The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Long-form audio understanding in September 2026: open-weights models are still ~18 points behind
An independent evaluation of Xiaomi's MiMo-V2.6-Flash-RL on hour-long audio, compared against published numbers for Gemini, Qwen3.5-Omni, Qwen3-Omni and other models. Includes per-question predictions and the serving recipe.
TL;DR
- MiMo-V2.6-Flash is the best open-weights long-audio model we know of. It scores 56.9% average accuracy on LongAudioSpan and hears every recording in full (up to 2.7 h). The previous best open result is Qwen3-Omni with thinking at 54.5%, on audio truncated to 80 minutes.
- It is still 15β18 points behind the closed frontier: Qwen3.5-Omni-Plus 74.6%, Gemini-3.1-Pro 72.3%.
- It is also behind Gemini 2.5 Pro (62.1%), a model from early 2025. On long audio, open weights are about 18 months behind the closed state of the art on this benchmark. The gap has widened: a year ago the best open model (Qwen3-Omni, 54.5) trailed Gemini 2.5 Pro by 7.6 points; today the best open model trails the best closed model by 17.7.
- On 60+ minute audio, MiMo (50.8%) scores below a text-only LLM reading an ASR transcript (53.4%). Hearing the audio currently gives an open model less than reading a transcript.
- "When was X said?" questions are close to chance: 28β39% at every duration, with 25% being chance.
- Thinking mode hurts on long audio. With thinking on, accuracy drops from 50.8% to 41.7% on 60+ minute audio, and 22% of those answers never finish (reasoning runs into the 16K token limit).
Why this matters
We are building a private, self-hosted pipeline for our own recordings. Voice memos, clip-on and room mics, and conversations that run for an hour or more land on one timeline with speakers, transcripts, and to-dos and calendar items extracted from them. Privacy is the point, so sending the audio to a hosted API is not an option. The question was simple: can an open-weights model do what Gemini does on 60β90 minutes of audio? As of September 2026, not yet.
Setup
| Benchmark | LongAudioSpan (arXiv 2608.26431), Accuracy mode: 1,080 four-option MCQs over 360 in-the-wild recordings, 50% English / 50% Chinese |
| Tiers | S 10β30 min (avg β16) Β· M 30β60 min (avg β42) Β· L 60+ min (avg β90, max 2.7 h) |
| Model | XiaomiMiMo/MiMo-V2.6-Flash-RL: 309B total / 15B active MoE, native audio encoder, 1M context, official MXFP4/FP8 checkpoint |
| Serving | vLLM (vllm/vllm-openai:mimo-v26-x86_64-cu130), tensor parallel 4 on 4Γ RTX PRO 6000 Blackwell |
| Audio | Full recording, never truncated. MiMo uses 6.25 audio tokens per second, so 1 h β 22.5K tokens and the longest file is 61.5K tokens |
| Decoding | Thinking off: temperature 0, max 512 tokens. Thinking on: temperature 0.6, top-p 0.95, max 16,384 tokens |
| Prompt | Audio, then question and options, then "Respond with only the letter (A, B, C or D) of the correct option." |
| Scoring | Option letter extracted from the reply (also accepts "the answer is X" and "ηζ‘ζ―X"). Replies with no option letter count as wrong, as in the paper |
Rubric (open-ended) and Chain (anchor) modes were not run. Rubric needs the paper's GPT-5.4 judge ensemble.
Results: Accuracy (%)
Numbers for all models except MiMo are from the LongAudioSpan paper (Table 3). "Input limit" is the audio cutoff each model ran with: open models in the paper were truncated to that length, so their M and L scores come partly from incomplete audio.
| Model | Open | Think | Input limit | S | M | L | Avg |
|---|---|---|---|---|---|---|---|
| GPT-5.4, question only (no audio) | β | β | β | 41.1 | 41.1 | 40.6 | 40.9 |
| GPT-5.4, reading ASR transcript | β | β | β | 64.8 | 60.7 | 53.4 | 59.7 |
| GPT-5.4, reading dense caption | β | β | β | 83.2 | 80.0 | 77.3 | 80.2 |
| Phi-4-Multimodal (5.6B) | β | β | 30 min | 33.7 | 34.6 | 34.2 | 34.2 |
| Voxtral-Small (24B) | β | β | 40 min | 42.5 | 40.0 | 31.9 | 38.1 |
| Audio Flamingo Next (8B) | β | β | 30 min | 43.3 | 38.9 | 35.0 | 39.1 |
| Voxtral-Mini (3B) | β | β | 40 min | 45.3 | 35.8 | 37.2 | 39.4 |
| Audio Flamingo Next (8B) | β | β | 30 min | 46.1 | 43.6 | 37.2 | 42.3 |
| MOSS-Audio (8B) | β | β | 50 min | 45.0 | 43.1 | 35.6 | 41.2 |
| Baichuan-Omni-1.5 (11B) | β | β | 80 min | 46.9 | 46.9 | 36.4 | 43.4 |
| MOSS-Audio (8B) | β | β | 50 min | 52.2 | 52.8 | 38.6 | 47.9 |
| Qwen3-Omni (30B-A3B) | β | β | 80 min | 53.9 | 54.4 | 46.2 | 51.5 |
| MiMo-V2.6-Flash (this eval) | β | β | none | 63.5 | 51.1 | 41.7 | 51.8 |
| Qwen3-Omni (30B-A3B) | β | β | 80 min | 56.9 | 57.5 | 49.2 | 54.5 |
| MiMo-V2.6-Flash (this eval) | β | β | none | 63.3 | 56.7 | 50.8 | 56.9 |
| Gemini-2.5-Pro | β | β | none | 67.5 | 64.2 | 54.7 | 62.1 |
| Gemini-3-Flash | β | β | none | 65.6 | 69.2 | 64.2 | 66.3 |
| Qwen3.5-Omni-Flash | β | β | none | 73.9 | 71.4 | 58.9 | 68.1 |
| Gemini-3.1-Pro | β | β | none | 76.7 | 74.4 | 65.8 | 72.3 |
| Qwen3.5-Omni-Plus | β | β | none | 79.2 | 74.4 | 70.3 | 74.6 |
The MiMo thinking-on row covers 1,054/1,080 questions: all M and L, and 334/360 S. The job hit its time budget.
Reading the table
- Open vs closed gap: about 18 points on average and about 20 on 60+ minute audio. The best open score (MiMo, 56.9) trails Qwen3.5-Omni-Plus by 17.7 points. On the L tier it trails by 19.5 points (50.8 vs 70.3).
- The best open model of September 2026 does not match Gemini 2.5 Pro. Gemini 2.5 Pro dates from early 2025 and beats MiMo on every tier (S +4.2, M +7.5, L +3.9).
- On long audio, the open models are closer to "no audio" than to the closed leaders. A text model that sees only the question scores 40.6 on L. MiMo adds about 10 points over that, Qwen3.5-Omni-Plus about 30.
- Transcript beats ears. GPT-5.4 reading an ASR transcript (59.7 avg, 53.4 on L) outscores every open audio model, MiMo included. The paper's conclusion applies directly: the hard part is extracting the few relevant facts from a long, redundant audio stream, not reasoning about them.
- Being able to take the whole file is not the same as using it. MiMo is the only open model here that ingests every file untruncated, and it leads the open field on L. But it still loses 12.5 points from S to L. Gemini-3.1-Pro loses 10.9 and Qwen3.5-Omni-Plus 8.9.
Where MiMo-V2.6-Flash fails (thinking off, per-question data in predictions/)
By question type. Temporal questions are the problem:
| Dimension | Accuracy | n |
|---|---|---|
| interpretation | 73.8 | 103 |
| summarization | 70.6 | 102 |
| recognition | 65.3 | 95 |
| causal | 64.9 | 114 |
| logical | 61.7 | 120 |
| counting | 57.4 | 94 |
| relational | 57.1 | 126 |
| tracking | 49.0 | 155 |
| localization ("at what timeβ¦") | 32.2 | 171 |
Localization is near chance at every duration: S 27.8, M 30.0, L 38.6. Tracking (following an entity over time) collapses on long audio: S 58.2, M 54.0, L 34.0. The model gets the gist of a long recording but cannot say when something happened. That is a problem for anything timeline-shaped.
By language. English 61.7 vs Chinese 52.2. The gap is similar on every tier (S 69.4/57.2, M 60.6/52.8, L 55.0/46.7). This is somewhat surprising for a model from a Chinese lab. The paper found that language does not separate the strongest models, only the weaker ones.
By genre. Lectures (61.6) and commentary (61.1) are easiest. Drama/variety (51.9) and live streams (54.8), with many speakers, overlap, music and noise, are hardest.
Thinking mode makes long audio worse, not better. Most models in the paper gain from thinking. MiMo loses:
| Tier | Thinking off | Thinking on | Replies that ran into the 16K-token limit (thinking on) |
|---|---|---|---|
| S | 63.3 | 63.5 | 26 / 334 |
| M | 56.7 | 51.1 | 63 / 360 |
| L | 50.8 | 41.7 | 78 / 360 |
The damage is concentrated in Chinese long audio (L/zh: 46.7 β 29.4) and in causal and counting questions. Median latency per question goes from 21 s to 26 s, but the 90th percentile goes from 40 s to 272 s. Replies that never terminate are the main failure: the model reasons in circles until it runs out of budget.
A second, informal test: interview transcription with speaker labels
Separately, we ran MiMo-V2.6-Flash on six of our own bilingual (German/English) interview recordings, 1β44 minutes long, with strong regional accents. We used the exact prompt we use with Gemini 2.5 Pro, which handles these recordings very well: a timestamped transcript with named speakers, confidence markers on uncertain words, and tags for music and sounds.
| Gemini 2.5 Pro | MiMo-V2.6-Flash | |
|---|---|---|
| Follows the requested output format | yes | no. It invents its own format, sometimes with Chinese labels |
| Speaker turns found (one interview, interviewer lines) | 38 | 20: interviewer turns merged into the main speaker |
| Music and sound tags (six files) | present | effectively none |
| Confidence markers | yes | none |
| Proper nouns (place, band, project names) | correct from context | misheard as similar-sounding common words |
| Timestamps | turn starts | fine-grained, within about 1 s of Gemini's |
| Failure modes | β | one file hit a "so, so, soβ¦" repetition loop until max tokens |
Conclusion from our use case: MiMo's timestamps are good and it takes a 44-minute file in one pass. But speaker identification, turn-taking and context switching, the things that make a transcript useful, fail. Gemini 2.5 Pro handles all of them.
Practical takeaways
- To self-host long-audio understanding today, use MiMo-V2.6-Flash with thinking off. It is the strongest open option, and it fits 2.7 h of audio in one request. Budget 4Γ 96 GB GPUs for hour-long audio with concurrency.
- Don't rely on an audio LLM for "who said what, when". Put dedicated diarization (e.g. NVIDIA Nemotron-3-Diarization), speaker embeddings and ASR with word timestamps first. Then give the LLM the transcript and the audio for interpretation. The transcript baseline above shows the text path is currently stronger.
- For researchers: the gap is in perception over long context, meaning localization, tracking and speaker continuity, not in reasoning. The benchmark to watch is how far L-tier accuracy falls from S-tier accuracy.
Serving notes for MiMo-V2.6 audio in vLLM
The mimo-v26 vLLM image rejects long audio out of the box. What was needed (all in code/run.sh):
VLLM_MAX_AUDIO_DECODE_DURATION_S=14400,VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=1024andVLLM_MAX_AUDIO_DECODE_BYTES=2147483648. The defaults of 600 s, 25 MB and 256 MB reject anything over 10 minutes.- Keep
image> 0 in--limit-mm-per-prompt. MiMo'sget_mm_max_tokens_per_item()does not report audio, so with image=0 and video=0 vLLM disables the encoder cache and every audio request fails. - The encoder cache ends up equal to
--max-num-batched-tokens(8192), which is too small for a 42-min file. Raising batched tokens to 64K runs out of memory at startup on 2 GPUs. Instead, floormax_tokens_per_mm_itemincompute_mm_encoder_budget(one-line patch, envVLLM_ENCODER_BUDGET_MIN=65536). - 2Γ RTX PRO 6000 (TP2) serves short audio but runs out of memory encoding hour-long files under concurrency.
4Γ (TP4) with
--gpu-memory-utilization 0.80, 16K batched tokens, 16 sequences and 8 client workers ran all 2,134 requests with zero errors at about 40 answers/min (thinking off). - With
--reasoning-parser mimoand thinking on, the answer can end up entirely inreasoning, withcontentempty. Read both.
Files
predictions/accuracy_thinking_off.jsonlandpredictions/accuracy_thinking_on.jsonl: one row per question with tier, language, genre, audio length, level and dimension, the predicted and correct letters, token counts, finish reason and latency. Raw model text is not included, because it quotes the benchmark audio.code/:run.sh(HF Jobs entrypoint),eval.py(client) andrescore.py(scoring).run.shoptionally copies in an exact-QKV loader for TP2 from diffbot's MiMo-V2.6 serving recipe. It is not needed at TP4 and not included here.
Caveats. This is a single run with our own prompt and answer extraction, not the paper's harness, so compare within about Β±2β3 points. With 360 questions per tier, the 95% interval is about Β±5 points. LongAudioSpan's questions and scripts are CC BY-NC-SA 4.0, and so is this derived data. The benchmark audio is not redistributed here.
September 2026. Compute: Hugging Face Jobs, about $45 in total, including failed setup runs.
- Downloads last month
- 86