Union Alpha vs DeepSeek V4.1 Flash: A One-Week Freebie Against Open Weights

Union Alpha is free today. DeepSeek V4.1 Flash is $0.15 per million input tokens — and it will still be there in December. That asymmetry is the entire comparison: one of these two models is an anonymous preview with a stated one-week window and no published weights, and the other is an MIT-licensed mixture-of-experts checkpoint you can download, audit, and self-host. If you are picking a model to build on rather than to play with, the free one is not the safer bet.

What Union Alpha actually is

Union Alpha surfaced on third-party model catalogues on 16 September 2026 and is now served on OrcaRouter's free tier. Its operator is anonymous — the provider field literally reads "Stealth" — and no lab has claimed it. That is not a knock on the model; it is the single most important fact about it, because everything else you can say is downstream of it.

What is verifiable is the endpoint's shape, because you can call it and read the metadata:

• Context window — 262,144 tokens (256K), with a maximum output of 131,072 tokens.

• Inputs and outputs — accepts text and images, emits text only. No video input, no audio.

• API surface — tools, tool_choice, response_format, temperature, top_p, and max_tokens. Structured JSON output works; strict JSON-schema enforcement does not. In practice tool_choice only honours "auto".

• Reasoning controls — none exposed. There is no effort dial, no thinking budget, no visible reasoning channel.

• Provenance — no weights, no licence, no parameter count, no architecture note, no stated knowledge cutoff. The tokenizer is filed as "Other".

• Price — $0 per million input and output tokens during the preview, rate-limited, with HTTP 429 once you hit your plan's cap.

The one performance claim attached to it, "frontier-level performance across a broad range of general-purpose tasks", comes from the operator and has never been independently reproduced. There is no leaderboard entry, no vendor benchmark table, and no model card behind it.

What DeepSeek V4.1 Flash actually is

DeepSeek V4.1 Flash launched on 10 September 2026, six days before Union Alpha appeared, and it is the opposite kind of object. It is a sparse mixture-of-experts model built on a new Causal Encoder-Decoder architecture, with a 552B-parameter backbone (16B active on output, 8B on input) plus 196B Engram conditional-memory parameters — roughly 763B total — trained on 45T tokens and released under the MIT licence.

Two engineering details matter more than the parameter count. Compressed Sparse Attention 2 combined with FP4 KV caching brings the cache down to roughly 890 bytes per token, about a quarter of what DeepSeek V4 Flash used, which is what makes a 1M-token context economically sane. And DeepSeek is folding its own product line into it: legacy V4 Flash and V4 Pro requests are being routed to V4.1 Flash at Flash pricing.

• Context window — 1,000,000 tokens, against Union Alpha's 262,144.

• Max output — 384,000 tokens per response, against 131,072.

• Inputs — text and image (native vision); output text only.

• Price — $0.15/M input, $0.60/M output, $0.003/M cached input off-peak; doubled to $0.30/$1.20/$0.006 during peak hours (Mon–Fri, 01:00–04:00 and 06:00–10:00 UTC). Weekends are always off-peak.

• Weights — MIT, downloadable, self-hostable.

• Observed speed — 215 output tokens per second, with a p50 time-to-first-token of 842 ms, a p95 of 2.40 s and a 1.2% error rate.

The benchmark picture is lopsided, and not in the way you would guess

DeepSeek publishes. Union Alpha does not. That is the whole of it.

V4.1 Flash ships with a full table, though it is largely DeepSeek's own reporting plus aggregator entries rather than a single independent audit: GPQA Diamond 90.9, HLE 63.9 with tools and 36.8 without, Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2, CyberGym 88.1, MathArena Apex 65.6, and an Artificial Analysis Intelligence Index of 40. Note the gaps inside that table — Terminal-Bench 4.0 comes in at 31.2, drastically below the 2.1 score, which is a reminder that vendor headline numbers cluster around whichever suite flatters the model.

For Union Alpha, exactly one independent measurement exists. The SMF Clearinghouse ran it through 157 tests of the Official A suite with reasoning off and scored it 136/157 — 86.6%, zero errors, ranking tenth of twenty-six entries. Reasoning 30/30, tools 2/2, coding 26/30, math 24/30, and a notably weak writing suite at 2/5. That is a real data point and it is the only one.

So the honest summary is: V4.1 Flash has many numbers, most of them vendor-flavoured but at least visible and dated. Union Alpha has one number, and it is a decent one.

Price is the only axis Union Alpha wins — and it wins it outright

Free beats $0.15/$0.60. There is no clever framing that changes this. For a week of evaluation work, batch experiments, or a throwaway agent loop you want to run a hundred times, Union Alpha costs nothing and V4.1 Flash costs real money at scale.

But "free" is carrying an unusual amount of weight here. The preview is rate-limited, so throughput is capped by your plan rather than by the model. There is a single endpoint behind it, which means no provider fallback if that endpoint wobbles. And the free window is the vendor's to close — no post-preview price has been announced.

This is also where routing changes the arithmetic. OrcaRouter passes provider list price through at 0% markup, so when DeepSeek moves its off-peak rates, the change is live on our side the same day rather than waiting on a repricing cycle. Both models sit behind the same key, which means you can run the comparison above as a config change rather than two integrations — and get Union Alpha for $0 on the free tier and V4.1 Flash on the paid path without a second contract.

Which one should you actually call?

The deciding question is not capability. It is whether you need the answer to survive next month.

Call DeepSeek V4.1 Flash if the work goes to production. The 1M-token context, the 384K output ceiling, the MIT weights, the dated benchmark table, and the fact that DeepSeek is consolidating its own legacy traffic onto it all point the same way: this is a model with a future and a paper trail. At $0.003 per million cached input tokens, a heavily cached long-context workload is close to free anyway.

Call Union Alpha if you are exploring, benchmarking against, or simply curious about a model that might be a fast tier of something much larger. Its vision input and 256K window are genuinely useful for document and screenshot work, and the Official A result suggests it is competent rather than embarrassing. Just keep it off any path where a sudden 429 or an unannounced shutdown costs you something.

And if you want to do both without re-plumbing anything, that is the case for a routing layer rather than a direct integration.

What would change this verdict

Three things, in order of likelihood.

A named lab claims Union Alpha. Precedent says this is plausible: the previous anonymous entry in this series, Ox Alpha, was revealed six days after appearing as Zhipu's GLM-5.3-Flash. Reveal brings a licence, a parameter count, and almost certainly a price.

An independent leaderboard scores it. Right now a single 157-test suite is the entire public record. One Artificial Analysis entry with a dated methodology would move Union Alpha from anecdote to data.

The free window closes quietly. This is the default outcome for stealth previews — Ox Alpha was pulled from the catalogue after its preview ended. If Union Alpha disappears without a post-preview price, "free" will have cost exactly one week of integration work, which is why pairing it with automatic failover is the prudent way to use it today.

The paid path is one key away: DeepSeek V4.1 Flash model page lists the off-peak $0.15/$0.60 rates with no per-token margin added, and the peak-hour doubling is itemised on the same page.