FreeVideo Runs 33B MiniMax H3 on 8GB VRAM: Benchmarks Show 10–30 Min per Clip

One‑Sentence Takeaway

FreeVideo proves a 33B model can start on 8GB VRAM, but 120 community runs show 10-second clips need 10–30 minutes on such cards — you save the API fee, you pay in time, SSD writes, and a community licence you must read.

What the Three Pieces Actually Do

People see "FreeVideo + MiniMax H3 + ComfyUI" and think three models. There is only one model — MiniMax H3 — the other two are the execution layer and the UI.

H3 (the engine) : 33B dense multimodal Transformer (plus frozen Qwen3-VL-32B text encoder), released 31 Jul 2024, weights open 3 Aug. Generates up to 15 s, 24 FPS, 32 kHz native stereo. Full weights ≈ 144 GB (DiT + text encoder + video/audio VAEs); third‑party estimates put minimum VRAM at 80 GB on 4 GPUs.

FreeVideo (the gearbox) : inference engine, not a model. Runs OpenVDN’s 8-step VDN-H3 (Video DeltaNet hybrid attention, arXiv:2609.20744). Open‑sourced 2 Oct 2024, Apache 2.0, iterated to v0.4.x in two weeks: quality tiers, Mac preview, embedded workflows, INT8 weights, Claude/Codex agent support, minute‑long video.

ComfyUI (the cockpit) : native H3 support merged same day (PR #15224). Two paths now: official native workflow or FreeVideo low‑VRAM path.

How 33B Fits in 8GB: Three Tricks, One Cost

Lighter engine : 8-step VDN-H3 uses far fewer sampling steps; Video DeltaNet’s "near‑frame exact softmax + bidirectional linear memory" hybrid caps VRAM growth on long sequences.

Weight streaming + async prefetch + tiled compute : community analysis splits model into ~50 Transformer blocks (~432.5 MB each), scheduled across VRAM / pinned host memory / disk — non‑resident layers read from disk on the fly.

Per‑GPU precision auto‑select : GeForce/RTX 30 → INT8; workstation/DC → FP8; auto‑detects attention kernels. Official claim: INT8 on RTX 3090 makes 5 s generation ~ 2.3× faster .

The cost : VRAM pressure moves to disk and RAM. Benchmarks confirm this becomes the real bottleneck.

Five‑Step Quick Start

Pick install route : launcher ( FreeVideo.exe Win, FreeVideo-Mac-arm64.dmg macOS, AppImage/ ./setup.sh Linux) auto‑installs ComfyUI, runtime, models, reuses existing model dirs, supports offline bundles. Or manual: git clone https://github.com/FlashML-org/FreeVideo.git into custom_nodes, restart ComfyUI (not browser refresh), open Workflow → Browse Templates → FreeVideo → FreeVideo-All-in-One, then finish in FreeVideo Settings.

Download all offline bundles : video model, text encoder, video/audio decoders, sampling cache (four packs); fresh ComfyUI needs "runtime" pack; "audio reference cache" optional. Missing any pack stalls generation.

Choose variant & input : FL2VA (frame‑conditioned): 0 img = text‑to‑video, 1 img = first/last frame, 2 img = locked first+last. Ref2VA (full reference): up to 9 images + 3 video + 3 audio clips (≤12 files). Start with FL2VA .

Dial down parameters first : four quality tiers (Light / Standard / Fine / Ultra) — pick Light , drop resolution to 960×544, duration to 5 s. Default output: 768p short side, 24 FPS, 5–15 s, native stereo. 2K not available locally — H3-Regenerate-2K and H3-Context-IR remain closed, API‑only.

Reproduce via "video‑carries-workflow" : since v0.2.3 drag generated video back onto canvas to restore prompt, seed, all params; v0.3.2 adds half‑res preview; community LoRAs plug in; v0.4.0 adds AI agent queue + timeline editing.

Mac users — read this : experimental, macOS 14+, Xcode CLI tools, Metal + MLX fused kernels. M5+ uses Metal 4 INT8 tensor units; older chips fall back to BF16 (slower). Only text, first frame, and second‑pass sampling supported (no last frame, no audio/video refs). Two gotchas: unsigned preview — verify shasum -a 256 against Release page, then "Open Anyway" in Privacy & Security; only tested on 24 GB unified M5.

What 120 Community Runs Actually Say

FreeVideo’s repo collects 120 self‑reported runs (end‑to‑end submit‑to-save time). Resolutions, durations, tiers, versions differ — raw comparison is meaningless. Author filtered to a single aperture : 10 s / 1344×768 / Light tier , used measured "time" column (not "optimized" estimates), dropped thermal‑throttled or GPU‑busy outliers, deduped same model+VRAM keeping fastest. 19 records → 13 distinct GPUs , versions 0.3.x–0.4.x, not a unified benchmark .

Three Takeaways from the Ladder

5.3× spread at same aperture : RTX 5080 16 GB top at 240 s (4.0 min) ; RTX 5060 8 GB bottom at 1279 s (21.3 min) . "8GB runs" writes the price tag right there.

VRAM isn’t the only variable : RTX 3070 Ti 8 GB hit 749 s, beating several 16 GB cards — it had 96 GB system RAM . Streaming weights lets host RAM backstop VRAM, confirming the next point.

Same card varies wildly : 5080 two runs 240 s / 312 s; 3090 two runs 377 s / 704 s (128 GB vs 64 GB RAM, different versions). Never treat a single run as a spec.

Four "off‑ladder" reference rows (excluded from ranking): M5 24 GB Ultra tier 10 s = 30509 s (8.5 h) — ~127× the ladder top. Mac vs PC split is tier + unified memory , not "fits or not".

Representative Cross‑Tier Records (self‑reported, versions 0.3.0–0.4.24)

RTX 4060 Ti 16 GB: Light 1920×1088 15 s → 3306 s (55 min); Ultra 1344×768 15 s → 9829 s (2.7 h) .

RTX 3070 Ti 8 GB: Ultra 10 s → 8501 s.

Three judgments more useful than "8GB runs" :

First latency driver is tier + resolution, not VRAM . Dropping Ultra → Light on same 4060 Ti cuts 15 s from 2.7 h to 55 min — bigger gain than swapping GPU.

SSD & RAM are the hidden bottlenecks . Footnotes: one machine model disk read ~60 MB/s, 1504 s total ≈ 900 s waiting on disk; another 80 MB/s, 1676 s total ≈ 570 s waiting; 16 GB RAM box "every step reads disk", 887 s total ≈ 460 s waiting. Adding RAM / moving models to NVMe often beats GPU upgrade — inevitable with weight streaming.

Mac is a separate curve; unified memory decides everything . Official M5 24 GB Light: 1344×768 10 s ≈ 42 min, 960×544 10 s ≈ 22 min, 512×512 1.6 s ≈ 3 min. Community M5 24 GB Ultra 10 s = 30509 s (8.5 h). M5 Max 128 GB Light 1664×960 3 s = 755 s (12.6 min). M4 Pro 24 GB Light 960×544 10 s = 3457 s (58 min). Matches earlier Mac buying series: on‑device generation depends on unified memory + bandwidth, not "fits" — 24 GB runs but only for short experiments; usable daily driver starts at 128 GB.

Two caveats: "optimized" column is config‑based estimate, not re‑run ; thermal throttling is silent killer — one 3070 laptop hit 89 °C, downclocked to 229 MHz avg, 5 s clip took 3821 s. Treat as build guides, not leaderboard.

Cost Accounting: What You Save vs What You Spend

Cloud list price (MiniMax CN API, JD Cloud docs cross‑checked): 768P ¥0.50/s, 2K ¥0.80/s, 480P ¥0.33/s. → 10 s 768P ≈ ¥5 , 100 clips ≈ ¥500 (linear projection).

Local marginal cost : 250 W system, ¥0.6/kWh, 17 min/clip → electricity ≈ ¥0.04/clip (negligible); hardware sunk; SSD wear from repeated reads.

Real cost is throughput : 8GB card 10 s clip 500–1600 s → ~ 2–6 clips/hour (proj.). Cloud parallelizes, delivers in seconds‑minutes unattended.

Decision matrix:

Need data sovereignty, unlimited rerolls, overnight batch, zero platform fees → local wins.

Need same‑day delivery, commercial volume, 2K native → API wins; ¥5/clip cheaper than your time.

"Free" saves API money, not time. That’s why v0.4.0 baked in AI‑agent queuing — local video is a nightly batch job, not an instant toy.

Four Traps Before You Build

"Free" ≠ "unrestricted" — dual licence : FreeVideo code = Apache 2.0. H3 weights = MiniMax H3 Community Licence: excludes US, EU, UK, Korea ; commercial use > $20M annual revenue requires separate written grant; commercial product must prominently mark "MiniMax H3"; distillation / synthetic‑data training restricted; generated videos are yours . Mainland China not excluded, but read terms yourself — not legal advice.

Local ≠ full H3 : open = H3-Base (768p audio/video). 2K regeneration + context understanding modules closed . Official ComfyUI full weights ~144 GB ≠ FreeVideo’s 8-step VDN-H3. Don’t judge local output by cloud 2K demos.

8GB = "minimum to launch", not "recommended" : README says 8GB VRAM + 16GB RAM. Community shows 6GB laptop doing 5.2 s clip (outlier) and 8GB box hitting 3800 s due to thermals. Same spec, 10× experience gap.

Don’t port GGUF habits : ComfyUI‑GGUF manual weight stitching exists, but H3 hidden dim 2688 not divisible by 256 → true K‑quant impossible; files labelled Q4_K_M often actually Q4_0/Q4_1 — pick by file size, not suffix . Text encoder Qwen3-VL-32B is the VRAM hog (quantized 14.6–27.1 GB). FreeVideo’s value is skipping that assembly + version alignment. (From third‑party tutorials & community analysis, not official README.)

Should Your Machine Spin Up? Seat Map

(See article image: five hardware buckets with recommended action.)

One‑Line Answer

FreeVideo putting 33B into 8GB is real; ComfyUI making it clickable and reproducible is real. But it trades ¥5/clip API fee for 10–30 minutes wait, a hammered SSD, and a community licence you must read. Don’t ask "can my card run it" — ask "am I willing to let it run all night". Yes → install tonight. No → pay the ¥5, keep your time for storyboards and QA.

Raw Data (Self‑Verify)

Community 120 runs (Chinese)

Community runs (English)

Official Mac timing notes

Project home & install

If WeChat links blocked, open in browser: github.com/FlashML-org/FreeVideo/tree/main/docs Ladder derived from above table filtered to "10 s / 1344×768 / Light", deduped same model+VRAM, not using official "optimized" estimates.

Sources: FlashML-org/FreeVideo GitHub README (CN/EN), Mac.zh-CN.md, community-results.md (120 self‑reported runs, non‑uniform, "optimized" = estimate); MiniMax H3 release info & Hugging Face model card; ComfyUI Wiki / comfy.org; MiniMax CN API pricing (JD Cloud + multiple price pages cross‑checked); Video DeltaNet paper arXiv:2609.20744; third‑party hardware & GGUF tutorials. "Projected" costs/throughput are linear estimates from public data. Licence terms per official text — not legal advice. No independent benchmark performed.