
Ling 3.0 Flash DSPark: The 1.36B Drafter That Quietly Shipped for Ling 3.0 Flash
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 87 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 151 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 190 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 86 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 246 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 97 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 138 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
Ling 3.0 Flash DSPark is a 1.36-billion-parameter draft model that Ant Group's inclusionAI lab uploaded to Hugging Face on August 9 with no press release, no blog post, and no launch thread — a repository that, at the time of writing, shows seven downloads and one like. It is not a new version of Ling 3.0 Flash, and it is not a model you will ever call directly. It is a speculative-decoding component: load it in front of Ling 3.0 Flash, and it proposes tokens for the big model to verify, letting the 124B/5.1B MoE emit several tokens per forward pass instead of one. Everything below that is not visibly true from the repository itself is labeled as unconfirmed, because the vendor has said nothing.
Ant Group announced Ling 3.0 Flash on July 27 and open-sourced its weights under MIT on August 7. The DSPark repository is a separate, quieter ship: created August 9, described only in a model card that identifies the checkpoint as "a DSpark speculator for Ling3," and touched again on August 21 with a config.json update. No announcement, no documentation page, no benchmark blog post. This piece collects what the repository actually says and marks the rest as open questions.
What a "DSpark speculator" is, without the jargon
Speculative decoding runs a small, cheap model ahead of the real one. The small model (the drafter) guesses the next handful of tokens; the big model (the target) checks the whole batch in a single forward pass and keeps the guesses it agrees with. When the drafter is right, one target forward pass yields several tokens, so throughput climbs without touching the target's weights.
DSpark is a version of that idea from DeepSeek's researchers — introduced in July 2026 and already deployed inside DeepSeek-V4 — and it is the algorithm SGLang calls DSPARK. Two things distinguish it from a plain drafter. It extends DFlash (a parallel-draft approach) with auxiliary features pulled from the target model itself, so the drafter gets information about what the big model is doing. And it carries a confidence head that scores each proposed token and stops the block early when a suffix looks unlikely to survive verification — so it does not spend a verify pass on tokens that are about to be thrown away.

The Ling-3.0-flash-dspark checkpoint implements exactly that: a speculator trained with SpecForge, sized to propose blocks of eight draft tokens, with a nine-token verify width once the target's bonus token is counted.
What the repository says, exactly
The model card is sparse but specific. Ling 3.0 Flash DSPark is a ~1.36B-parameter draft model (1,363,707,905 parameters, BF16) with five full-attention transformer layers at hidden size 2,560 and multi-head attention with 32 query heads and 32 key/value heads. It reads auxiliary features from five layers of the target — the card lists layers 1, 11, 23, 29 and 35 — and it pairs that with a confidence head described as a "vanilla Markov head, rank 256." Its position embedding tops out at 262,144, the same 256K context the target carries.
• Parameters — 1,363,707,905 (~1.36B), BF16, roughly 2.7 GB of weights.
• Architecture — 5 full-attention layers, hidden size 2,560, MHA with 32 query and 32 key/value heads; tagged qwen3.
• Target hooks — auxiliary features from Ling 3.0 Flash layers 1, 11, 23, 29 and 35.
• Confidence head — a vanilla Markov head, rank 256.
• Block — 8 draft tokens; verify width 9 including the target bonus token.
• Context — 262,144 position embeddings, matching the target's 256K native window.

That last row matters for deployment planning: the drafter is built for the same long-context serving profile as the target, so it does not shrink the context you can speculate over.
The acceptance numbers, read carefully
The card reports one set of performance figures: mean acceptance length, defined as "the mean number of tokens accepted per speculative verification step, including the target bonus token." Across nine workloads the macro mean is 5.29 — roughly five of the nine positions accepted per verify step, which is the number that would drive a speedup. The per-workload spread shows where the drafter is strong:
• Code and math — GSM8K 6.40, MATH-500 6.29, HumanEval 6.57, MBPP 6.34, LiveCodeBench 5.33.
• Reasoning — AIME 2025 5.56.
• Conversational — MT-Bench 3.92, Alpaca 3.51, Arena-Hard-v2 3.72.
Three caveats attach to those numbers. They are acceptance lengths, not end-to-end speedups: the card publishes no tokens-per-second figures, so "5.29 accepted tokens" has not yet been converted into a "x-times faster" claim for this model. They are vendor-reported and unreproduced — nobody outside inclusionAI has published a measurement. And they are context-dependent: acceptance length varies with workload, sampling temperature and the exact target checkpoint, so treat the code-and-math column as the strong case and the chat column as the realistic low end.
What is confirmed, and what is not
Confirmed by the public repository: the checkpoint exists and is downloadable; it is a DSpark speculator for Ling 3.0 Flash; it was trained with SpecForge; it is served through SGLang's DSPARK algorithm; and it was created on August 9 and updated as recently as August 21.
Not confirmed, and worth being precise about: the license. The card's frontmatter reads license: other, with no license text attached — so unlike the MIT-licensed Ling 3.0 Flash weights, the DSPark speculator's terms are currently unstated. Also unconfirmed: that the acceptance lengths reproduce on other hardware, what they imply in end-to-end tokens per second, whether any inference provider serves the drafter (nothing suggests one does), and whether an official announcement is coming. None of these are accusations; the release is under two weeks old. They are the difference between a promising component and a verified one.
Running it
In practice you need an SGLang build with DSPARK support — the algorithm landed in SGLang v0.5.16 in July, around the DSpark integration driven by DeepSeek-V4. Launch the server against the target model and name the drafter:
sglang serve --trust-remote-code --model-path <LING3_MODEL_PATH> --tp-size <TP_SIZE> --speculative-algorithm DSPARK --speculative-draft-model-path <LING3_DSPARK_MODEL_PATH>
The command is from the model card itself; the block size is read from the drafter's config. There is no llama.cpp path yet — the hybrid KDA + MLA architecture needs a custom fork even for the main model, so the speculator is an SGLang story for now.

Whether it is worth running
Ling 3.0 Flash DSPark only does work if you are serving Ling 3.0 Flash yourself. It is a local component: roughly 2.7 GB of extra BF16 weights alongside a 124B target, an SGLang build that supports the algorithm, and hardware with the memory to spare. The payoff is a higher acceptance length than a typical small drafter, in exchange for the memory and the moving parts. For teams already operating the target on their own GPUs at interactive, latency-bound loads, the drafter is a cheap thing to measure — two flags, greedy decoding, a look at the tokens-per-second delta. For anyone calling Ling 3.0 Flash through a hosted API, it is irrelevant: you cannot attach a local drafter to someone else's serving stack.
The model-card honesty gap is the main thing to budget around. A mean acceptance of 5.29 is a strong number — DeepSeek-V4's DSpark speculator reported acceptance lengths in the 3.2–3.6 range per the LMSYS write-up, and Liquid AI's LFM2.5-2.6B drafter averaged 4.83 — but acceptance length is not speedup, and until someone publishes end-to-end throughput on this exact pairing, the size of the win is an estimate.
And this is where the self-host-versus-hosted mix becomes the real decision. A drafter is one more self-hosted piece in front of a model that is also available as an API — and keeping those two options comparable is exactly the job a routing layer does. One API across 200+ models, automatic failover when a provider degrades, and provider list prices passed through at 0% markup let you run your own Ling-3.0-flash server for the latency-critical path while routing the rest of the traffic to hosted endpoints, and switch by configuration instead of by rewrite. The DSPark drafter does not change where Ling 3.0 Flash runs; it changes how fast it runs when it is on your hardware.
The open questions to watch
Watch for four things. An official announcement or blog post would fill in the missing context around the release. A published tokens-per-second figure — from inclusionAI or anyone else — would turn the 5.29 acceptance number into an actual speedup. A license change from other to a named license would clear the biggest blocker to production use. And a provider starting to serve Ling 3.0 Flash with speculation enabled would change the calculus for people who do not self-host at all.
Until then, Ling 3.0 Flash DSPark is what its repository says it is: a real, downloadable, 1.36B DSpark speculator for Ling 3.0 Flash, shipped without fanfare, with strong but unverified acceptance numbers and no published speedup. That is a workable basis for an experiment — and an honest description of a component nobody has independently measured yet.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
