Open-Weight Picks
The same question as the best-in-class snapshot, restricted to models you can download — where the licence matters more than the benchmark, and half the names in circulation are wrong.
Best in Class answers “what’s the best model for this job”. This page answers a narrower one: what’s the best model I can download, run, and ship a product on. Dated 23 September 2026, and stale the moment it’s published.
One rule governs everything below. With hosted models you read the pricing page; with open weights you read the licence, and the licence is the spec. Apache-2.0 and MIT mean do what you like. Anything else — non-commercial, source-available, research-only — is a constraint on your product, not a footnote.
| Task | Pick | Licence | Needs |
|---|---|---|---|
| Cut-outs, background removal | BiRefNet | MIT | Modest GPU; native up to 2048² |
| Segment / erase by description | SAM 3.1 (Meta) | See repo | GPU |
| Speech recognition | SenseVoice (Alibaba) | Apache-2.0 | CPU-viable |
| Speech synthesis | CosyVoice 3 (Alibaba) | Apache-2.0 | 0.5B — small |
| Music | ACE-Step 1.5 XL | MIT | <10 s/song on a 3090 |
| Image → 3D | TRELLIS.2 (Microsoft) | MIT | 24 GB VRAM |
| Image generation, quality | FLUX.2 [dev] | Non-commercial | 4090 at 4-bit |
| Image generation, commercial | Qwen-Image / FLUX.2 [klein] | Apache-2.0 | Single GPU |
| Video generation | Wan (Alibaba) | Apache-2.0 | 24 GB for the big tier |
| Document OCR | PaddleOCR-VL-1.6 | See repo | Runs on CPU |
| OCR, ultra-light | GLM-OCR (Z.ai) | See repo | 0.9B — edge-deployable |
| Coding | Qwen3.8-27B / GLM-5.3 | Apache-2.0 (Qwen) | 27B at 4-bit on one card |
The licence traps
FLUX.2 [dev] is the one people get wrong, and the distinction is worth stating precisely because it cuts the other way from what you’d assume:
The weights are under the FLUX [dev] Non-Commercial License — non-commercial, non-production use only. The images you generate with them may be used personally, scientifically and commercially.
So you can sell the picture; you cannot build the service. If you need the model inside a
product, Black Forest Labs’ own FLUX.2 [klein] (4B) is Apache-2.0, as
is Alibaba’s Qwen-Image — which is also the better pick when the image has to contain
legible text.
“MIT” on BiRefNet means the licence, not the institution. BiRefNet comes from Peng Zheng and co-authors (CAAI AIR'24), not from MIT the university. The confusion is everywhere in Korean-language roundups, and it matters only because it sends you looking for the wrong paper.
Names that don’t survive a check
The list circulating in video roundups has several transcription casualties. If you’re searching for these, search for the second column:
| Circulating as | Actually |
|---|---|
| “MetaSeg 3.1” | SAM 3.1 — Meta’s Segment Anything. Text prompts arrived in SAM 3 |
| “TRELLIS 2 (3B)” | TRELLIS.2, and it’s 4B (microsoft/TRELLIS.2-4B) |
| “GLM-OCR 0.9, a 9B model” | GLM-OCR is the 0.9B model — the 0.9 is the size |
| “Imigen” | No model by that name checks out. The Apache-2.0 slot it describes is filled by Qwen-Image or FLUX.2 [klein] |
Two claims worth tempering
SenseVoice is not straightforwardly the best open ASR for Korean. Apache-2.0 and genuinely fast — 50–100× real time — and it does support Korean. But its documented strength is Mandarin and Cantonese, and independent comparisons have Whisper Large-v3 ahead of it on Japanese and Korean. If Korean is your primary language, benchmark both on your own audio before committing. For Korean TTS, CosyVoice 3 is the safer recommendation: 0.5B, Apache-2.0, Korean among its nine languages, voice cloning from 3–10 seconds.
“Best open coding model” is contested, not settled. As of September the live open-weight composite has Qwen3.8 Max first and GLM-5.3 second, while Kimi K3 leads the frontend code arena. GLM-5.3 does lead on terminal and agentic benchmarks, which may well be the axis you care about — but the flat claim that it tops the coding board doesn’t hold. For running locally on one card, Qwen3.8-27B is the practical answer: dense, Apache-2.0, and the smallest footprint at the top of the self-hosting tables.
The two sizes worth internalising
Two numbers decide most of this list for a home machine:
- 24 GB VRAM is the gate. TRELLIS.2 and the large Wan tiers sit behind it; a 4090 or better is the entry ticket. Below that, quantisation and offload get you into the room — Wan drops to 6–8 GB at 480p with GGUF and CPU offload — at a cost in speed.
- Under 1B parameters is where the pleasant surprises are. GLM-OCR at 0.9B and CosyVoice 3 at 0.5B both compete with far larger models, and OCR is the one area where open models genuinely beat the paid ones — see PaddleOCR-VL at 96.3 on OmniDocBench, on a CPU.
For running any of these locally, start at llmfit to size the model to your hardware, and Ollama to serve it.
Next
Services worth running on your own hardware → Self-Hosted
Last updated 23 Sep 2026, 00:00 UTC.