cd /news/artificial-intelligence/cloudflare-clairic-open-weights-upda… · home › topics › artificial-intelligence › article
[ARTICLE · art-172220] src=explainx.ai ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Cloudflare Clairic Open Weights: Updated 27B and New 9B Clairic-Flash

Cloudflare released open-weight Clairic (27B) and Clairic-flash (9B) extraction models on November 18, 2026, hosted on Workers AI and published on Hugging Face, graduating Clairic from preview and adding a 9B flash tier. Cloudflare reports both new models scored 1,263/1,263 correct fields on internal tracked assertions versus 1,033/1,263 for the preview, and 576/576 on synthetic receipt fields versus 369/576 for preview, with image extraction errors dropping from 207 to zero on an internal mixed-input set. Cloudflare measured p50/p95 full HTTP response times of 358/736 ms for Clairic and 155/328 ms for Clairic-flash on fresh text fields, and 453/976 ms and 305/605 ms respectively on fresh image fields, charging input tokens only with no separate output-token meter.

key takeaway

Clairic-flash (9B) returns fresh text fields at 155 ms p50 versus 358 ms for the 27B Clairic, and both models bill input tokens only with no separate output-token meter.

by read7 min views3 publishedNov 18, 2026
Cloudflare Clairic Open Weights: Updated 27B and New 9B Clairic-Flash
Image: Explainx (auto-discovered)

Cloudflare released open-weight Clairic and Clairic-flash on November 18, 2026, hosted on Workers AI and published on Hugging Face, with the Workers AI Superintelligence team branding the launch. The update graduates Clairic from preview, adds a 9B flash tier, and re-architects extraction for PDFs, images, tables, and dates.

If you already use System One models to shrink chat bills, this is the extraction-specific release that pairs with Clef for routing and decisions.

What did Cloudflare announce? #

The November 18 blog post names two deliverables:

Model Role Workers AI ID
Clairic (27B) Flagship typed extraction @cf/cloudflare/clairic
Clairic-flash (9B) Faster, cheaper extraction @cf/cloudflare/clairic-flash

Both ship as open weights and managed endpoints. Cloudflare says the mission is purpose-built small models that augment expensive LLM usage rather than replace frontier chat models entirely.

What is System One extraction? #

System One is Cloudflare's label for narrow models that return typed values — numbers, dates, enums, spans — instead of free-form text. Our earlier claimed-vs-verified Clairic preview piece walked the preview API.

The November 18 release keeps the schema-first contract:

  • Confidence scores andcharacter spans per field
  • Structured fields you define in the request
  • No separate output-token meter ; Cloudflarecharges input tokens only

That pricing shape matters when you extract hundreds of fields from invoices, leases, or support tickets.

Architecture changes since preview #

Cloudflare says preview feedback pushed a full redesign:

  • Parallel slot autoregressive generation for ordinary text and number fields, so each fieldstops at its own end marker while the batch advances together
  • Special heads forestimates, forecasts, durations, calculated dates, and arithmetic
  • PDF and image ingestion beyond thereceipt photo demo from preview

On internal tracked assertions, Cloudflare reports 1,263/1,263 correct fields for both new models versus 1,033/1,263 for preview. On synthetic receipt fields, it reports 576/576 for Clairic and Clairic-flash versus 369/576 for preview.

Treat those tables as vendor evals, not independent benchmarks.

Cloudflare Clairic internal eval scores for text image PDF and receipt extraction

Image: Cloudflare

Receipt and image extraction example #

Cloudflare Clairic open-weight model receipt extraction comparison versus preview

Image: Cloudflare Cloudflare's side-by-side coffee shop receipt shows preview returning null for the merchant name while Clairic returns Coffee House with all five fields filled. Image extraction error counts moved from 207 errors to zero on an internal mixed-input set, per the blog.

For operators, the lesson is validate on your document distribution. Receipt OCR is not the same as scanned contracts or noisy mobile photos.

Latency: is Clairic fast enough for online flows? #

Cloudflare measured full HTTP response time on a self-hosted server and published p50 / p95 in milliseconds:

Eval set Clairic p50 / p95 Clairic-flash p50 / p95
Fresh text fields 358 / 736 ms 155 / 328 ms
Fresh image fields 453 / 976 ms 305 / 605 ms
Commercial leases 691 / 803 ms 379 / 519 ms
Loan pricing 395 / 407 ms 188 / 192 ms

Clairic-flash targets sub-second responses on several sets below 400 ms p50. Full Clairic still lands under one second on many rows, which Cloudflare contrasts with typical chat completions on similar workloads.

Pricing and the December 1 promo #

Cloudflare moved the 9B preview endpoint to Clairic-flash and placed the 27B model on the main Clairic ID.

| Model | List input price after promo | Promo through Dec 1, 2026 |

|---|---|---|
| [Clairic 27B](https://developers.cloudflare.com/workers-ai/models/clairic?ref=explainx) | $0.24 per million input tokens | $0.10 | 

| Clairic-flash 9B | $0.10 per million input tokens | $0.10 |

If you pinned preview Clairic at $0.10, Cloudflare says the endpoint upgrades in place to better extraction without a schema migration. Cost-sensitive jobs should test Clairic-flash before December 1, when 27B returns to $0.24. Compare this shape to OpenAI Decisions API and Microsoft Decision-1.1 when you route versus extract.

Open weights on Hugging Face #

Cloudflare follows Clef's open-weight playbook: publish weights for research and fine-tuning, keep managed inference on Workers AI.

Reasons builders download weights:

- **Air-gapped** document pipelines
- **Fine-tunes** on**industry-specific forms**
  • Latency experiments onyour own GPUs

Reasons to stay on Workers AI:

  • No kernel tuning
  • Billing integration withWorkers andR2
  • Automatic model router updates

Check license files on Hugging Face before redistributing derivatives.

Competitive context: extraction is crowded #

Cloudflare compares against Liquid AI D1.2 with Apple Vision OCR and Scaledown Extract in its tables. It admits no public extraction leaderboard exists yet.

That honesty matters. Extraction is becoming a commodity layer as agents ingest email, PDFs, and tickets. Jev and Microsoft's Decision models sit adjacent but optimize decisions, not field filling.

If you run agents on Cloudflare, validate every extracted field before it triggers a tool call when untrusted documents enter the loop.

How to try Clairic this week #

  1. Open Workers AI docs for**@cf/cloudflare/clairic** and**@cf/cloudflare/clairic-flash** .
  2. Port one production form withfive to ten fields and comparepreview-era accuracy.
  3. Load-test****Clairic-flash ifp95 under 400 ms is ahard SLA .
  4. Download Hugging Face weights only ifself-host TCO beats**$0.10/M** promo math onyour volume .

Workers AI Superintelligence team #

Cloudflare reframes its frontier research group as the Workers AI Superintelligence team, responsible for Clef, Clairic, and unannounced models.

For edge developers, the takeaway is Cloudflare building models, not only hosting others' checkpoints. That affects router defaults and which open weights get first-class Workers integrations.

PDFs, leases, and multi-page line items #

Cloudflare's PDF support is the largest user-visible upgrade since preview. You can request supplier, line items, and totals when fields span pages. The cap is four images or PDF pages per request, which pushes designers toward chunking very large packets.

For **commercial leases**, Cloudflare's latency table shows **691 ms p50** on full Clairic and **379 ms** on flash. **Loan pricing** rows are **tighter** — **395 ms** versus **188 ms** — suggesting **numeric tables** compress well in the **9B model**.

If you migrate from **general chat extraction**, benchmark **three document types** you actually ship: **invoices**, **contracts**, and **support exports**. Cloudflare's **tracked customer assertions** row is the closest proxy to **messy production**.

Self-hosted vs Workers AI: a cost sketch #

Open weights enable private cloud deployment, but total cost is GPU hours + engineer time, not only $/M tokens.

A rough decision tree:

  • Under a few million fields per month — stay onWorkers AI during the**$0.10 promo** , especially if you already pay forWorkers andR2 .
  • Strict data residency — downloadHugging Face weights and run onyour VPC , acceptinghigher p95 until you tune kernels.
  • Burst-heavy batch — compareClairic-flash****throughput againstbatching onyour own vLLM cluster.

Cloudflare's input-only pricing simplifies math: you do not multiply by output tokens when schemas return large JSON.

Pairing Clairic with agents and routers #

Most production stacks should not send every user message to Clairic. Typical layering:

  1. Clef or Decision model — classify intent and pick tools
  2. Clairic — filltyped slots fromattachments
  3. Frontier chat model — synthesizenarrative answers only when needed

That pattern mirrors Microsoft IQ plugins and OpenAI Decisions API stories from the same month.

Eval tables: how to read Cloudflare's numbers #

Cloudflare publishes fresh text, mixed-input, customer assertions, and synthetic receipt rows. Clairic and Clairic-flash tie on several perfect scores, while preview lags.

When vendors compare against Liquid D1.2 + OCR and Scaledown Extract, note different OCR pipelines. Apple Vision on D1.2 is not identical to Clairic's native image heads.

Until a neutral leaderboard exists, treat comparisons as directional and run your own 200-field golden set.

Security and untrusted documents #

Extraction models read attacker-controlled PDFs. That is indirect injection surface when an agent auto-invokes Clairic on email attachments.

Combine schema validation with agent guardrails when untrusted files enter the loop. Typed JSON does not stop malicious field content from flowing downstream.

What developers are saying #

The Hacker News thread was quiet at launch, with few hands-on reports yet. When builders post latency numbers from real PDFs, treat them as complements to Cloudflare's self-hosted measurements.

Migration checklist from Clairic preview #

If you integrated @cf/cloudflare/clairic during preview, Cloudflare says weights and schemas roll forward without renaming your Workers binding. Still run this checklist before December 1:

  • Re-benchmark ten production documents; previewnull fields may disappear.
  • Switch latency-sensitive routes toClairic-flash and comparep95 against yourSLA .
  • Pin Hugging Face revisions if youself-host , becauseopen-weight tags can move independently ofWorkers AI defaults.
  • Update cost dashboards when27B returns to $0.24 per million input tokens.

Document schema versions in your repo so finance can reconcile November promo spend against December list prices.

Pricing and model IDs reflect Cloudflare's November 18, 2026 announcement. Promo pricing ends December 1, 2026 unless Cloudflare extends it.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @cloudflare 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/cloudflare-clairic-o…] indexed:0 read:7min 2026-11-18 · —