Ant's Ling-3.0-flash-VL scores 42 on the Artificial Analysis index
Ant Group's open model team spent the day pushing Ling into vision, while the local-inference crowd shipped two things you can actually run tonight: a denser weight format for ternary models and the first serious head-to-head of 27B quants on a 16GB card.
Ant Ling released Ling-3.0-flash-VL, a vision-language version of its 124B hybrid-linear MoE, and claims it scores 42 on the Artificial Analysis Intelligence Index — four points above the 38 the text-only Ling-3.0-flash earned in August. The model keeps the parent's economics: 124 billion total parameters with 5.1 billion active per token, on a 42-layer backbone that alternates Kimi Delta Attention linear-attention layers with gated MLA layers at 5:1, plus a vision tower feeding in through a two-layer MLP projector and VideoRoPE for time encoding. Ant is pitching a 1M-token context and a cluster of capabilities — visual perception, STEM reasoning, document intelligence, multimodal agent work, frontend coding from screenshots, and medical report interpretation — with demos showing the model generating a site, rendering it in a browser, comparing the result to a design reference and revising the code.
The interesting claim is the one buried in the numbers. Ant's own follow-up post argues that adding vision raised the text-only intelligence score rather than diluting it, which would mean the multimodal training is doing real work instead of just bolting a camera onto a language model. That is a testable, and so far untested, claim: the score is the company's, not an independent run, and as of this evening there is no Ling-3.0-flash-VL entry on Hugging Face for anyone to check it against. Ant open-weighted everything before it, including the text-only Ling-3.0-flash, so the weights are likely coming — but a release post plus a benchmark chart is not a release. Until it lands, the 42 is a claim, not a result.
A llama.cpp fork adds a base-3 packing format that shrinks ternary model weights by roughly 22% with no accuracy cost. The format, Q2_B3 or "B3S," exploits a fact general quantizers ignore: models like BitNet b1.58 only ever store -1, 0 or +1 per weight times a block scale, so packing those three states in base 3 — five trits to a byte — gets below the 2 bits per weight a generic format spends. On a 128-weight block it costs 28 bytes, or 1.75 bits per weight, against roughly 2.25 for a standard 2-bit representation. In practice a 27B ternary model's weights drop from about 7.6 GB to about 5.9 GB; a 9B drops from 2.5 GB to 2.0 GB. The catch is that it only works on models that are already ternary — run a normal FP16 model through it and quality collapses — and the author has verified AMD ROCm and CPU paths only. CUDA and Metal compile but are untested, and there are no perplexity or tokens-per-second tables yet, just the author's note that his own runs showed no meaningful slowdown.
The framing here matters more than the single-card saving. Ternary weights have been a research curiosity since Microsoft's BitNet papers because the memory story was compelling and everything else — training stability, kernel support, quality at scale — lagged. Formats like this are the unglamorous plumbing that makes the memory story pay off, and 1.75 bits per weight sits only about 10% above the information-theoretic floor for three states. The open question is upstreaming: llama.cpp already merged ternary support a few weeks ago, and the author doesn't expect this fork to be accepted.
A LocalLLaMA benchmarker ran 21 different quantizations of Qwen3.8 27B on a single RTX 5080 and ranked them by how far each drifts from the full-precision model. The test used KL divergence against a high-precision reference on the author's own C code rather than a generic perplexity set. The winner that fits in 16GB is a 14.5 GB IQ4_XS quant, scoring 0.056 mean KLD and agreeing with the reference model's top token choice 95.8% of the time. The cheapest options are much worse: a 8.2 GB 2-bit variant drifts to 0.893 KLD and 85.7% agreement. The two best quants overall, at roughly 16.4 GB, don't fit at all — which is the whole point of the exercise for anyone on a 16GB card.
Two details are worth more than the ranking. First, the spread between the best and worst near-identically-sized quants is large: at roughly 13GB, an entry from one quantizer posts 0.093 KLD while another lands at 0.147 — a real quality difference for the same VRAM. Second, the author's own caveat that a 16GB card only fits about 60k context with the winning quant, and only with aggressively quantized KV cache. Size on disk is the number everyone shops by; the context you can actually hold is the one that decides whether the model is usable. The model itself has been the local community's default 27B for a month — One week with Qwen 3.8 27B — the local LLM community's verdict — and if the tradeoff between bits and quality is new territory, AI 101 — What is model quantization? covers the mechanics.
What to watch: whether Ant ships Ling-3.0-flash-VL weights to Hugging Face, and whether anyone independent confirms that 42.
Which matters more to you right now — the headline intelligence score, or the weights actually being downloadable? Tell us in the comments.
Sources: llama.cpp ternary B3S fork (GitHub) · Ternary Q2_0 repacker (GitHub)
