# Ling-3.0-flash

> Source: https://aiwiki.ai/wiki/ling_3_0_flash
> Updated: 2026-09-05
> Fact-checked: 2026-09-05
> Categories: AI Models, Chinese AI, Large Language Models, Mixture of Experts, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Ling-3.0-flash." aiwiki.ai, 5 Sept 2026. https://aiwiki.ai/wiki/ling_3_0_flash
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Ling-3.0-flash** is an open-weight [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) language model from [inclusionAI](https://aiwiki.ai/wiki/inclusionai), the open-source AI initiative of [Ant Group](https://aiwiki.ai/wiki/ant_group), and the first model of the Ling 3.0 generation. The Ant Ling team announced it on July 23, 2026 with a free API on [OpenRouter](https://aiwiki.ai/wiki/openrouter), Ant Group issued a press release on July 27, and the weights appeared on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the [MIT license](https://aiwiki.ai/wiki/mit_license) on August 2, 2026.[1][2][3][4] The model has 124 billion total parameters and activates about 5.1 billion per token. It was built from the start of pretraining on a hybrid-linear attention stack, alternating five Kimi Delta Attention (KDA) layers with one gated [multi-head latent attention](https://aiwiki.ai/wiki/multi-head_latent_attention) (MLA) layer, a design Ant calls "native" in contrast to earlier Ling models that migrated to hybrid attention, and it routes each token to 8 of 512 experts, an activation ratio of 1/64.[2][5][19] Ant Group describes it as a fast, cheap "execution node" for agent workflows rather than a replacement for its trillion-parameter flagships.[1]

The Ling 3.0 line grew through August and September 2026. Ant released a 7.9-billion-parameter sibling, Ling-3.0-tiny; pretrained, mid-trained, and merged base checkpoints for both sizes; FP8, INT4, FP4, and GGUF quantizations; a speculative-decoding draft model; and three domain variants built on the flash model: Ling-3.0-flash-Fin for finance (weights September 3), Ling-3.0-flash-VL for vision (announced September 4), and Ling-3.0-flash-Sante for health and medicine (September 4).[4][6][7][8][9]

## Release timeline

| Date (2026) | Event | Source |
| --- | --- | --- |
| July 23 | Ant Ling posts "Today, we're releasing Ling-3.0-flash"; OpenRouter lists `inclusionai/ling-3.0-flash:free` | [3][30] |
| July 27 | Business Wire press release from Hangzhou: free API on OpenRouter and Vercel AI Gateway through August 3, weights to follow | [1] |
| August 2 | `inclusionAI/Ling-3.0-flash` repository created on Hugging Face, MIT license | [4] |
| August 4 | `Ling-3.0-flash-fp8`, `-int4`, and `-fp4` quantized checkpoints | [4] |
| August 6 | Ling-3.0-tiny listed on OpenRouter | [10] |
| August 9 | `Ling-3.0-flash-dspark` speculative-decoding draft model | [4][11] |
| August 10 | Ling-3.0-tiny weights, plus `-fp8` and `-int4` variants, on Hugging Face | [4][6] |
| August 11 to 20 | `-base-30T`, `-base-midtrain`, and `-base` checkpoints for tiny and flash (repositories created August 11, last updated August 20) | [4][12] |
| August 27 | Ling-3.0-flash-Fin announced; OpenRouter lists `ling-3.0-flash-fin` and a free tier | [7][13] |
| August 30 to 31 | Official `Ling-3.0-tiny-GGUF` and `Ling-3.0-flash-GGUF` repositories | [4][14] |
| September 3 | `Ling-3.0-flash-Fin` weights on Hugging Face, with the FinFIRST evaluation dataset | [4][15][16] |
| September 4 | Ling-3.0-flash-VL announced (18:04 UTC); Ling-3.0-flash-Sante announced (19:14 UTC) and listed on OpenRouter and Vercel AI Gateway | [8][9][17][18] |

The press release is dated four days after the model was already public. The OpenRouter model page and the Ant Ling post both give July 23 as the release date, so that is the date used in this article; the Business Wire text of July 27 is the first formal company statement.[1][30][3]

## Architecture

Ant Group calls Ling-3.0-flash a "native hybrid-reasoning" model with a "native hybrid-linear attention architecture". Both words carry a specific meaning in Ant's materials. Native hybrid reasoning means one checkpoint serves both a thinking mode, which emits a chain of thought before answering, and a non-thinking mode, switched per request with an `enable_thinking` flag; thinking is on by default.[2][5] Native hybrid-linear means the linear-attention layers were present from the first pretraining step, rather than being retrofitted onto a full-attention model. Ant's developer documentation describes the change as a transition "from migration-based hybrid linear architecture to native hybrid linear design".[19]

### Attention

The model interleaves Kimi Delta Attention with gated multi-head latent attention at a ratio of five to one: 35 KDA layers and 7 gated MLA layers across 42 transformer layers.[2] KDA is a [linear attention](https://aiwiki.ai/wiki/linear_attention) mechanism introduced by [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai) in the October 2025 [Kimi Linear](https://aiwiki.ai/wiki/kimi_linear) paper; it extends the gated delta rule of [Gated DeltaNet](https://aiwiki.ai/wiki/gated_deltanet) with a diagonal, channel-wise forget gate so that each feature channel keeps its own decay rate.[20] Moonshot's own flagship, [Kimi K3](https://aiwiki.ai/wiki/kimi_k3) of July 2026, is likewise built on KDA with gated MLA layers (69 KDA and 24 gated MLA).[28] Ant's press release credits KDA with "fine-grained diagonal gating in Delta Rule state updates" and says it replaced the Lightning Attention used in the previous Ling generation, allowing the model to "retain critical information more precisely" over long documents and codebases.[1]

Because only one layer in six keeps a conventional [KV cache](https://aiwiki.ai/wiki/kv_cache), the memory that grows with context length is much smaller than in a full-attention model of the same depth; the KDA layers carry a fixed-size recurrent state instead.[5][20] The configuration file shows the MLA layers use a KV compression rank of 512, 32 attention heads, and a partial rotary embedding (a 64-dimension rotary component alongside a 128-dimension non-rotary component in each 192-dimension query/key head, base frequency 6,000,000), with QK normalization enabled.[21] Unlike Kimi Linear, which drops positional encoding from its full-attention layers, Ling-3.0-flash keeps [rotary position embedding](https://aiwiki.ai/wiki/rotary_position_embedding) in its MLA layers (`use_mla_nope: false`).[21]

### Mixture of experts

Every non-dense layer holds 512 routed experts plus one shared expert, and the router picks 8 routed experts per token. That is a 1/64 activation ratio, halved from the 1/32 that Ant's press release cites for the previous generation.[1][2] The first two layers are dense. The router uses a sigmoid score function with the aux-loss-free "noaux_tc" top-k method and an expert bias, and the model carries one multi-token-prediction layer that inference engines can use for speculative decoding.[21][5]

The Hugging Face model card gives the following summary. Rows marked with an asterisk are additional values read from the repository's `config.json`.[2][21]

| Item | Ling-3.0-flash |
| --- | --- |
| Architecture | Hybrid-linear MoE (`BailingMoeV3ForCausalLM`, model type `bailing_hybrid`) |
| Parameter scale | Total 124B, activated 5.1B |
| Transformer layers | 35 KDA + 7 gated MLA (5:1), 42 in all |
| Dense layers | 2 |
| Routed experts | 512 |
| Shared experts | 1 |
| Activated experts | 8 routed + 1 shared |
| Attention heads | 32 |
| Hidden size | 2560 |
| Expert intermediate size | 768 |
| Dense intermediate size | 6144 |
| Vocabulary size | 157,184 |
| Context training schedule | 8K, then 32K, then 256K |
| Maximum position embeddings* | 262,144 |
| MLA KV compression rank* | 512 |
| Rotary embedding* | Partial (64 of 128 dimensions), base 6,000,000 |
| Multi-token-prediction layers* | 1 |
| Router* | Sigmoid, aux-loss-free top-k (`noaux_tc`), 8 expert groups |

Hugging Face's safetensors metadata reports about 127.5 billion stored parameters for the repository (BF16 plus a small number of FP32 tensors), slightly above the 124B headline figure.[22]

### Context length

The native context window is 256K tokens (262,144 positions), reached through a three-stage schedule of 8K, 32K, and 256K during training. Ant's press release says the model "can seamlessly scale to 1M tokens", and its developer documentation lists the window as "native 256K, extendable up to 1M"; hosted endpoints on OpenRouter and Vercel AI Gateway expose 262,144 tokens.[1][2][19][23]

### Ling-3.0-tiny

Ling-3.0-tiny is the small member of the generation: 7.9 billion total parameters and 1.3 billion activated per token, aimed at local and edge deployment. It uses the same hybrid design at a 3:1 ratio (three KDA layers followed by one MLA layer in each four-layer block) with 128 routed experts, 8 routed plus 1 shared active per token.[6] The base-model card gives 18 KDA and 6 gated MLA layers, one dense layer, 16 attention heads, a hidden size of 1,536, an expert intermediate size of 512, a dense intermediate size of 4,608, and the same 157,184-token vocabulary as the flash model; the configuration file adds 131,072 maximum positions.[29][24] The SGLang recipe extends that to 262,144 tokens with [YaRN](https://aiwiki.ai/wiki/yarn) scaling, and OpenRouter's listing (created August 6) gives a 262K window, though it had no serving endpoint as of September 5, 2026.[6][10] Ant says the model has been validated on NVIDIA DGX Spark, Apple Silicon MacBooks, and the Mac mini, reaching about 100 to 105 tokens per second on DGX Spark and 86 to 90 tokens per second on an M4 Pro MacBook in FP8, with roughly 8.34 GiB of peak memory at an 8K context.[6]

## Base checkpoints and training

For both sizes Ant published three pre-post-training checkpoints on Hugging Face: a pretrained checkpoint (`-base-30T`), a mid-trained checkpoint (`-base-midtrain`), and a merged checkpoint (`-base`). The cards label the `-base-30T` checkpoints as the pretrained stage and do not spell out what the suffix denotes.[12] The merged checkpoints use Warmup-Stable and Merge (WSM), a schedule from a July 2025 Ant paper that replaces the usual learning-rate decay phase with weighted merging of checkpoints from the stable phase. Ant argues this leaves the base model better suited to continued pretraining and lets researchers explore different decay profiles offline.[12][25] The cards also state that Ling-3.0-tiny-base and Ling-3.0-flash-base "share the same training recipe", so that experiments validated on the tiny model can be scaled to the flash model.[12]

No Ling 3.0 technical report is linked from the model cards or the developer documentation, and an arXiv search for the model name returned nothing on September 5, 2026; the architectural and training details above come from the model cards, the configuration files, the press release, and the developer documentation. Post-training details are limited to Ant's statements that the flash model was trained in "over 10,000 interactive environments" for coding, general, and deep-research agent tasks, with "enhanced self-correction and long-horizon planning mechanisms".[1][2]

## Benchmarks

All results in this section are Ant-reported unless stated otherwise. Ant's model card presents most of its benchmark comparisons as images, which this article does not transcribe; the figures below come from the machine-readable evaluation files Ant added to the repository, from the text of the model card, and from Artificial Analysis.

### Ant-reported results

The `Ling-3.0-flash` repository includes an `.eval_results` directory with five scores that Ant attributes to its own model card.[26]

| Benchmark | Ant-reported value |
| --- | --- |
| [AIME](https://aiwiki.ai/wiki/aime) 2026 (MathArena) | 93.2 |
| HMMT February 2026 (MathArena) | 87.0 |
| [Humanity's Last Exam](https://aiwiki.ai/wiki/humanity_s_last_exam) | 22.7 |
| SWE-bench Multilingual (percent resolved) | 72.4 |
| SWE-bench Pro | 56.6 |

The model card's text says the model "performs strongly" on SWE-Bench Pro, SWE-Bench Multilingual, Tau3-banking-AA, MCP-Atlas, and SkillsBench, and lists the evaluation settings: thinking mode on, sampling at temperature 0.6, top-p 0.95, and top-k 20; SWE-bench runs through the OpenHands harness with a 256K context and 32K output budget; [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1 under the Artificial Analysis protocol with the Terminus 2 harness, a two-hour timeout, and three runs per task; [BrowseComp](https://aiwiki.ai/wiki/browsecomp) in both a single-agent ReAct setup with context summarization at 64K tokens and an internal multi-agent harness.[2] One of the named benchmarks, AntSWEBench, is described as an internal software-engineering benchmark covering Java, JavaScript, and Python.[2]

The July 23 announcement made a comparative claim: "With 1/8 of the total and 1/12 of the active parameters, it matches or beats our 1T flagship model on most benchmarks shown."[3] The model card puts the same point as 124B and 5.1B being about 12.4 percent and 8.1 percent of the previous flagship, Ring-2.6-1T, and says the model "matches or outperforms its predecessor across key benchmarks".[2] The press release goes further, saying the model "matches or surpasses industry-leading models with two to three times its parameter scale" on foundational reasoning, instruction following, and long-context processing.[1] None of these comparisons has an independent counterpart yet.

For Ling-3.0-tiny, Ant's model card reports a score of 25 on the Artificial Analysis Intelligence Index v4.1.1 and 16 on the Artificial Analysis Agentic Index, and says that in Artificial Analysis testing the model produced over 160 output tokens per second with about 18 seconds of end-to-end latency for a 500-token response.[6]

### Base model evaluation

The base-checkpoint cards are the one place where Ant publishes a full comparison table as text rather than as an image. Ant compares the merged `Ling-3.0-flash-base` with five other base models on what it calls "a self-built comprehensive benchmark suite" spanning knowledge, code, math, reasoning, and long context. The table reproduces Ant's figures; the comparison models and their parameter counts are as Ant lists them.[12]

| Domain | Benchmark (metric, shots) | Ling-3.0-flash-base (124B A5.1B) | Ling-2.5-flash-base (104B A7.4B) | NVIDIA-Nemotron-3-Super-base (120B A12B) | Step-3.5-Flash-base (196B A11B) | Hy3-preview-base (295B A21B) | MiMo-V2.5-base (310B A15B) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Knowledge | CCPM (EM, 0-shot) | 90.00 | 78.38 | 80.77 | 83.90 | 88.64 | 88.05 |
| Knowledge | ARC-C (EM, 0-shot) | 96.27 | 95.59 | 96.27 | 92.88 | 94.58 | 95.59 |
| Knowledge | AGIEval (Acc, 0-shot) | 77.50 | 72.39 | 67.08 | 74.27 | 77.61 | 71.84 |
| Knowledge | SimpleQA-Verified (Acc, 5-shot) | 26.90 | 15.90 | 21.20 | 27.80 | 24.20 | 19.10 |
| Knowledge | MMLU-Pro (EM, 5-shot) | 67.74 | 61.36 | 63.89 | 63.65 | 66.44 | 64.90 |
| Knowledge | CEval (EM, 5-shot) | 89.76 | 87.86 | 78.12 | 89.36 | 90.74 | 88.60 |
| Code | HumanEval-Plus (Pass@1, 0-shot) | 81.10 | 80.49 | 78.05 | 78.05 | 79.27 | 76.22 |
| Code | CruxEval (Pass@1, 1-shot) | 78.50 | 76.50 | 72.44 | 71.69 | 82.94 | 70.06 |
| Code | MultiPL-E (Pass@1, 1-shot) | 70.79 | 68.07 | 57.28 | 63.21 | 45.77 | 53.57 |
| Code | [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench) 2408-2505 (Pass@1, 1-shot) | 40.09 | 33.04 | 38.99 | 35.68 | 35.90 | 35.50 |
| Code | BigCodeBench (Pass@1, 0-shot) | 52.19 | 50.35 | 25.53 | 49.74 | 52.02 | 50.18 |
| Code | FullStackBench (Pass@1, 3-shot) | 51.16 | 48.43 | 49.61 | 50.86 | 54.95 | 50.39 |
| Code | LCBench 2301-2502 (Pass@1, 3-shot) | 57.76 | 57.51 | 54.22 | 49.67 | 46.94 | 45.12 |
| Math | MATH500 (Acc, 4-shot) | 79.00 | 74.00 | 70.80 | 62.80 | 60.60 | 70.00 |
| Math | OlympiadBench (Acc, 3-shot) | 47.89 | 39.31 | 54.52 | 34.34 | 41.27 | 35.09 |
| Math | TheoremQA (Acc, 5-shot) | 61.10 | 57.95 | 72.05 | 56.03 | 60.96 | 51.23 |
| Math | OmniMath (Acc, 3-shot) | 47.65 | 36.20 | 38.55 | 28.46 | 33.33 | 27.03 |
| Reasoning | CommonSenseQA (EM, 5-shot) | 89.93 | 87.55 | 86.24 | 88.62 | 86.98 | 84.60 |
| Reasoning | BBH (EM, 3-shot) | 89.17 | 84.72 | 89.00 | 87.19 | 76.00 | 86.22 |
| Long-context | LongBench (Acc, 0-shot) | 52.62 | 42.81 | 31.05 | 21.47 | 20.58 | 29.57 |
| Long-context | LEval (Acc, 0-shot) | 83.24 | 76.89 | 58.65 | 60.30 | 66.29 | 73.29 |

On Ant's numbers the flash base has the top score on 13 of the 21 rows, ties Nemotron-3-Super-base on ARC-C, and trails on seven: Nemotron-3-Super-base leads on OlympiadBench and TheoremQA, Hy3-preview-base on AGIEval, CEval, CruxEval, and FullStackBench, and Step-3.5-Flash-base on SimpleQA-Verified. The largest margins in Ant's favor are on the two long-context tests. Ant reports the same suite for `Ling-3.0-tiny-base` against Ling-2.5-mini-base (16B A1.4B) and the [Qwen3.5](https://aiwiki.ai/wiki/qwen3_5) 9B and 4B base models, where the tiny base is ahead on most code rows, behind Qwen3.5-9B-base on most knowledge, reasoning, and long-context rows, and behind its own predecessor on MATH500 (65.60 versus 68.40) and LCBench (41.59 versus 46.16).[29]

### Independent evaluation

[Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) lists Ling 3.0 Flash as an open-weights reasoning model released in August 2026 with a 262K context. As of September 5, 2026, its page reports a score of 27 on the Artificial Analysis Intelligence Index v4.2 (ten evaluations including GDPval-AA v2, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, and AA-LCR v1.1), ranking first of 63 models in its comparison class of open-weight models between 40B and 150B parameters. The same snapshot gives an output speed of roughly 320 to 330 tokens per second at the time of access (the figure changes between measurements), a cost of $0.03 per Intelligence Index task (first in class) at reference prices of $0.075 per million input tokens and $0.22 per million output tokens, and 240 million output tokens generated across the index, which the site calls "very verbose" against a median of 170 million.[27] Vercel's AI Gateway model page states that the model "scores 38 on the Artificial Analysis Intelligence Index"; the two numbers likely reflect different index versions, and this article reports both with their sources.[23][27]

The HealthBench, BrowseComp, and medical-benchmark chart that Ant published for Ling-3.0-flash-Sante is covered on the [Ling-3.0-flash-Sante](https://aiwiki.ai/wiki/ling_3_0_flash_sante) page.

## Variants

| Model | Date (2026) | Parameters | Availability (as of September 5, 2026) | License |
| --- | --- | --- | --- | --- |
| Ling-3.0-flash | API July 23; weights August 2 | 124B total / 5.1B active | Hugging Face, [ModelScope](https://aiwiki.ai/wiki/modelscope), OpenRouter, Vercel AI Gateway, Ant Ling API | MIT |
| Ling-3.0-flash-fp8, -int4, -fp4 | August 4 | 124B / 5.1B (FP8 blockwise E4M3; INT4 compressed-tensors W4A16; MXFP4) | Hugging Face | MIT |
| Ling-3.0-flash-dspark | August 9 | 1.36B draft model | Hugging Face | "other" |
| Ling-3.0-tiny | Ant Ling API August 7 (per Ant's changelog); OpenRouter listing August 6; weights August 10 | 7.9B / 1.3B | Hugging Face (BF16, FP8, INT4), Ant Ling API; the OpenRouter listing had no active endpoint as of September 5, 2026 | MIT |
| Ling-3.0-flash-base-30T, -base-midtrain, -base | Created August 11, updated August 20 | 124B / 5.1B | Hugging Face | MIT |
| Ling-3.0-tiny-base-30T, -base-midtrain, -base | Created August 11, updated August 20 | 7.9B / 1.3B | Hugging Face | MIT |
| Ling-3.0-tiny-GGUF, Ling-3.0-flash-GGUF | August 30, August 31 | Same as base models | Hugging Face | MIT |
| Ling-3.0-flash-Fin | Announced August 27; weights September 3 | 124B / 5.1B | Hugging Face (BF16), OpenRouter (paid and free), Vercel AI Gateway | MIT |
| Ling-3.0-flash-VL | Launched September 4 on the Ant Ling API and chat (per Ant's changelog) | Not stated | Ant Ling API and chat; no weights on Hugging Face or ModelScope as of September 5, 2026 | Not stated |
| Ling-3.0-flash-Sante | September 4 | 124B / 5.1B | OpenRouter (free), Vercel AI Gateway (free through October 4); not on Hugging Face | Not stated |

Dates for Hugging Face repositories are the `createdAt` and `lastModified` values reported by the Hugging Face API on September 5, 2026.[4] The SGLang cookbook lists the quantization formats.[5] The draft model's card carries a `license: other` tag rather than MIT.[11]

### Quantized and community checkpoints

Ant's own quantizations are FP8 with blockwise E4M3 expert weights, INT4 in the compressed-tensors W4A16 format, and an MXFP4 checkpoint that mixes MXFP4 routed experts with block-FP8 dense and shared projections.[5] Official [GGUF](https://aiwiki.ai/wiki/gguf) conversions for [llama.cpp](https://aiwiki.ai/wiki/llama_cpp) followed at the end of August.[14] Community conversions appeared within days of the weights: GGUF builds from bartowski, AtomicChat, bloomer010, and others, [MLX](https://aiwiki.ai/wiki/mlx) builds for Apple Silicon, abliterated variants, and expert-pruned "REAP" versions that cut the model to 97B, 73B, or 46B total parameters at the same 5B active count.[4] As of September 5, 2026, Hugging Face counted more than 266,000 downloads for the AtomicChat GGUF alone, more than the official BF16 repository.[4]

### Ling-3.0-flash-dspark

DSpark is a speculative-decoding draft model for the flash checkpoint, released on August 9. Ant describes it as an extension of DFlash with target-model auxiliary features and a confidence head that decides how many draft tokens to propose. The draft has 1.36 billion parameters, five full-attention layers, a hidden size of 2,560, and a block size of 8 draft tokens (a verify width of 9 including the target's bonus token). It was trained with SpecForge and is served with SGLang or llama.cpp.[11]

| Workload | Acceptance length (Ant-reported) |
| --- | --- |
| GSM8K | 6.40 |
| MATH-500 | 6.29 |
| AIME 2025 | 5.56 |
| HumanEval | 6.57 |
| MBPP | 6.34 |
| [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench) | 5.33 |
| MT-Bench | 3.92 |
| Alpaca | 3.51 |
| Arena-Hard-v2 | 3.72 |

Acceptance length is the mean number of tokens accepted per verification step; the macro mean across the nine workloads is 5.29.[11]

## Domain variants

### Ling-3.0-flash-Fin

Ling-3.0-flash-Fin is a finance-tuned version of the flash model, announced by Ant Ling on August 27, 2026 and listed on OpenRouter the same day, with weights published on Hugging Face on September 3.[7][13][4] The model card calls it "the first finance-enhanced model in the Ant Ling family", developed "with leading financial institutions and domain experts" through continued training on financial data, and keeps the 124B/5.1B architecture and 256K context of the base model.[15] Ant's stated targets are end-to-end financial research (retrieval, evidence review, calculation, modeling, and report preparation), source-grounded search that prioritizes authoritative sources, multi-document reasoning across annual reports and filings, and spreadsheet and valuation workflows.[15] TechNode reported the launch on August 28, noting a one-month free API period through OpenRouter for finance professionals and developers.[16]

Ant says it evaluated the model on FinFIRST, FinSearchComp Verified, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench, and tau3-Banking, and describes it as "competitive with both similarly sized models and substantially larger general-purpose models"; the results themselves are published as an image.[15][7] FinFIRST, the source-grounded retrieval benchmark, was released alongside the model as a Hugging Face dataset under Apache 2.0.[16] The card recommends temperature 1.0 (rather than the base model's 0.6) and states that valuation results and investment conclusions "require professional review and do not constitute investment advice".[15] OpenRouter listed the paid endpoint at $0.06 per million input tokens and $0.18 per million output tokens on September 5, 2026, alongside a free tier.[13]

### Ling-3.0-flash-VL

On September 4, 2026 at 18:04 UTC, Ant Ling announced Ling-3.0-flash-VL, "built on Ling-3.0-flash with visual understanding and visual agent capabilities", and said it "performs well across visual perception, STEM reasoning, document intelligence, multimodal agent tasks, frontend coding, and medical report interpretation".[8] The post included a benchmark image but no link. As of September 5, 2026, no repository named Ling-3.0-flash-VL exists on Hugging Face or ModelScope, and OpenRouter does not list the model; Ant's developer changelog of September 4, 2026 records the model as launched on Ant's own chat interface and on an OpenAI- and Anthropic-compatible API, so this article treats it as a hosted release without open weights.[4]

### Ling-3.0-flash-Sante

Ling-3.0-flash-Sante, announced about an hour later on September 4, is a health and medicine variant of the flash model. Ant Ling described it as built for "medical reasoning, professional healthcare tasks, deep research, and evidence-based retrieval", and said it showed "leading performance among open-source models" on MedXpertQA-Text, DiagnosisArena-MCQ, AFUMED-Drug, HealthBench Professional, and BrowseComp.[9] OpenRouter lists it as a free model with a 262,144-token context and 32,768-token maximum completion, and Vercel's AI Gateway offers it free through October 4, 2026, after which the standard model id begins billing.[17][18] The weights were not on Hugging Face as of September 5, 2026.[4] The Ling-3.0-flash-Sante page covers the model, its reported benchmark chart, and the chart's caveats.

## Serving and deployment

Ant built the flash model's serving story around [SGLang](https://aiwiki.ai/wiki/sglang). The model card links a cookbook with a launch-command generator for BF16, FP8, INT4, and MXFP4 checkpoints on Hopper and Blackwell GPUs, and a prebuilt Docker image, `lmsysorg/sglang:dev-Ling-3.0-flash`.[2][5] The recommended low-latency recipe runs BF16 with tensor parallelism of 4 on 141 GB-class cards (H20-3e, H200) or four-GPU Blackwell nodes, and 8 on 80 GB H100 or H800 cards; FP8 recipes pair tensor parallelism with expert parallelism because the blockwise E4M3 expert weights only shard cleanly at TP2.[5] Two speculative-decoding paths are offered: the built-in multi-token-prediction layer (`--speculative-algorithm NEXTN`) or the external DSpark draft, which requires a "ReplaySSM" ring for the KDA verify step.[5] Ling-specific reasoning and tool-call parsers (`ling3`) return the chain of thought in a separate `reasoning_content` field.[5]

The third SGLang recipe, and the one Ant highlights in its own materials, is HiCache with [Mooncake](https://aiwiki.ai/wiki/mooncake). The model card says the model "natively integrates the SGLang HiCache + Mooncake hierarchical caching architecture", with physical dual pools and a cluster-shared L3 cache, to avoid recomputing prefixes during long-horizon agent sessions; Ant reports a 60 percent to over 80 percent reduction in time to first token on long inputs from this setup.[2][1] That figure is Ant's own measurement.

For [vLLM](https://aiwiki.ai/wiki/vllm), the model card gives a `vllm serve` recipe with `--mamba-cache-mode align`, the `ling3` parsers, and a multi-token-prediction speculative configuration with three speculative tokens; at the time of the card it required building vLLM from source.[2] Ling-3.0-tiny can also run through [Ollama](https://aiwiki.ai/wiki/ollama) by building from a pull request that adds the `bailing_moe_v3` architecture; Ant notes the support is limited to MLX on Apple Silicon and was not yet in an official Ollama release.[6]

Ant recommends temperature 0.6, top-p 0.95, and top-k 20 for the flash model and temperature 1.0 for the tiny and Fin models, with thinking enabled.[2][6][15] The model card names [Claude Code](https://aiwiki.ai/wiki/claude_code), Kilo Code, Qwen Code, [Hermes Agent](https://aiwiki.ai/wiki/hermes_agent), and [OpenClaw](https://aiwiki.ai/wiki/openclaw) as agent frameworks in which the model "delivers a strong user experience", and Ant's developer site publishes integration guides for four of them.[2][19]

### Hosted access and pricing

OpenRouter created the `ling-3.0-flash` listing on July 23, 2026. As of September 5, 2026 the `:free` variant had no serving endpoint; the paid `inclusionai/ling-3.0-flash` endpoint (served by Novita) remained live with a 262,144-token context, and Vercel AI Gateway still offered a free routing id, `ling-3.0-flash-free`.[30] [Vercel](https://aiwiki.ai/wiki/vercel)'s AI Gateway lists the paid model at $0.06 per million input tokens and $0.18 per million output tokens, with cached reads at $0.01 per million, and a free routing id, `ling-3.0-flash-free`, that serves the same model; Vercel's page gives a 256K context and up to 32K tokens per response.[23] Ant's own API, documented at developer.ant-ling.com, offers OpenAI- and Anthropic-compatible endpoints and claims a peak of 1,000 tokens per second with a time to first token under 100 milliseconds; those are the company's figures.[19]

## Position in the Ling family

Ant's developer documentation lays out the Ling lineage as five generations: Ling 1.0 (March 2025), which validated MoE training on non-flagship accelerators; Ling 2.0 (October 2025), the generation of the trillion-parameter [Ling-1T](https://aiwiki.ai/wiki/ling_1t) and [Ring-1T](https://aiwiki.ai/wiki/ring_1t), with end-to-end FP8 training; Ling 2.5 (February 2026), with high-throughput decoding and long-context work; Ling 2.6 (April 2026), which the company says completed the chain "from logical reasoning to task execution"; and Ling 3.0 (July 2026), which "transitioned from migration-based hybrid linear architecture to native hybrid linear design" with lower expert activation ratios and KDA.[19] The immediate predecessors are Ling-2.6-1T, a one-trillion-parameter model with 63 billion activated parameters and a 1M context that pairs MLA with linear attention, and its reasoning sibling Ring-2.6-1T, which the Ling-3.0-flash card uses as its comparison point.[19][2]

Ling-3.0-flash is not a flagship in the sense of Ling-1T or Ling-2.6-1T. The press release frames it as one half of a "planning-execution separation" in which large models handle deep planning and Ling-3.0-flash handles high-frequency execution "as a cost-controllable, fast, and highly stable execution node".[1] Ant's model selection guide accordingly recommends Ling-3.0-flash for most general workloads and latency-sensitive services, Ling-2.6-1T for ultra-long documents and complex multi-hop agent chains, and Ling-3.0-tiny for on-device and offline use.[19] The lineage also shows a move toward sparser activation: Ant's press release puts the previous generation's expert activation ratio at 1/32 and Ling 3.0's at 1/64.[1][2]

## Reception

Coverage of the flash launch outside Ant's own channels was thin. The July 23 post and the August 2 weights each drew one Hacker News submission with a few points and no comments, and a third submission covered a red-teaming test of Ling-3.0-tiny; TechNode covered the Fin variant on August 28; and the Sante and VL announcements on September 4 surfaced first through OpenRouter and Vercel listings.[31][16][17][18] The most substantive third-party signal is the Artificial Analysis ranking described above, which as of September 5, 2026 placed the model first for intelligence and cost per task among open-weight models in its size class while flagging its verbosity.[27] Adoption is visible mainly through the inference ecosystem: SGLang shipped a dedicated image and cookbook, vLLM carries `ling3` reasoning and tool-call parsers, and community quantizers produced GGUF, MLX, and pruned builds within the first week of the weight release.[5][2][4]

## References

1. [Ant Group Unveils Ling-3.0-Flash Delivering Top-Tier Performance at a Fraction of the Parameter Scale](https://www.businesswire.com/news/home/20260726584441/en/) - Business Wire (Ant Group), July 27, 2026. Syndicated copy: [Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/ant-group-unveils-ling-3-060000298.html).
2. [inclusionAI/Ling-3.0-flash model card](https://huggingface.co/inclusionAI/Ling-3.0-flash) - Hugging Face (inclusionAI), accessed September 5, 2026.
3. [Ant Ling (@AntLingAGI) post announcing Ling-3.0-flash](https://x.com/AntLingAGI/status/2080351022028095681) - X, July 23, 2026.
4. [Hugging Face Hub API, models by author inclusionAI](https://huggingface.co/api/models?author=inclusionAI&search=Ling-3.0&sort=createdAt) - Hugging Face, accessed September 5, 2026 (repository creation dates, last-modified dates, licenses, download counts).
5. [Ling-3.0-flash cookbook](https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-flash) - SGLang Documentation, accessed September 5, 2026.
6. [inclusionAI/Ling-3.0-tiny model card](https://huggingface.co/inclusionAI/Ling-3.0-tiny) - Hugging Face (inclusionAI), accessed September 5, 2026.
7. [Ant Ling (@AntLingAGI) post introducing Ling-3.0-flash-Fin](https://x.com/AntLingAGI/status/2093022087069958492) - X, August 27, 2026.
8. [Ant Ling (@AntLingAGI) post announcing Ling-3.0-flash-VL](https://x.com/AntLingAGI/status/2095935971556782372) - X, September 4, 2026.
9. [Ant Ling (@AntLingAGI) post introducing Ling-3.0-flash-Sante](https://x.com/AntLingAGI/status/2095953758148853892) - X, September 4, 2026.
10. [inclusionAI: Ling 3.0 Tiny (free)](https://openrouter.ai/inclusionai/ling-3.0-tiny:free) - OpenRouter, accessed September 5, 2026.
11. [inclusionAI/Ling-3.0-flash-dspark model card](https://huggingface.co/inclusionAI/Ling-3.0-flash-dspark) - Hugging Face (inclusionAI), accessed September 5, 2026.
12. [inclusionAI/Ling-3.0-flash-base model card](https://huggingface.co/inclusionAI/Ling-3.0-flash-base) - Hugging Face (inclusionAI), accessed September 5, 2026.
13. [inclusionAI: Ling 3.0 Flash Fin (free)](https://openrouter.ai/inclusionai/ling-3.0-flash-fin:free) - OpenRouter, accessed September 5, 2026; paid-tier pricing from the [OpenRouter models API](https://openrouter.ai/api/v1/models), same date.
14. [inclusionAI/Ling-3.0-flash-GGUF model card](https://huggingface.co/inclusionAI/Ling-3.0-flash-GGUF) - Hugging Face (inclusionAI), accessed September 5, 2026.
15. [inclusionAI/Ling-3.0-flash-Fin model card](https://huggingface.co/inclusionAI/Ling-3.0-flash-Fin) - Hugging Face (inclusionAI), accessed September 5, 2026.
16. [Ant Group launches finance-tuned Ling model, plans to open-source it next week](https://technode.com/2026/08/28/ant-group-launches-finance-tuned-ling-model-plans-to-open-source-it-next-week/) - TechNode (TechNode Feed), August 28, 2026. FinFIRST dataset: [inclusionAI/FinFIRST](https://huggingface.co/datasets/inclusionAI/FinFIRST) - Hugging Face, last modified September 3, 2026.
17. [inclusionAI: Ling 3.0 Flash Sante (free)](https://openrouter.ai/inclusionai/ling-3.0-flash-sante:free) - OpenRouter, accessed September 5, 2026.
18. [Ling 3.0 Flash Sante is now available on AI Gateway for free](https://vercel.com/changelog/ling-3-0-flash-sante-is-now-available-on-ai-gateway-for-free) - Vercel Changelog, September 4, 2026.
19. [Models: Ling](https://developer.ant-ling.com/en/docs/models/ling) - Ant Ling developer documentation (Ant Intelligence (Hang Zhou) Technology Co., Ltd.), last updated September 4, 2026.
20. [Kimi Linear: An Expressive, Efficient Attention Architecture](https://arxiv.org/abs/2510.26692) - arXiv (Kimi Team, Moonshot AI), October 30, 2025.
21. [inclusionAI/Ling-3.0-flash config.json](https://huggingface.co/inclusionAI/Ling-3.0-flash/blob/main/config.json) - Hugging Face (inclusionAI), accessed September 5, 2026.
22. [Hugging Face Hub API, inclusionAI/Ling-3.0-flash](https://huggingface.co/api/models/inclusionAI/Ling-3.0-flash) - Hugging Face, accessed September 5, 2026 (safetensors parameter metadata).
23. [Ling 3.0 Flash API, Pricing and Playground](https://vercel.com/ai-gateway/models/ling-3.0-flash) - Vercel AI Gateway, accessed September 5, 2026.
24. [inclusionAI/Ling-3.0-tiny config.json](https://huggingface.co/inclusionAI/Ling-3.0-tiny/blob/main/config.json) - Hugging Face (inclusionAI), accessed September 5, 2026.
25. [WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-training](https://arxiv.org/abs/2507.17634) - arXiv (Changxin Tian, Jiapeng Wang, Qian Zhao, Kunlong Chen, Jia Liu, Ziqi Liu, Jiaxin Mao, Wayne Xin Zhao, Zhiqiang Zhang, Jun Zhou), July 23, 2025.
26. [inclusionAI/Ling-3.0-flash, .eval_results directory](https://huggingface.co/inclusionAI/Ling-3.0-flash/tree/main/.eval_results) - Hugging Face (inclusionAI), accessed September 5, 2026 (aime_2026.yaml, hmmt_feb_2026.yaml, hle.yaml, swe-bench_multilingual.yaml, swe-bench_pro.yaml).
27. [Ling 3.0 Flash: Intelligence, Performance and Price Analysis](https://artificialanalysis.ai/models/ling-3-0-flash) - Artificial Analysis, accessed September 5, 2026.
28. [moonshotai/Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3) - Hugging Face (Moonshot AI), accessed September 5, 2026.
29. [inclusionAI/Ling-3.0-tiny-base model card](https://huggingface.co/inclusionAI/Ling-3.0-tiny-base) - Hugging Face (inclusionAI), accessed September 5, 2026.
30. [inclusionAI: Ling 3.0 Flash (free)](https://openrouter.ai/inclusionai/ling-3.0-flash:free) - OpenRouter, accessed September 5, 2026.
31. [Hacker News search results for "Ling-3.0"](https://hn.algolia.com/api/v1/search?query=Ling-3.0&tags=story) - Hacker News (Algolia search API), accessed September 5, 2026.