AuK-Flash vs MathForm-8B: Two Quiet Specialists on Borrowed Qwen Brains

AuK-Flash and MathForm-8B have more in common than either lab probably realizes, and none of it is the task they do. MathForm-8B is OpenBMB's 8B autoformalization model — an Apache-2.0 fine-tune of Qwen3-8B that turns natural-language mathematics into Lean 4 statements a compiler can check. AuK-Flash is Tencent's distilled speech model — an MIT-licensed diffusion generator for speech synthesis and editing that routes its instructions through a Qwen2.5-Omni-3B encoder before it makes sound. One converts prose math into proof-checkable formal text; the other converts text and instructions into audio. They share the same architectural strategy from opposite directions: neither trains a general language brain from scratch — each wraps an existing open model and spends the real effort on its output modality. Both shipped the same way too — weights quietly on Hugging Face, no press release — and both are exactly the kind of narrow, self-hosted release that a frontier-model benchmark cannot capture.

The pattern that unites them

Look at the two release trails side by side and they are near-mirrors. Tencent's AuK-Flash repository is dated August 21 on Hugging Face (the sibling AuK base, August 18); the open-source announcement — code, weights and demo links — landed only this week, dated September 7 on the Hugging Face card and September 9 on the linked GitHub repo. OpenBMB's MathForm-8B repository appeared August 14 with no announcement at all. In both cases the model card is the launch, the weights are ungated under a permissive license, and there is no hosted API to try — you download the checkpoint and run it on hardware you control. For a reader deciding what to adopt, that shared shape matters more than the difference in domain: both are bets that a tightly-scoped open model can serve a niche a generalist serves badly, and both ask you to supply the evaluation the vendor did not.

The honest scoreboard

Because the two models do not overlap on any benchmark, the scoreboard is a portrait of two different bets rather than a ranking. Sourcing matters on every row — AuK-Flash's claims are Tencent card statements days old and unreproduced; MathForm-8B's figures are OpenBMB-reported from arXiv 2608.14221 and unreproduced:

• What it is — AuK-Flash: Tencent's ~1.5B speech generation and editing model, distilled to a 4-step diffusion variant. MathForm-8B: OpenBMB's 8B autoformalizer turning informal math into Lean 4.

• Output — AuK-Flash: 24 kHz audio — speech, edited speech, separated sources. MathForm-8B: Lean 4 statements with a header and named theorem.

• Construction — AuK-Flash: diffusion transformer distilled to fixed NFE 4 / CFG 0, with Qwen2.5-Omni-3B as the instruction encoder. MathForm-8B: Qwen3-8B fine-tuned with SFT then RL on the ~367K-example FormalVerse dataset.

• Input language — AuK-Flash: Chinese and English (per the model card). MathForm-8B: English, in the register of math problem statements.

• Release — AuK-Flash: repos August 18/21; open-source announced September 7–9. MathForm-8B: repo August 14; no announcement as of writing.

• Evidence — AuK-Flash: none yet beyond the card's capability list. MathForm-8B: vendor-reported 88.06% Pass@8 under syntax check and 72.37% under consistency check, with FATE-H 63% and FATE-X 37% on the hard consistency sets.

The scoreboard in six rows: both are open-weights specialists with borrowed Qwen brains and thin independent evidence — they differ in what they make, not in how they were released.

AuK-Flash: speech as the specialist's output

The "borrowed brain" claim for AuK-Flash is literal, not rhetorical. The checkpoint Tencent publishes contains the diffusion transformer and layer-fusion weights; the model that actually understands your instruction — Qwen2.5-Omni-3B — is downloaded separately and loaded at runtime, and so is the BigVGANFlowVAE that turns the latent back into a 24 kHz waveform. What Tencent trained is therefore not a language model at all but a conditional flow-matching generator: given a semantic plan from the Qwen encoder, it renders audio in a handful of steps. AuK-Flash is the four-step distilled version of that generator, and its repo — about 6.1 GB for the Flash checkpoint in BF16 plus a 637 MB VAE — ships with a CLI, a Python engine, Gradio and ComfyUI nodes, and a fine-tuning script. The capability list is wide for a speech model: zero-shot and instruction TTS, content and lyric editing, pitch/speed/volume and emotion/timbre changes, de-accenting, whisper conversion, enhancement, and source separation, all behind one natural-language instruction interface in Chinese and English. Whether each works as described is untested outside Tencent; the architecture claim — a small Qwen that understands, a distilled diffusion model that makes sound — is visible in the repo itself.

The tencent/AuK-Flash repository page — MIT-licensed weights for the distilled 4-step speech model, with the Qwen2.5-Omni-3B encoder and VAE loaded from separate files at runtime.

MathForm-8B: formal proof as the specialist's output

MathForm-8B is the same pattern one domain over. OpenBMB took Qwen3-8B — a generalist trained across 119 languages — and spent its entire capability on one output shape: a Lean 4 theorem statement derived from a natural-language problem. The SFT stage on FormalVerse teaches the informal-to-formal mapping; an RL stage then sharpens it against the Lean compiler's verdict, which is why the model reports not a vague quality score but a Pass@8 under syntax check and a stricter pass under semantic consistency check. The vendor numbers are strong — 88.06% and 72.37% across six benchmarks, beating the paper's 7B–32B specialized baselines — and just as unreproduced as AuK-Flash's claims. The repo is Apache-2.0, served through Transformers, vLLM or SGLang, and gives up essentially everything Qwen3-8B is known for in exchange: no 119-language general chat, a roughly 16K-token output budget for Lean, and no documented ability outside formalization.

The openbmb/MathForm-8B repository — Apache-2.0 weights for the autoformalizer, listing Qwen3-8B as its base model. This is MathForm-8B's entire public surface.

Verification is the theme neither benchmark captures

Put the two outputs next to each other and a strange symmetry appears. MathForm-8B's whole design exists because natural-language math has no correctness signal — the Lean compiler supplies one, so the model is trained against a pass/fail verdict it can actually feel. AuK-Flash's domain has the same problem with a different solution: there is no compiler for "does this clip sound like the speaker, in a whisper, with the cough removed," so the pass/fail signal is a human ear. Neither aggregate benchmark measures that — an Elo on text or a quality score on synthesized speech tells you little about whether the specialist's core promise (verified Lean, or editable, cloneable speech) holds in your workflow. The practical consequence is that both models must be evaluated the way the vendor could not evaluate them: MathForm-8B by running its output through Lean, AuK-Flash by listening to its output on your own data. Until someone does that, the headline claims on both cards are directions, not results.

Who each is for, and who should wait

• Choose MathForm-8B when your workload ends at formalization — converting problem banks, building Lean corpora, feeding proof automation — and you want the strongest open specialist at this weight. It is not a general model, so pair it with one for everything around the formal step.

• Choose AuK-Flash when your workload ends in audio — a voice to speak your text, a recording to edit, a mix to separate — and you want an open model that treats generation and editing as one operation. Budget for the Qwen encoder beside it, and for your own listening test.

• Wait on both if you need a proven, supported piece of infrastructure: neither has independent results, neither has a hosted API, and both are days-to-weeks old. The pattern to steal is architectural — a small borrowed language brain plus a specialist output head is far cheaper to train and iterate than a full generalist — and it is a pattern you can adopt in your own fine-tuning whether or not you ever run these two checkpoints.

Both releases also illustrate why a routing layer earns its place in a mixed stack: when your pipeline crosses from a general reasoning model into a specialist like one of these, the point of the router is that you do not commit the architecture to a single answer. One API across 200+ models, automatic failover if a provider degrades, and provider list prices passed through at 0% markup means you can keep the generalist on the default path and call the specialist only for the subset of requests that need it — and swap the specialist the day a better one ships.

The comparison to hold onto is not AuK-Flash against MathForm-8B — they will never compete for the same request. It is the two of them against the assumption that a frontier generalist is the only model worth running. Speech editing and verified formalization are both domains where a small, focused, open model with a borrowed brain can beat a much larger one that was never pointed at the problem — which is precisely the bet both Tencent and OpenBMB just published, and precisely what still needs independent proof.