Implementation
Dharma AI: a custom vocabulary cut a 4B grader's cost 77%
Dharma AI says rebuilding the tokenizer of its Qwen3-4B essay grader cut measured inference cost 77.0% and lifted its bench score from 0.703 to 0.810.
Small models
1 article and 2 community signals on Small models for people running AI in production, written and curated by Wren with every source linked.
Updated Oct 22, 2026
Articles
What practitioners are saying
All signals →Cloudflare: Clairic, a preview extraction model on Workers AI that returns typed values with no output tokens
Cloudflare says Clairic returns extracted, computed, estimated and forecast values with confidence scores in one pass, billing input tokens only. In its damaged-claim test a Clairic, Clef and GLM-5.3 cascade took 19.3 s against 61.7 s and cost $6,032 against $11,740 per million claims. TSA forecasts were off 3.3% on average over 21 days.
Why it matters Vendor-reported on one task, with no Clairic price or public accuracy benchmark; Cloudflare warns arithmetic and forecasting are weak. Test it on 200 of your own documents first.
Cloudflare adds Clef-omni and cuts Clef-flash price to $0.038 per million input tokens
Cloudflare says Clef-omni, an open-weight decision model taking text, image, audio and video, launched at $0.15 per million input tokens, and that Clef-flash fell from $0.09 to $0.038 per million while its hosted context window dropped from 64k to 24k tokens. It reports Clef up to 2.0x faster after moving to SGLang. All figures are Cloudflare's own.
Why it matters A reported price cut that comes with a smaller hosted context window; if you route classification or tool-choice calls to Clef-flash, check your longest requests against 24k tokens (Cloudflare says 0.24% of its requests exceed that).