Issue Framework 001 / Evaluation method
Small Model Review
Less model worship. More useful systems.
A Recollective-owned publicationAI-assisted production / release governed by Recollective
Latest story01
Recollective owns and operates this editorial channel
Evaluation Methods
A 209M-Parameter Forecast Still Needs a Local Baseline
Timestral’s benchmark position is a reason to test it on a real forecasting task, not a substitute for the baseline that already informs the decision.
Published October 21, 20266 min readSmall Model Review Editorial · a Recollective channel
Channel archive
More from Small Model Review
02
Evaluation Methods
Inference Cost Is Only One Line of the Small-Model Bill
A compact checkpoint can still leave substantial work in data preparation, review, escalation, and operations; a five-card review ties those costs to useful outcomes.
October 21, 2026 · 6 min read03Workflow Tests
Can a Tiny Vision Model Read the Receipt?
SmolVLM makes compact image understanding tangible. The useful comparison starts with the smallest detail you cannot afford to lose.
October 15, 2026 · 4 min read04Evaluation Methods
Four Labels, One Small Model
Before building a conversational assistant for the inbox, consider the quieter job of putting a message in the right pile.
October 15, 2026 · 4 min read05Workflow Tests
A Browser-Local Model Needs a Device Matrix
Browser execution moves performance, support, updates, and the data boundary onto the user’s hardware. A five-card review makes the operating point reproducible.
September 14, 2026 · 7 min read06Evaluation Methods
The Long-Context Reliability Test Needs Three Positions
A context window tells you how much input a model can accept. It does not tell you whether the answer-bearing detail will be found, cited, or worth the review time.
September 10, 2026 · 5 min read07Evaluation Methods
Open Weights Are Not a Local Deployment Plan
An open-weight file becomes a responsible local system only after its memory, terms, runtime, and data boundaries are documented.
September 4, 2026 · 6 min read08Evaluation Methods
Quantization Needs Three Scoreboards
Smaller weights can free memory, but a trustworthy choice requires separate evidence for resources, latency, and the task itself.
September 4, 2026 · 5 min read09Workflow Tests
The Escalation Cascade Needs a Ledger
A small model is useful only when the router knows what it may accept, when to escalate, and what a human still must review.
September 4, 2026 · 7 min read10Review Framework
Smaller Models, Better Questions: A Review Framework for Everyday AI Work
The useful question is rarely “Which model wins?” It is what level of capability, oversight, and cost this job requires.
August 29, 2026 · 7 minute frameworkEditorial purpose
An evaluation desk for studying models as parts of a complete workflow. We care about task fit, review burden, operating cost, and the shape of failure more than leaderboard mythology.
01
Task fit
Representative work is tested before a tool is praised for general capability.
02
Review burden
A fast draft can be expensive if plausible mistakes demand line-by-line reconstruction.
03
Failure shape
Obvious failures are often safer than polished errors that slip through review.