Ling-3.1-flash-1019: Smarter, and Hungrier

Last checked 2026-10-31. Prices and scores move; everything below is dated and sourced. Quick facts verified against the official Hugging Face repo (config.json, model card) and Artificial Analysis's live model page on the day of writing.

What shipped, exactly

On October 27, Ant Group's Ling team (inclusionAI) published Ling-3.1-flash-1019 — a post-training refresh of Ling-3.1-flash, the 560B-total / 25B-active hybrid-linear MoE we profiled at its original open-weights release. The license stays MIT. The team simultaneously dropped Ling-3.1-flash-base in BF16 — an instruction-unfinished base for continued pretraining, RL, or domain fine-tuning — which is the part enterprise people actually care about and most releases skip.

What did not change: the skeleton. We pulled config.json from the official repo and it confirms the same 72-layer hybrid layout (roughly 63 linear-attention KDA layers plus 9 gated full-attention layers, first 4 layers dense), 8 experts active per token, and the bailing_hybrid architecture tag. This is a post-training iteration, not a new model.

The scores, and the asterisk

The model card's claims, which we verified as written on the card itself:

• Artificial Analysis Intelligence Index v4.3.2: 41 → 44
• AA-Briefcase (complex office tasks): ~1400 → 1524 Elo
• AutomationBench-AA: 62% → 68%
• CritPt (physical reasoning): 18% → 26%
• SciCode: 54% → 57%

Here's the asterisk, and we're not going to pretend it away: when we loaded Artificial Analysis's public page for this model on October 31, the displayed Intelligence Index was 29, ranked #4 of 22 — not 44. AA appears to have rebased its v4.3.2 display since the card was written (GLM-5.3-Flash's live page shows a similarly deflated figure versus numbers circulating at its release). We don't know which snapshot will be canonical next week, so we report both: the card says 44 on v4.3.2; the live page says 29 at #4/22. What's stable across both: the direction is up, and the tier is frontier-adjacent for open weights. Chinese coverage additionally claims #4 among same-scale open-weight models and, for the base model, wins on 32 of 47 benchmarks against DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base — both plausible, both single-sourced, neither re-verified here.

The part that decides your bill: verbosity

Ling-3.1-flash was already flagged — by us and by AA — as a talkative model. The 1019 refresh does not fix this. It sets a record. From AA's live page, rendered October 31:

• 240M output tokens per Intelligence Index run — #22 of 22, the most verbose model in the table
• $3.19 cost per Intelligence Index task at list pricing ($0.30 in / $0.90 out / 80% cache discount)
• For contrast, GLM-5.3-Flash in the same table: 120M tokens, $0.58 per scored task, at $0.15 in / $0.50 out

Read that third bullet again. Two models in the same intelligence tier, one costs 5.5× more per unit of scored work — not because its sticker price is 5.5× higher (it's ~2×), but because it talks twice as much to finish the same job. Chinese coverage at release measured the same jump on the old snapshot: ~145K output tokens per task, +40% over the predecessor, pushing per-task cost from $1.01 to $1.49 before AA's rebasing muddied the comparison. The direction is consistent everywhere we look.

Verbosity on a benchmark harness is not the same thing as verbosity in your agent loop. If your prompts ask for one-line JSON verdicts, your mileage will differ from AA's. But if you're routing open-ended reasoning through this thing, budget for the tokens before you budget for the price.

Context window: read the config, not the card

The model card advertises a "1M-token context window." The config file says max_position_embeddings: 262144 — 256K native. Both can be true: the 1M figure requires enabling YaRN-style RoPE scaling, which costs you some fidelity at the far end and requires inference-stack support (the team recommends SGLang for this series). If you need 512K+ contexts in production, test at your target length before committing; the out-of-the-box window is the 256K one.

Deployment reality check

Nothing about deployment got cheaper. The BF16 weights run to several hundred gigabytes, so "self-host" means a multi-card server regardless of the 25B active parameter count. The API route (list $0.30/$0.90 per 1M, $0.06 cached input per transcribed pricing) is the realistic path for most teams, and it is where the verbosity tax lands hardest. Watch for third-party hosts (OpenRouter and friends) reselling below list — that was true for the predecessor and is the cheapest way to A/B the 1019 against your current model without a commitment.

Who should switch

Already on Ling-3.1-flash: yes, upgrade. Every measurable score moved up, the license didn't tighten, and the base model release gives you an exit path into custom fine-tuning if the ecosystem matures. Meter your token burn for a week afterward.

Choosing a cheap workhorse today: the AA table now argues for GLM-5.3-Flash on economics — same intelligence tier, half the price, half the talk. We ran that comparison in full here (GLM 5.3 Flash vs DeepSeek V4.1 Flash), and GLM's cost-per-scored-task lead is the number that should anchor the decision, not the sticker prices.

Need the base for continued training: Ling-3.1-flash-base in BF16 under MIT is genuinely rare — most "open" releases don't ship a retrainable base. If you're building a domain model and the 560B size fits your budget, this is the interesting artifact of the release.

Bottom line

We wrote two weeks ago that Ling-3.1-flash would become the default cheap-and-open pick only when a checkpoint trimmed token output while holding the score. The 1019 is the opposite of that checkpoint: smarter, and hungrier. It's a good upgrade for the installed base and a misleading bargain for everyone else — the sticker price is a fraction of your real cost, and AA's #22-of-22 verbosity ranking is the number that will show up in your invoice, not in the launch post.

• Ling-3.1-flash at open-weights release — where the verbosity thesis started
• GLM 5.3 Flash vs DeepSeek V4.1 Flash — the cheaper-per-scored-task alternative, priced out
• Qwen3-8-Flash vs DeepSeek 4.1 Flash — the budget-tier comparison

Sources: Hugging Face inclusionAI/Ling-3.1-flash-1019 — model card (MIT, 560B/25B, score deltas, "1M context" claim) and config.json (max_position_embeddings 262144, 72 layers, bailing_hybrid), pulled 2026-10-31; Artificial Analysis model page rendered 2026-10-31 (#4/22, $3.19/task, 240M verbosity, $0.30/$0.90, 80% cache discount); API list pricing and base-model benchmark claims as transcribed from Chinese coverage (智东西/pandaily) — single-sourced, not independently re-verified. GLM-5.3-Flash comparison figures from our October 31 price check and AA's live page the same day.