Overview
abab6.5 and abab6.5s are a pair of large language models that MiniMax announced on 17 April 2024 as the abab6.5 series. abab6.5 is the flagship: a Mixture-of-Experts (MoE) chat model with roughly one trillion total parameters and a 200,000-token context window. abab6.5s is its lighter, faster sibling — MiniMax says it was built with the same training techniques and data but optimized for efficiency, processing nearly 30,000 words in about one second. Both were offered through MiniMax's own products (such as the Hailuo AI assistant) and its open API platform rather than as downloadable weights.
At launch MiniMax positioned abab6.5 as the point where its core model capabilities — across knowledge, reasoning, math, coding, and instruction following — began to approach the leading proprietary models of early 2024, namely OpenAI's GPT-4, Anthropic's Claude-3, and Google's Gemini-1.5. MiniMax did not publish a numeric benchmark table for the series, so the comparison was framed qualitatively rather than with head-to-head scores. The series did headline one concrete long-context result: in 891 'needle-in-a-haystack' tests, where an unrelated sentence is hidden inside a long 200k-token document, abab6.5 retrieved the answer correctly every single time (891/891).
abab6.5 / abab6.5s are now a legacy line. MiniMax moved on to the open-weight MiniMax-01 series (MiniMax-Text-01 and MiniMax-VL-01, January 2025) and then the M-series reasoning models (MiniMax-M1, M2, M3). The abab6.5 variants no longer appear on MiniMax's current API pricing page, and MiniMax never released their weights, so they are best understood as a historical, China-built, trillion-parameter MoE milestone rather than a model you would deploy today. They were text-only and aimed primarily at Chinese- and English-language applications.
| Released | 2024-04-17 |
|---|---|
| License | Proprietary (closed; hosted API only — weights never released) |
| Weights | API only |
| Parameters | abab6.5: ~1 trillion total parameters (Mixture-of-Experts). abab6.5s: not separately disclosed (described as the same training recipe and data, optimized for efficiency). Active-parameter-per-token counts were never published. |
| Context | 200k tokens |
| Max output | Not officially published |
| Architecture | Mixture-of-Experts (MoE). The abab6.5 series builds on the MoE direction MiniMax took with its earlier abab6 model (described by MiniMax as one of the first MoE-based large language models in China). MiniMax disclosed abab6.5 as a roughly trillion-parameter MoE chat model with a 200,000-token context window; abab6.5s shares the same training techniques and data but is tuned for higher throughput, processing close to 30,000 words/characters in under a second. MiniMax did not publish layer counts, expert counts, routing details, or active-parameter figures for this line. |
| Knowledge cutoff | Not officially stated |
| Modalities | Text |
| Status | Retired / legacy. Announced 17 April 2024; superseded by the MiniMax-01 series (January 2025) and the later M-series (M1/M2/M3). The abab6.5 variants no longer appear on MiniMax's current API pricing page. |
Strengths
- Large 200,000-token context window — among the longer commercial context lengths available in early 2024
- Strong long-context retrieval at launch: a perfect 891/891 on MiniMax's 200k-token needle-in-a-haystack test
- Trillion-parameter Mixture-of-Experts design, building on MiniMax's early MoE work in China
- abab6.5s offered high throughput (around 30,000 words per second) for latency-sensitive use
- Bilingual focus — designed for both Chinese- and English-language applications
Best for
- Long-document chat and analysis needing a large context window (historical, via MiniMax's hosted API)
- Chinese- and English-language conversational assistants (e.g. MiniMax's Hailuo AI)
- High-throughput, latency-sensitive chat where abab6.5s traded some capability for speed
- A reference point for understanding MiniMax's model lineage before MiniMax-01 and the M-series
How to access
| Provider | Model ID |
|---|---|
| MiniMax ↗ | abab6.5-chat / abab6.5s-chat (legacy) |
abab (legacy) — every version
The full lineage of the abab (legacy) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| abab6.5 / abab6.5scurrent | 2024-04-17 | 200K | Proprietary |
| abab6 | 2024-01 | 32K | Proprietary |
| abab5.5 / abab5.5s | 2023 | 16K | Proprietary |
FAQ
What is the difference between abab6.5 and abab6.5s?
abab6.5 is the flagship — a roughly trillion-parameter Mixture-of-Experts chat model with a 200,000-token context window. abab6.5s was built with the same training techniques and data but optimized for efficiency: it shares the 200k-token context yet runs much faster, processing close to 30,000 words in about one second. In short, abab6.5 favors capability and abab6.5s favors speed.
Are abab6.5 and abab6.5s still available?
Not really. They are a legacy line from April 2024. MiniMax has since shipped the open-weight MiniMax-01 series (January 2025) and the M-series reasoning models (M1, M2, M3), and the abab6.5 variants no longer appear on MiniMax's current API pricing page. The weights were never released, so there is no way to self-host them.
How good was abab6.5 at long-context retrieval?
MiniMax highlighted one concrete result: across 891 'needle-in-a-haystack' tests — where an unrelated sentence is hidden inside a long 200,000-token document — abab6.5 answered correctly every single time, a perfect 891/891. MiniMax did not publish a full numeric benchmark table for the series.
Were abab6.5's weights open-sourced?
No. abab6.5 and abab6.5s were closed, proprietary models available only through MiniMax's products and hosted API. MiniMax did not begin releasing open weights until the later MiniMax-01 series in January 2025.