Ant Ling 3.0 Flash Review 2026: The 124B Open-Source Agent That Out-Scores Trillion-Parameter Flagships
Verdict first: Ant's Ling-3.0-Flash is the open-source value dark horse of August 2026 — a 124B-parameter model that activates only 5.1B tokens per query yet leads open models on SWE-bench Pro (56.6%), ties Claude-Sonnet-4.6 on function calling (BFCL-v4 73.0%), and lands a 38 on the Artificial Analysis Intelligence Index. It is purely MIT-licensed, cheap enough to be the least expensive measured model at its intelligence tier, and small enough to run on a single DGX Spark. The catch: the 2.5x price discount ends August 31, and some benchmark figures are still awaiting independent reproduction.
Ant Ling-3.0-Flash — Ant Group's open-source agent flagship (124B total / 5.1B active, MIT license)
What Is Ling-3.0-Flash?
Ling-3.0-flash is the next-generation native hybrid-reasoning model from Ant Group's inclusionAI (Bailing/百灵), announced July 23–24, 2026 and fully open-sourced on August 7, 2026 on Hugging Face and ModelScope.
Ant Group is the parent company of Alipay — one of the largest fintech companies in the world — and the Ling family is its core model line for general-purpose MoE + tool calling + agents (alongside Ring for deep reasoning and Ming for multimodal).
Ling-3.0-Flash is built for production-grade agents: with just 124B total parameters and a tiny 5.1B active per token, it claims to match trillion-parameter flagships. The media framed it as "Ant's 124B execution model" — released the same day Nvidia's Jensen Huang was championing open source.
Key Specs at a Glance
| Spec | Ling-3.0-Flash |
|---|---|
| Release / Open-sourced | July 23-24, 2026 · Open-sourced Aug 7, 2026 (MIT) |
| Architecture | Sparse MoE — 124B total, 5.1B active (~4%), 512 routed + 1 shared expert, 8 active per token |
| Attention | 35× Kimi Delta Attention + 7× gated MLA, 5:1 alternating (hybrid attention) |
| Context window | Native 256K/262K, extendable to 1M |
| License | Pure MIT — no acceptable-use or revenue-threshold restrictions |
| Quantized versions | base / FP8 / FP4 / INT4 |
| Agent training | 10,000+ interactive training environments (coding / general / deep-research) |
| API | OpenAI-compatible at https://api.ant-ling.com/v1 |
Benchmarks: Small Model, Flagship Scores
Ling-3.0-Flash's headline story is small parameters punching far above their weight:
| Benchmark | Score | Context / Rival |
|---|---|---|
| AA Intelligence Index | 38 | +24 over prior-gen Ling 2.6 Flash (14); ties MiMo-V2.5 & Qwen3.6 27B |
| SWE-bench Pro | 56.6% | Open-source leader; beats trillion-param Ring-2.6-1T |
| ArtifactsBench | 77.0% | 10+ points ahead of 2nd place |
| BFCL-v4 (function calling) | 73.0% | Tied #1 with Claude-Sonnet-4.6 |
| MiniAppBench | 25.3% | Mini-app / artifact generation |
| GDPVal v2-AA (professions Elo) | 1108 | Prior-gen Ling 2.6 Flash: 545 |
| LIFEBench | 77.3% | Ranks first |
| τ3-Bench Banking | 27–28% | 2nd among flash-tier open models (behind DeepSeek V4 Flash) |
| Hallucination | 97% → 44% on wrong answers | Now abstains more instead of fabricating |
Speed: ~353 tok/s on Artificial Analysis; >1100 tok/s in high-performance deployments; time-to-first-token drops 60–80% on long inputs (SGLang HiCache + Mooncake tiered caching).
The catch: the AA Intelligence Index is independently measured, but SWE-bench Pro and some other figures are vendor-reported and early independent reproduction is still in progress. Treat "56.6%" as an upper bound for now.
Pricing & the Aug 31 Discount Window
Ling-3.0-Flash is the least expensive model measured at Intelligence Index ≥ 38 on a per-token basis:
| Item | Price |
|---|---|
| Input (first-party API, AA) | $0.075 / M tokens (~¥0.5) |
| Output (first-party API, AA) | $0.22 / M tokens (~¥1.6) |
| Cache hit | 80% off |
| Ant platform list price | ¥0.40 input / ¥1.20 output per M |
| AA per-task cost | ~$0.04 (~¥0.27) |
| OpenRouter | Has a ling-3.0-flash:free tier |
⚠️ Promotional window: Ling Studio is running a 2.5x discount (25% of list price) from August 7 00:00 to August 31 24:00, 2026, reverting to list price on September 1. OpenRouter ran a one-week free window that ended August 3.
Bottom line: now through end of August is the golden window to test it. Budget production workloads against the post-Sept-1 list price.
Open Source & Three Deployment Paths
The pure MIT license is the differentiator vs. many "open-weights-but-restricted" models — no acceptable-use clauses, no revenue-threshold triggers. You can download, modify, commercialize, and redistribute freely, with base / FP8 / FP4 / INT4 quantizations included.
| Path | Use case | Notes |
|---|---|---|
| Cloud API | Rapid validation / shipping agent apps | OpenAI-compatible, lowest barrier |
| Single-box private | Data cannot leave the domain | MXFP4 / INT4 runs end-to-end on one NVIDIA DGX Spark |
| High-performance | Latency-sensitive services | >1100 tokens/s on specified GPU configs |
What We Liked
- ✅ Extreme intelligence density — 124B total / 5.1B active punches to trillion-parameter level
- ✅ Open-source leader on SWE-bench Pro (56.6%) + #1 tie on BFCL-v4 (73.0%) — great for agents & tool calling
- ✅ Pure MIT license — no restrictions, commercial-friendly, quantized versions included
- ✅ Ultra-low price — cheapest measured model at Index ≥38; 2.5x discount through Aug 31
- ✅ Single-box private deployment — one DGX Spark, data never leaves the building
- ✅ Speed — >1100 tok/s high-performance; much faster TTFT on long inputs
- ✅ Ant fintech pedigree — production-grade reliability and a compliance path
What We Didn't Like
- ❌ Vendor-reported benchmarks — SWE-bench Pro 56.6% still awaiting independent reproduction
- ❌ Discount expires Aug 31 — cost resets to list price in September
- ❌ Prior-gen credibility — Ling-2.6-flash's claimed SWE-bench Verified 61.2% lacked independent confirmation; trust must be rebuilt
- ❌ 5.1B active is light for deep reasoning — complex long-horizon tasks still favor Ring-2.6-1T or closed flagships
- ❌ Young ecosystem — fewer third-party tools & community assets than Qwen / GLM / DeepSeek
Ling-3.0-Flash vs DeepSeek V4 Flash vs Qwen vs GLM
| Dimension | Ling-3.0-Flash | DeepSeek V4 Flash | Qwen3.6 27B | GLM-5 |
|---|---|---|---|---|
| AA Intelligence Index | 38 | 52 | 38 | — |
| SWE-bench Pro | 56.6% (leader) | — | 50.2% | — |
| Total params | 124B (5.1B active) | Large MoE | 27B | 743B |
| Open source | ✅ MIT | ✅ | ✅ | ✅ |
| Input / M tokens | $0.075 | ¥1 | — | ¥4 |
| Single-box deploy | ✅ DGX Spark | — | ✅ | ❌ (too large) |
Bottom line: choose DeepSeek V4 Flash for the absolute highest agentic ceiling; choose Ling-3.0-Flash for the balanced value pick — cheap, MIT, capable, and deployable on one machine.
Who Should Use Ling-3.0-Flash?
- 🎯 Developers building agents, coding and function-calling apps — BFCL-v4 #1 tie, SWE-bench Pro open leader
- 🎯 Enterprises with data-sovereignty requirements — DGX Spark single-box private deployment
- 🎯 Budget-sensitive teams — cheapest measured model at its intelligence tier, plus 2.5x discount now
- 🎯 Engineers studying hybrid-attention architectures — KDA + MLA 5:1 stack is genuinely novel
Skip it if you need the strongest deep multi-step reasoning (use Ring-2.6-1T or a closed flagship), you need a mature plugin ecosystem today, or your production budget relies on the 2.5x discount that ends August 31.
Final Verdict: 8.5 / 10
| Category | Score |
|---|---|
| Intelligence / coding | 8.5 / 10 |
| Value for money | 9.0 / 10 (higher with the 2.5x discount) |
| Open source / deployment | 9.5 / 10 (pure MIT + 3 paths) |
| Speed | 8.5 / 10 |
| Ecosystem maturity | 6.5 / 10 |
| Reliability / transparency | 7.0 / 10 (partly self-reported) |
Ling-3.0-Flash is the open-source value surprise of August 2026. A 124B model with 5.1B active parameters delivers the best SWE-bench Pro score among open models, ties a Claude flagship on function calling, ships under a pure MIT license, runs on a single DGX Spark, and undercuts everything else at its intelligence tier. Grab it in the 2.5x-discount window before August 31 — and budget September at list price.
Frequently Asked Questions
Is Ling 3.0 Flash from Ant Group (Alipay)?
Yes. The Ling family comes from Ant Group's inclusionAI (Bailing/百灵) team — Ant is the parent company of Alipay. Ling is its core agent-model line, alongside Ring (deep reasoning) and Ming (multimodal).
Is Ling 3.0 Flash free?
The weights are free (pure MIT — download, modify, commercialize freely). The API is paid: $0.075 input / $0.22 output per million tokens, with an 80% cache-hit discount and a 2.5x promo through August 31. OpenRouter has had a free tier.
Is Ling 3.0 Flash better than Qwen or DeepSeek?
It leads open models on SWE-bench Pro and ties #1 on BFCL-v4 — excellent for agents and tool calling, at a lower price than DeepSeek V4 Flash. But its AA Intelligence Index (38) trails DeepSeek V4 Flash (52), and it's not the model for the hardest deep-reasoning tasks.
Can I run Ling 3.0 Flash locally?
Yes — that's a headline feature. FP4/INT4 versions run end-to-end on a single NVIDIA DGX Spark; bigger GPU configs hit 1100+ tokens/s. Weights are on Hugging Face and ModelScope.
When does the Ling 3.0 Flash discount end?
Ling Studio's 2.5x discount (25% of list) runs August 7–31, 2026. Prices return to list on September 1, 2026.
Sources
- ITHome — Ant open-sources Ling-3.0-flash: 124B params, 5.1B active
- Zhitong Finance — Ling-3.0-flash officially open-sourced
- Alibaba Cloud Dev — native hybrid reasoning, 124B vs 1T flagships
- TechFlow — Ling 3.0 Flash hits 38 on Artificial Analysis
- 80aj — 124B MoE breaks 1100 tokens/s
- OrcaRouter — what we know (and don't) about Ling-3.0-Flash
- ITBear — multi-version release & limited-time discount
- Alibaba Cloud Dev — Ling-2.5-1T trillion-parameter open source
Related Reviews
- DeepSeek V4 Pro Review 2026 — The Agent Flagship That Got 5x Smarter
- Qwen3.8-Max Review 2026 — Alibaba's 2.4T Open-Source Frontier
- Zhipu GLM-5 Review 2026
Disclaimer: Some benchmark figures are vendor-reported as of August 16, 2026 and subject to change. The Ling Studio 2.5x discount ends August 31, 2026. This article is for informational purposes; verify current pricing and capabilities before making purchasing decisions.
没有评论:
发表评论