Alex Getmanearly AI signals
EN

MiniMax M3

MiniMax

Released Updated
Context window
1M
1,000,000 tokens
Input / Output
image, text, video → text

Timeline

  1. · Weights version · evidence link
Input / output price · USD / 1M tokens
ModelEvidence linksInput / output price
minimax-m3
standard · text
opencode-go$0.30 / $1.20
  1. · Evidence links · GitHub search · MiniMax · commits · rodrigueztaylor87/rod · @rodrigueztaylor87
  2. · Evidence links · GitHub search · MiniMax · commits · SilverKnightKMA/paseo-plugins · @SilverKnightKMA
  3. · Evidence links · GitHub search · MiniMax · commits · SilverKnightKMA/paseo-plugins · @SilverKnightKMA
    Excerpt
    now carry the FULL pi reference '<pi-provider>/<model>' (magpie/minimax/MiniMax-M3),
  4. · Evidence links · GitHub search · MiniMax · commits · Lore-Hex/quill-router · @jperla
  5. · Evidence links · GitHub search · MiniMax · pulls · xiaoY233/Chat2API · @octo-patch
    Excerpt
    feat(minimax): add MiniMax-M3 model entry
  6. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @Vincent-PRO-AI
    Excerpt
    [Model][MiniMax-M3] SM120 Blackwell NVFP4 KV cache Triton reader
  7. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @Pb314314
    Excerpt
    use sparse MLA (dense MLA, DeepSeek-V4 / V4.1, MiniMax-M3, non-MLA models) are not affected; the reason for each is in "Per-model check".
  8. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @cjackal
    Excerpt
    - Pixtral, Cohere Compass and MiniMax-M3 also call `get_cu_seqlens_padding_logics`, so every encoder-CG model whose `cu_seqlens` comes from `maybe_recompute_cu_seqlens` takes its padding from one place. Without
  9. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @MrGeva
    Excerpt
    [Perf][MiniMax-M3] Two-shot fused MoE finalize + all-reduce + norm + MXFP8 up to 512 tokens
  10. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @MrGeva
    Excerpt
    [Perf][MiniMax-M3] Raise the fused MoE finalize + all-reduce cap from 32 to 128 tokens
  11. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @afriedri
    Excerpt
    short prompts at high concurrency is higher on MiniMax-M3 (TP8: up to +21% at 1024 tokens, conc 64; TP4: +14% at 1024
  12. · Evidence links · GitHub search · MiniMax · pulls · vllm-project/vllm · @wolfcomos
    Excerpt
    Add `minimax_m3_msa_decode_fp16_dot`, an opt-in optimization for MiniMax-M3 Triton sparse decode on SM100-family GPUs. It uses FP16 dot operands with FP32 accumulation and selects splits targeting
  13. · Evidence links · GitHub search · MiniMax · pulls · viztor/dsh-opencode-patch · @viztor
    Excerpt
    | `opencode-go/minimax-m3` | info | `[KNOWN_DEFECT]` toggle-only model offered Minimal…Max | Off / On (D1) |
  14. · Evidence links · GitHub search · MiniMax · pulls · thequanhcm/quanpt1 · @thequanhcm
  15. · Evidence links · GitHub search · MiniMax · pulls · tenstorrent/tt-metal · @mdragulaTT
  16. · Evidence links · GitHub search · MiniMax · pulls · tenstorrent/tt-metal · @mdragulaTT
    Excerpt
    The MiniMax-M3 vision encoder on one Blackhole chip. It turns one image, as the media module's `prepare` resizes it, into its vision tokens: bf16 `[n_tokens,
  17. · Evidence links · GitHub search · MiniMax · pulls · sgl-project/sglang · @alphabetc1
  18. · Evidence links · GitHub search · MiniMax · pulls · openJiuwen-ai/jiuwenswarm · @octo-patch
  19. · Evidence links · GitHub search · MiniMax · pulls · jundot/omlx · @begcdn
    Excerpt
    On M2 Ultra (`applegpu_g14d`) the MiniMax-M3 sparse prefill path crashes once a prompt crosses 4096 tokens: `msa_prefill_supported` gates only on shape (`D=128`, `block_size=128`, `qhead_per_kv=16`), so
  20. · Evidence links · GitHub search · MiniMax · pulls · iv-org/invidious · @Fijxu
  21. · Evidence links · GitHub search · MiniMax · pulls · inference-sim/inference-sim · @sriumcp
    Excerpt
    the substrate's layer 8 and 9 price changes (MiniMax-M3 routers replicated in fp32, MXFP8 scale bytes, dsv4-nvfp4 MTP experts at MXFP4, a shared lm_head for
  22. · Evidence links · GitHub search · MiniMax · pulls · inference-sim/blis-vllm · @sriumcp
    Excerpt
    (`vllm/model_executor/models/config.py:43-55`) for every graph that is not DeepSeek-V4; MiniMax-M3's pool at the same memory is 72645 blocks, not 72893, because the catalog graph now holds
  23. · Evidence links · GitHub search · MiniMax · pulls · inference-sim/blis-vllm · @sriumcp
    Excerpt
    the figures on the H200 rows and the MiniMax-M3 points are development-split, in-sample numbers.
  24. · Evidence links · GitHub search · MiniMax · pulls · goonercbot/hermes-agent · @goonercbot
    Excerpt
    - Real configured MiniMax-M3 request through corrected Hermes adapter returned exactly `RECOVERY_OK`, end_turn, original request unchanged.
  25. · Evidence links · GitHub search · MiniMax · pulls · ghuntley/how-to-build-a-coding-agent · @octo-patch
    Excerpt
    Anthropic-compatible defaults to the standalone workshop agents, with MiniMax-M3.1-Flash as the default and MiniMax-M3 as the alternative.
  26. · Evidence links · GitHub search · MiniMax · pulls · boelabs/ai-gateway · @jhisdev
    Excerpt
    - `bun run test` fails: `vercel.test.ts` pins `vercel/minimax/minimax-m3` reasoning as `fixed ["high"]`; the regenerated snapshot now says `openai_effort ["none","high"]`. Tests may not be edited by
  27. · Evidence links · GitHub search · MiniMax · pulls · ashhart/TensorFold · @ashhart
    Excerpt
    CUDA decode attention (`window_attention`, shared by Qwen3.5/3.8, Llama, MiniMax-M3, Spark 2.5, Step5 and LFM): `window_rows` merged every chunk's partial for every row, 128 chunks at
  28. · Evidence links · GitHub search · MiniMax · pulls · anomalyco/models.dev · @rjuez00
    Excerpt
    - `MiniMax-M3` is no longer listed and a request using that ID returns `model=MiniMax-M3.1-Flash`, so the old entry is removed rather than kept as an
  29. · Evidence links · GitHub search · MiniMax · pulls · ROCm/aiter · @zhiding512
    Excerpt
    On gfx1250, every MiniMax-M3 sparse attention call (decode, MTP verify, prefill) runs FlyDSL `pa_decode` (`pa_fwd_asm` does not support gfx1250), and dense layers use it with
  30. · Evidence links · GitHub search · MiniMax · pulls · ROCm/aiter · @RElbers
  31. · Evidence links · GitHub search · MiniMax · pulls · ROCm/ATOM · @yhl-amd
    Excerpt
    under `multi`, which is the prefill-side setup for MiniMax-M3 PD with UMBP:
  32. · Evidence links · GitHub search · MiniMax · pulls · ROCm/ATOM · @rbrugaro-amd
    Excerpt
    does for the target (vLLM already did this). MiniMax-M3's DSpark draft now validates this config instead of building its own, and declares `draft_skips_online_quant` so its
  33. · Evidence links · GitHub search · MiniMax · pulls · Quarkgluonmixture/quarkspace · @github-actions[bot]
    Excerpt
    - `intelligence` — minimax-m3 (default) intelligence: catalog 45.4, archive 18.9
  34. · Evidence links · GitHub search · MiniMax · pulls · Portkey-AI/models · @polease
  35. · Evidence links · GitHub search · MiniMax · pulls · Pinvou/pinvou-agent · @asto18089
  36. · Evidence links · GitHub search · MiniMax · pulls · Pb314314/vllm · @Pb314314
    Excerpt
    use sparse MLA (dense MLA, DeepSeek-V4 / V4.1, MiniMax-M3, non-MLA models) are not affected; the reason for each is in "Per-model check".
  37. · Evidence links · GitHub search · MiniMax · pulls · NVIDIA/TensorRT-LLM · @ziyixiong-nv
    Excerpt
    sampler, speculative and QSA tests; the MXFP8 and MiniMax-M3 dispatch tests; GSM8K with DSpark (TP4);
  38. · Evidence links · GitHub search · MiniMax · pulls · NVIDIA-NeMo/Megatron-Bridge · @pruprakash
    Excerpt
    seeing `TEDotProductAttention`; GREEN 4/4, full file 86 passed, Gemma4/MiniMax-M3/Kimi-K2.5-VL/GPT-OSS unit tests 100 passed. E2E: gpt-oss-20b card inference (TP2 PP2 EP2, `--attention-backend local`) on this branch
  39. · Evidence links · GitHub search · MiniMax · pulls · EvolvingLMMs-Lab/lmms-eval · @sahirema
    Excerpt
    sample by frame rate ignore it, e.g. vLLM's MiniMax-M3 and Qwen3-VL loaders. On those, `max_frames_num` has no effect: long clips are sampled at the loader's
  40. · Evidence links · GitHub search · MiniMax · pulls · overnightworks/speech · @dependabot[bot]
  41. · Evidence links · GitHub search · MiniMax · pulls · maximhq/bifrost · @jimaldon
    Excerpt
    model catalog datasheet already carries `tencent/deepseek-v4-pro`, `tencent/deepseek-v4-flash` and `tencent/minimax-m3`; live-only models are listed with blank pricing and can be priced via overrides.
  42. · Evidence links · GitHub search · MiniMax · pulls · maximhq/bifrost · @octo-patch
    Excerpt
    API. New text examples and integration defaults select `MiniMax-M3.1-Flash`, with `MiniMax-M3` as the other current choice. Default management model suggestions put those models first and
  43. · Evidence links · GitHub search · MiniMax · pulls · iv-org/invidious · @Fijxu
  44. · Evidence links · GitHub search · MiniMax · pulls · iv-org/invidious · @Fijxu
  45. · Evidence links · GitHub search · MiniMax · pulls · iv-org/invidious · @Fijxu
  46. · Evidence links · GitHub search · MiniMax · pulls · cbc02009/container-images · @mahoro-ai
    Excerpt
    `/v1/chat/completions` request to a messages-protocol Go model (e.g. `opencode_go/minimax-m3`, `qwen3.8-max`, `claude-haiku-5-5`) dies mid-stream with `TypeError: 'NoneType' object is not an iterator` → `APIConnectionError` 500 →
  47. · Evidence links · GitHub search · MiniMax · pulls · MrGeva/vllm · @MrGeva
    Excerpt
    320 questions, concurrency 32, same node, together with MiniMax-M3 model changes):
  48. · Catalogue listing · GitHub · MiniMax-AI/minimax-code · code
    Excerpt
    afety margin. */ export function computeCompactionTriggerAt(input: { modelId?: string; contextWindow: number; perTurnMaxTokens: number; reserveTokens: number; safetyMarginTokens: number; }): number { if ( input.modelId === 'MiniMax-M3' && (input.contextWindow === 512_000 || input.contextWindow === 1_000_000) ) { return Math.floor(input.contextWindow * 0.9); } const perTurnMaxTokens = Number.isFinite(input.perTurnMaxTokens) && input.perTurnMaxTokens > 0 ? input.perTurnMaxTokens : 0; const reserve = Math.max(input.reserveTokens, perTurnMaxTokens + input.safetyMarginTokens); return Math.max(1, input.contextWindow - reserve); }

    This code reference does not establish a working model identifier.

Specifications

MiniMax M3 is a model from MiniMax with open weights, released on 1 Jun 2026. The context window is 1,000,000 tokens.

Context window
1,000,000
Input
image · text · video
Output
text
Weights
Open

Questions

What is the context window of MiniMax M3?

The context window is 1,000,000 tokens.

Are the weights of MiniMax M3 publicly available?

Yes. MiniMax publishes the weights.

When was MiniMax M3 released?

On 1 Jun 2026.