Ant open-sources a finance model built for research, not chatter

Share
Ant open-sources a finance model built for research, not chatter

Two open-weight releases landed within hours of each other today, and both are about the same unglamorous problem: making a model do a real job end to end instead of answering plausibly. Ant Group put the weights for its finance-tuned Ling model on Hugging Face, and the local-inference crowd got a speedup that doubles throughput on a consumer card.


Ant Group released the open weights for Ling-3.0-flash-Fin, a finance-tuned model that trades breadth for verifiable work. The model keeps the Ling-3.0-flash architecture — 124 billion total parameters with about 5.1 billion active per token, and a 256K context window — but adds continued training on financial data, and it ships under an MIT license with 64 checkpoint shards on Hugging Face. Ant built it with financial institutions and domain experts, and the pitch isn't chat: it's annual reports, multi-document research, valuation modeling, and spreadsheet workflows where a wrong unit or a wrong reporting period ruins the answer. The company is blunt about the ceiling — its own card says key assumptions, valuation results, and investment conclusions still need professional review, and that long-horizon workflows need more validation.

The more interesting artifact may be the one next to it. Ant also open-sourced FinFIRST, an evaluation set built around the idea that financial research fails in the middle, not at the end: each expected answer is broken into independently checkable criteria, so an agent gets credit for correct intermediate work and gets flagged when it lands on the right number through the wrong source. Model builders shipping their own graded eval alongside the weights is a healthier pattern than a marketing chart, and it's the part competitors will have to match.

What makes this worth watching is the timing. Domain-tuned open weights are where the real deployment pressure is sitting right now — most firms don't need a better general assistant, they need one that reconciles a 10-K against an earnings call and shows its work. A 124B model with 5.1B active parameters is also cheap enough to actually run, which is the difference between a benchmark entry and something a bank puts behind a firewall.


Microsoft capped Xbox Cloud Gaming at 15 hours a month for Game Pass Ultimate, and opened pay-as-you-go access to everyone else. The limits land in November: 15 hours for Ultimate at $22.99 a month, 10 hours for Premium at $14.99, and 5 hours for Essential at $9.99 — replacing the unlimited cloud access all tiers have had for the past year. Extra hours go on sale through the Xbox Store, and Microsoft says unused hours roll over; it hasn't said what they cost. Non-subscribers will also be able to buy time to stream games they already own.

The framing from Xbox is refreshingly direct: "the cost of providing cloud gaming grows as more people use it and play for longer." That is the same sentence every AI product team is now writing about inference, and the 4% figure Microsoft cites — the share of Game Pass subscribers who stream more than the new cap — is the tell. This isn't a price hike disguised as a feature; it's usage-based pricing arriving in a subscription that promised unlimited, because a small tail of heavy users was carrying the cost.


Local inference got a big free speedup: ik_llama.cpp now supports the multi-token-prediction head that ships inside Qwen3.8-Flash-Next. The checkpoint carries a 2.6B draft head that public converters were throwing away; with it loaded, the model proposes its own next tokens and then verifies them, so output is identical to running without it. One contributor measured 45 to 90 tokens per second on a 5090 with 128GB of system memory on coding traffic, and 9.5 to 12.5 on a 12GB 4070 — meaning the trick works on a mid-range card, not just a workstation. The code is on main, no fork required, and existing quantized files can pick up the head without a re-download.

It is not a uniform win, and the caveats are the useful part. On an RTX Pro 6000, code generation went from 85 to 113 tokens per second while prose dropped from 83 to 59: acceptance rates run 93–99% on code and only 60–65% on ordinary writing, so a drafting head trained on code drafts code well. That gap is the whole story of speculative decoding right now — the speedup follows the predictability of the text, not the size of the GPU.

What to watch: whether Ant's open evaluation set becomes the template other domain-model builders have to publish against.

Would you trust a finance model's valuation work if it shipped its own graded eval — or does the benchmark need to come from someone else? Write to us via our contact page.

Sources: ik_llama.cpp PR #2369 (GitHub)

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Tencent open-sources Hunyuan Hy4 preview — 770B params, 1M context

Tencent open-sources Hunyuan Hy4 preview — 770B params, 1M context

China's open-weight race just got its heaviest new contender, and Tencent is pitching it as a productivity workhorse rather than a benchmark-chaser. Tencent is open-sourcing Hy4 preview, a 770 billion-parameter Mixture-of-Experts model with a 1 million-token context window. The architecture activates only 49 billion parameters per token — a 78-layer stack where each layer beyond the first holds 256 routed experts plus one shared expert, with eight experts firing per token.

AllSpark's Iris beats every open-weight search agent — and says its scaffolding did the work

AllSpark's Iris beats every open-weight search agent — and says its scaffolding did the work

A paper out of AllSpark Research claims the strongest open-weight deep-research agents in their size classes, and the authors' most interesting admission is not a benchmark number: it is that half the story belongs to the harness wrapped around the model. AllSpark released Iris-mini and Iris-pro, two open-weight search agents it says beat every comparable open model on multi-hop browsing benchmarks. The recipe is the unusual part.

Unitree open-sources a 6B humanoid brain that beats GPT-6 Astra

Unitree open-sources a 6B humanoid brain that beats GPT-6 Astra

Robot makers have spent two years shipping better bodies than brains. This week one of them open-sourced the missing half. Unitree released UnifoLM-WLA-1.0, a 6B-parameter humanoid foundation model trained on about 2,500 hours of real-robot data that runs 64 tasks from a single checkpoint — 54 desktop, 10 whole-body — across two-finger grippers and several five-fingered dexterous hands, and it published the weights, code and datasets instead of a demo reel.

Ornith 1.5 open models match Claude Opus 4.8 on coding benchmarks

Ornith 1.5 open models match Claude Opus 4.8 on coding benchmarks

An open-source model family just posted numbers that rival one of the most capable closed models in the world — and it's available under an MIT license. Ornith 1.5, released today by the Ornith AI team, is a three-model family spanning 9 billion dense parameters up to a 397 billion mixture-of-experts flagship. The big headline: the 397B MoE scores 86.1 on Terminal-Bench 2.1 (Terminus-2) and 86.0 on SWE-bench Verified, placing it on par with Claude Opus 4.8's 85.0 and 85.8 respectively.

Qwen3.8-27B hits 1M downloads as small open models close the frontier gap

Qwen3.8-27B hits 1M downloads as small open models close the frontier gap

Alibaba's latest open-source release is turning heads: Qwen3.8-27B, a 27-billion-parameter model that runs on personal computers, has crossed one million downloads within days of launch and is outperforming models several times its size on independent benchmarks. The model scored 52 on the Artificial Analysis Intelligence Index — placing it above every medium-class open model (40B–150B parameters) and tying with DeepSeek V4 Flash, which ranks fifth among all large models.