workload routes. The 21,166,255,008-byte Q4_K_M file matches the SHA-256 published at a pinned OpenBMB revision.
Decision / why kept or replaced
This is the model we currently select on the Studio for this workload. That makes it our Studio choice, not a claim that it beats every model or hardware configuration. Earlier article trials are recorded separately below.
Limits
Four configured slots each have 16,384 tokens of effective context: 65,536 total is not a 65,536-token per-request allowance. This review checked the deployed identity and artifact without new inference; it is not a fresh comparative quality or speed benchmark.
Runtime reports llama.cpp build b1-358a420. The inspected server uses no reasoning budget.
The repository head now contains different bytes. Use the pinned file link to identify the file we verified, rather than assuming today’s main branch is identical.
Ornith 1.0 35B
Q4_K_M · Ollama
First-party testedTested with failures
Test / record date: Historical run records reviewed October 20, 2026
Coding, instruction following, reasoning and structured output
Evidence
248 trial receipts were reconciled with 248 actual, nonempty HTTP 200 responses. Coding and reasoning critical failures are present in the grading records.
Decision / why kept or replaced
The run history supports a tested entry, not a clean all-purpose recommendation. Returning an answer did not mean the answer passed.
Limits
Studio attribution is recorded; exact historical chip/RAM is not. Do not carry these results over to Ornith 1.5.
These records prove more than downloaded weights or a configured name. They still do not establish every field needed for a hardware-specific recommendation.
16 saved, nonempty response files were checked by checksum on the execution host. The retained 1,709-word draft also matches its recorded checksum.
Decision / why kept or replaced
Actual generation occurred. Human factual review was still pending; a returned draft is not a quality pass.
Limits
This October 16 trial’s exact checkpoint and hardware binding remain unresolved. The current Studio entry is separately verified; its identity is not retroactively assigned to this earlier trial. The source link is a model-family reference only.
Configuration & evidence notes for MiniCPM — earlier article trial, Recorded runtime label: minicpm-v47
The generated response files were verified on their execution host, not assumed missing because they were absent on another machine.
The recorded model profile is minicpm-v47. The exact weights must be bound before a reproducible setup guide or hardware recommendation.
Credit belongs to the model authors, quantizers and runtime developers. These are upstream references, not Workman test receipts. An upstream model card is not a claim that its current weights match every historical run; exact pinned artifacts are labeled separately.
Our recorded pair choice describes an October 19 adoption decision and an October 20 identity capture. It is not a live availability indicator.
Tested, screened, adopted and sustained operation are different claims. A retained historical decision is not proof of today’s routing.
Only configurations backed by first-party run evidence earn a tested entry. Staged weights, configured aliases and zero-call failures do not.
Check the exact artifact’s current license before business use. Successful generation does not establish commercial permission.
No throughput leaderboard or universal recommendation is inferred across these different packs, runtimes and workloads.
We checked the underlying run records and saved outputs. Production prompts, business inputs, private hosts and detailed operating playbooks stay private.
