Ling-3.1-flash-1019 Is Open Source Now: 574 GB, MIT, and a 256K Reality Check

At 05:46 UTC on 27 October, the Ling team at Ant Group created a Hugging Face repository called Ling-3.1-flash-1019 and filled it with 54 safetensors shards totalling 573.8 GB under an MIT licence. Beside it sits Ling-3.1-flash-base, a 1.14 TB BF16 checkpoint intended for further training, and behind both sits Ling-3.1-flash, the model this build is a post-training revision of. Five days earlier Ant had published only the score: 44 on the Artificial Analysis Intelligence Index v4.3.2, up from 41, with weights and API promised. Today the promise has a file listing, which makes this the first day anyone outside Ant's evaluation pipeline can check the number instead of reading it.

That gap is the story, and it cuts both ways. A lab that announces a benchmark before it ships an artifact is asking to be taken on trust; a lab that then ships the artifact five days later, at 574 GB and under a permissive licence, has at least handed over the means to disagree. The interesting questions are no longer about the score. They are about what is inside the download, what it costs to serve, and whether the 1M-token context on the model card is the context you actually get.

What is actually in the repository

The 1019 repository is not a thin adapter or a delta. It is a full checkpoint: 54 weight files at roughly 10.7 GB each, a tokenizer, a chat template, and custom modelling code (BailingMoeV3ForCausalLM) that requires trust_remote_code. The shipped config declares an FP8 quantisation scheme (E4M3, 128×128 block scaling) on top of the hybrid-linear MoE architecture the Ling line has used since 3.1:

• Total and active parameters — 560B total, 25B active per token, the same figures as Ling-3.1-flash.

• Layers — 72 in total: 63 KDA linear-attention layers and 9 gated MLA layers in a 7:1 ratio, plus 4 dense layers.

• Experts — 512 routed, 8 activated per token, 1 shared.

• Vocabulary — 157,184 tokens; hidden size 4,096.

• Footprint — 573.8 GB of safetensors across 54 shards, by our count from the repository file listing on 27 October.

The base model is a different kind of artifact. Ling-3.1-flash-base was created on 25 October and holds 121 files and 1.14 TB of BF16 weights — roughly twice the size of the aligned checkpoint because it has not been quantised to FP8. Its card describes it as the base checkpoint of Ling-3.1-flash, for continued pre-training, domain adaptation and post-training research, and it says plainly that it is not intended for direct chat or safety-critical use without alignment. Both repos carry the MIT tag.

44, and the part Ant did not measure

Artificial Analysis lists Ling-3.1-flash-1019 at 44 on its Intelligence Index, ranked fourth of 122 models on the page we read today, and its model page states that the evaluations were measured independently by Artificial Analysis. That is the strongest form the number has taken so far, and it is worth separating from the rest of the launch table.

The per-benchmark deltas come from Ant's own model card, not from an independent re-run: AA-Briefcase rising from roughly 1,400 to 1,524 Elo, AutomationBench-AA from 62% to 68%, CritPt from 18% to 26%, and SciCode from 54% to 57%. Two of those benchmarks — AA-Briefcase and AutomationBench-AA — are components of the Artificial Analysis index, so the index number and the vendor table are not independent evidence of each other. They are the same claim at two resolutions. Treat the 44 as measured by AA and the before/after pairs as vendor-reported until someone publishes a harness and a seed.

There is a second asymmetry. At the time of measurement, Artificial Analysis recorded no publicly available API for the 1019 build, and its output-speed field for the model is still blank. So the benchmark is repeatable today only by downloading 574 GB and running it yourself. That is not a complaint about the release; it is a description of what "open" changes. Before today, nobody could reproduce the score. After today, anybody with the hardware can.

The three points are bought with tokens

The index is a quality measure, not an efficiency one, and the efficiency side of this release is the least flattering part of it. Artificial Analysis measured 350M output tokens for the 1019 run against 220M for Ling-3.1-flash — about 59% more generated tokens to complete the same suite — and ranks the new build 48th of 48 on verbosity, with the median model at 140M. Its cost to run the index is $1.49 per task against $1.01 for the predecessor.

Some of that is arithmetic you can do from Ant's own rate card. At $0.90 per million output tokens, 350M tokens is $315 of output against $198 for 220M — a $117 difference before a single input token or cache read is counted. For a team that budgets per task rather than per token, the "free" checkpoint upgrade is closer to a 47% increase in the bill for a three-point index gain. Whether that trade is good depends entirely on what a failed task costs you, which is a question no leaderboard answers.

The context window is 256K in the config

The model card says 1M-token context. The shipped configuration says something else. In all three repositories — 1019, Ling-3.1-flash and the base — config.json sets max_position_embeddings to 262,144 and rope_scaling to null, with a RoPE theta of 6,000,000. A stock SGLang or vLLM load of these weights therefore serves 256K positions, not a million, unless the operator supplies a scaling configuration that the repository does not contain.

This is our reading of the published config files, not a statement Ant has made. It is possible the hosted Ling API applies a scaling scheme server-side and genuinely serves 1M; the open weights do not ship one. It is also not a new discrepancy — the 15 October Ling-3.1-flash release carried the same 1M claim over the same 262,144 setting. Anyone sizing hardware around a million-token workload should test the ceiling on their own stack before committing, because the number that matters is the one the server accepts, not the one in the introduction.

Serving it: eight H200s, four GB300s, and one copy-paste trap

Ant recommends SGLang and publishes a day-0 image, lmsysorg/sglang:dev-ling-3.1-flash, with MTP speculative decoding and Hopper decode optimisations. The documented reference configuration is eight H200 GPUs at TP8 plus EP8; on four GB300s the card says to use --tp 4 --ep 4. A vLLM recipe is included as well, with tensor parallelism of 8, expert parallelism, prefix caching and three speculative tokens. Two environment flags — SGLANG_ENABLE_FP8_BLOCK_GEMV and SGLANG_ENABLE_FUSED_KDA_OUTPUT_QUANT — are claimed to give roughly 10% faster batch-1 decode on Hopper.

The footprint explains the hardware list. At 573.8 GB of FP8 weights, the checkpoint leaves about 550 GB of headroom on eight H200s (1,128 GB of HBM) and about 578 GB on four GB300s (1,152 GB). That is enough for weights and a working KV cache, not enough to be careless with long contexts.

There is also a trap in the Quickstart that is worth flagging because it will cost someone a day. The 1019 model card's own SGLang command passes --model-path inclusionAI/Ling-3.1-flash, and its client example sends "model": "Ling-3.1-flash". Followed verbatim, the recipe loads the 41-point predecessor, not the 44-point build. Point the path at inclusionAI/Ling-3.1-flash-1019 explicitly and pin the served model name, or your first benchmark of "1019" will be a benchmark of the old checkpoint.

What Ling-3.1-flash-base is for — and what it is not

The base release is the more consequential of the two for anyone who trains models rather than calls them. A 1.14 TB BF16 checkpoint under MIT, with the same 560B/25B architecture, gives a lab a legitimate starting point for continued pre-training, domain adaptation, supervised fine-tuning, preference optimisation, distillation experiments and long-context MoE research without beginning from a chat-aligned model whose refusal behaviour it would have to undo.

What it does not give you is the recipe for 44. The card describes the base as the base checkpoint of Ling-3.1-flash, not a pre-post-training snapshot of the 1019 build, and no post-training data, preference set or RL configuration ships with either repository. If you want the measured model, download the measured model. If you want to build your own, the base saves you the pre-training bill but not the alignment work — and the gap between a well-tuned base and a vendor's best post-trained checkpoint is usually wider than a three-point index delta suggests.

Should you take the 574 GB?

• Download it if you already operate eight H200s or four GB300s, your workload is agentic or multi-step, and an MIT licence removes a legal obstacle you actually have.

• Download it if you need to reproduce the 44 yourself — that is now possible, and it was not on 22 October.

• Skip it if you were counting on 1M context out of the box; test the 256K config first.

• Skip it if your economics are per task and latency-sensitive. The 59% token increase is a real throughput cost, and no independent output-speed figure for 1019 exists yet.

• Start from the base only if you have a post-training pipeline. It is a foundation, not a shortcut.

The honest framing of OrcaRouter's role here: we do not serve Ling-3.1-flash-1019. What we sell is the other side of the decision — one key over 200+ hosted models at provider list price with 0% markup, so the question "does 1019 beat what I already run" can be answered without signing a separate contract for every baseline. Automatic failover and the routing DSL matter once you have picked a winner; they do not help you pick one.

What to watch next

Three things will decide whether this release ages well. The first is independent throughput: Artificial Analysis has no output-speed figure for 1019, and the first third-party provider to publish tokens per second will tell you more about production fitness than another index point. The second is whether a 1M-context scaling configuration appears in the repository or in the SGLang cookbook; until it does, the card's headline number is a claim with a 256K implementation underneath it. The third is what the community does with the base — a fine-tune that beats 44 on a narrow domain would be the strongest evidence that opening these weights was worth more than the score it shipped with.