
Hy4 flash vs MiniMax M3.1 Flash: Two Cheap Open-Weight Models, One With a Score
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 138 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 175 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 86 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 120 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 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 · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 116 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Here is the rare matchup where both models are downloadable, both are pitched at the same "cheap and fast" tier, both take images, both hold a million tokens, and one of them has a number from somebody other than its maker. MiniMax-M3.1-Flash has been on Artificial Analysis since its 19 October 2026 release and currently reads 29 on Intelligence Index v5.0.1, at $0.15 per million input tokens and $0.60 per million output. Hy4 flash has been public since 16 November 2026, went open-weight on 17 November, costs about half that on input and less than half on output, and has no Artificial Analysis entry at all.
So this page cannot tell you which model is smarter, because only one of them has been measured. What it can do is something more useful for anyone actually choosing: put the two release artefacts side by side, because these are genuinely comparable objects — same tier, same openness, same month — and the differences that survive are concrete ones about licensing, serving cost and what each vendor was optimising for.
Start with what both vendors actually shipped
Both releases include an instruction-tuned model, open weights, and a base or quantised sibling. Neither is a closed API with a download bolted on.
• MiniMax-M3.1-Flash — 428B total parameters, 23B active, mixture-of-experts, reasoning model, text and image in with text out, 1M-token context, weights on Hugging Face under the MiniMax Community License, which permits commercial use with restrictions.
• Hy4 flash — 220B total parameters (155.5B backbone plus 63.5B Engram plus a 0.43B vision encoder and a 0.85B drafter), 12B active, mixture-of-experts, text and image in with text out, 1,048,576-token context, weights on Hugging Face as Hy4 flash, Hy4 flash-FP8 and Hy4 flash-Base under the Tencent Hy License.
• What MiniMax did not ship — no separate base checkpoint and no base-model benchmark table; the 428B model is the artefact.
• What Tencent did not ship — no independent score, no published agentic benchmark suite for the instruction model, and no Transformers release (the integration lives on pull request 50903).
The parameter counts deserve a caveat rather than a comparison, because they are not measuring the same thing. MiniMax's 428B/23B is a conventional dense-parameter count. Tencent's 220B includes 63.5B of Engram n-gram embedding tables — parameters that are hashed and looked up rather than multiplied — so the two totals are not a like-for-like measure of model size any more than a library's shelf count is a measure of how much it has read. The active-parameter figures are the fairer comparison at 12B against 23B, and they predict what the price cards actually show.
The price rows, and the one place they invert
Both rate cards are the vendors' own list prices, not measurements.
• Input — Hy4 flash $0.070 against MiniMax-M3.1-Flash $0.15. Hy4 flash is 2.1x cheaper.
• Output — $0.281 against $0.60. Hy4 flash is 2.1x cheaper.
• Cached input — $0.003 against $0.015. Hy4 flash is 5x cheaper.
• Measured cost per Index task — $0.49 for MiniMax-M3.1-Flash per Artificial Analysis; not measured for Hy4 flash.
The last row is the interesting one, because it shows why per-token price is a bad proxy for per-task cost. MiniMax-M3.1-Flash's $0.49 per completed Index task is charged at a $0.15 input rate — more than double Hy4 flash's — and it still comes out cheaper per task than models with lower token prices, because the number of tokens a model spends getting to an answer varies more than the rate does. Tencent's claim that RL post-training cut its average output length by 50% is the same idea from the other direction: the cheapest model is usually the one that stops talking first.
So the honest reading of the price comparison is: Hy4 flash's rates are lower by a factor of two, and nobody knows whether that translates to a factor of two per task until an independent harness runs both. Tencent's own WorkIndex figures put Hy4 flash at 1/8 the per-task cost of DeepSeek-V4.1-Flash-1017, which is a different comparison on a different suite. Do not carry it across.

Speed: one measured, one claimed
• MiniMax-M3.1-Flash — 132 output tokens per second, 50.64 seconds to first answer token, 14.2-ish minutes of wall clock per Index task if you extrapolate from the $0.49 and 132 t/s figures. All Artificial Analysis.
• Hy4 flash — over 150 tokens per second decoding claimed by Tencent for the FP8 build on a tensor-parallel-8 node, and a 6.7-minute median end-to-end completion time in Tencent's own 85-user blind test.
Those are not comparable and the difference is instructive. MiniMax's number is a throughput measurement on a standardised harness; Tencent's is a serving claim for a specific container image on specific hardware, plus a task-completion median from an internal population. Neither vendor is being dishonest. They are answering different questions, and the second one — how long until the user has their answer — is the one that decides whether a product feels fast.
The 50-second time to first token on MiniMax-M3.1-Flash is worth flagging on its own. That is a reasoning model doing its thinking before it says anything, and it is fine for batch work and bad for anything interactive that streams. Hy4 flash has the same structural issue in a milder form: reasoning_effort defaults to high, and the documented alternative is no_think. There is no middle setting on Tencent's model, where MiniMax exposes a fuller reasoning configuration. If your latency budget needs "think a little," that is a real gap.
Licences: two different ways of being not-quite-open
Neither checkpoint is open source. Both are open weight with conditions, and the conditions are shaped differently enough that they will decide different deals.
• MiniMax Community License — permits commercial use with restrictions; the licence text governs deployment, and the model card carries it rather than an OSI-approved licence.
• Tencent Hy License v1.0 — Apache 2.0 plus a Section 10.2 that requires a separate written licence if you run a model-as-a-service business above $50 million in worldwide gross revenue in any twelve-month period, or serve more than 100 million monthly active users.
• Tencent's exemptions — a team that never resells access pays nothing extra, a product that exposes the model only inside a fixed feature rather than open prompting is not MaaS at all, and forwarding a request to somebody else's hosted endpoint does not count. Those three sentences are why a European SaaS company can usually read this licence once and move on, whereas the MiniMax text asks for a judgement call about what "restrictions" means for its own deployment.
• Tencent's trigger is revenue-linked, not territory-linked — there is no excluded jurisdiction, which distinguishes it from licences that bar the EU, the UK or South Korea outright.
Tencent's document is unusually readable for this genre because the additional terms are isolated in one numbered section that says exactly what it does to the Apache grant. Whatever else is true of the release, the licence was drafted by somebody who expected engineers to read it.

The score gap, and why "29" is not the whole story
Artificial Analysis measures MiniMax-M3.1-Flash at 29 on v5.0.1. If you have read anything else about this model on this blog, you may have seen 42 or 43 — those were correct on the previous index build, and the 27 October 2026 rebase moved every model on the board. Quoting 43 today and 29 for a competitor would be comparing two instruments. On the current build, 29 is the number, and it places MiniMax-M3.1-Flash in the same broad band as Gemini 3.8 Flash's 31 and below Claude Haiku 5.5's 36.
For Hy4 flash there is no band. The closest available evidence is Tencent's blind test, where it scored 2.79 out of 4 for user satisfaction against Claude Haiku 5.5's 2.81. If you are willing to make an inference — and it is only an inference — a model that lands within noise of a 36-index model on office tasks is probably not far off that band either. That is a hypothesis to test, not a score to cite.
What would settle it is one shared harness run. Both models are downloadable, which means the barrier is compute, not access, and the community that benchmarks Chinese open-weight models will get to it. Until then, the asymmetry is the fact: one of these has been publicly measured and one has not, and no amount of architecture description closes that.

Routing: one of these two you can call today
This is the practical difference, and it happens to run through our own catalogue. MiniMax-M3.1-Flash is on OrcaRouter as minimax/minimax-m3.1-flash, at $0.15 per million input tokens and $0.60 per million output with a $0.015 cache read — MiniMax's own list price, passed through at 0% markup, so a vendor price change is live here the same day. Multi-modal input is text, image and video, and the context window is 1,048,576 tokens.
Hy4 flash is not on the catalogue. What is there from Tencent is tencent/hy4-preview, the 770B/49B model Hy4 flash was distilled from, at $0.834 and $2.501 per million with a $0.042 cache read. That is a genuinely useful third model for this decision rather than a consolation: if you are weighing whether the Hy family's behaviour suits your workload, the preview tells you in an afternoon, and the open weights tell you what running the cheap one yourself would cost.
The routing DSL is where this gets more interesting than a two-way choice. A pipeline can send short interactive calls to the cheaper route and long document work to whichever model won your own eval, in one endpoint, with automatic failover across providers so that a model in its first week of availability is not a single point of failure. That posture — put the new model on the traffic you can afford to lose — is the right one for any checkpoint this young, and it is the same posture you would want for a self-hosted FP8 deployment behind an internal gateway.
Which one, for which job
Choose MiniMax-M3.1-Flash if you need a measured model now, if your inputs include video, if you need a reasoning-effort setting between "on" and "off," or if you want to be on a platform where the vendor's price is already live rather than a card you have to model yourself. Its 428B/23B shape and 132 tokens per second make it a workhorse, and the fact that it has been on the board for a month means other people have already found its failure modes in public.
Choose Hy4 flash if cost per token is the binding constraint, if you want to own the deployment, if you need a base checkpoint to fine-tune, or if the licence question is the one that decides — Tencent's document has no territorial exclusion and a clean internal-use carve-out, and for a European team that can be the whole answer. Its 12B active parameters and 1M-token context at 1/6 the KV cache of its predecessor are an argument that the second price, the one measured in GPUs, is unusually low.
Wait, in both cases, if the decision is expensive. One of these models has never been independently benchmarked and the other has never been run against it. The comparison that will actually settle this — both on one harness, scored on the same tasks, with cost per completed task beside the score — does not exist yet. When it does, the parameters and licence rows above will still be true and the score row will finally mean something.
Compared in this article4
Detected from this article · Benchmarks: Artificial Analysis · updated daily
