the AI bench
VERIFIED AUGUST 2026
All models

MODEL · MICROSOFT · 3.8B (MMDIT 48-BLOCK + FLUX.2 SEMANTIC VAE + MULTI-LAYER GPT-OSS TEXT FEATURES)

Microsoft Lens (3.8B, MIT) — WITHDRAWN

⚠️ **Withdrawn by Microsoft — you can no longer download this.** As of our August 3, 2026 check, all three Hugging Face repos (`microsoft/Lens`, `-Turbo`, `-Base`) return 401 to anonymous access, `github.com/microsoft/Lens` returns 404, and no repo matching "lens" appears among the 534 repos Microsoft currently lists on the Hub. We are keeping this page up rather than deleting it because we recommended Lens-Turbo as a planner pick from May 22 to August 3, and a reader who acted on that deserves to find out what happened instead of hitting a dead link. What it was: Microsoft's first foundational text-to-image model, a 3.8B MMDiT trunk paired with FLUX.2's semantic VAE and multi-layer features from a frozen GPT-OSS text model, shipped as a three-step ladder (`Lens-Base` 50-step supervised, `Lens` 20-step RL-tuned, `Lens-Turbo` 4-step distilled) under MIT. Note the pattern: it was uploaded and pulled once in mid-May before the May 25 public release, so this is the second withdrawal.

License: MIT · Context: Up to 1,440 × 1,440 native; aspect ratios 1:2 to 2:1 · Released: May 19, 2026 (paper arxiv 2605.21573; HF model card finalized May 18)

The decision in five lines

The call
Consider — runnable locally, family reference
Best for
Local evaluation and family reference
Runs on
23 hardware picks fit (cheapest: Intel Arc B580 12 GB · $249)
Watch out
If you already have the weights locally, the MIT licence you received them under still stands; a withdrawal does not retroactively revoke a licence already granted.
Evidence
Estimated · last verified August 2026

3.8B (MMDiT 48-block + FLUX.2 semantic VAE + multi-layer GPT-OSS text features)
PARAMETERS
IMAGE GEN
TYPE
Up
CONTEXT
~8 GB BF16 on A100/V100 fallback; MXFP4 native on Hopper+ for ~half
VRAM AT Q4

Where we recommend this

This model isn’t currently in an active planner slot. See the runner notes below if you’re running it anyway.

The call

⚠️ **Withdrawn by Microsoft — you can no longer download this.** As of our August 3, 2026 check, all three Hugging Face repos (`microsoft/Lens`, `-Turbo`, `-Base`) return 401 to anonymous access, `github.com/microsoft/Lens` returns 404, and no repo matching "lens" appears among the 534 repos Microsoft currently lists on the Hub. We are keeping this page up rather than deleting it because we recommended Lens-Turbo as a planner pick from May 22 to August 3, and a reader who acted on that deserves to find out what happened instead of hitting a dead link. What it was: Microsoft's first foundational text-to-image model, a 3.8B MMDiT trunk paired with FLUX.2's semantic VAE and multi-layer features from a frozen GPT-OSS text model, shipped as a three-step ladder (`Lens-Base` 50-step supervised, `Lens` 20-step RL-tuned, `Lens-Turbo` 4-step distilled) under MIT. Note the pattern: it was uploaded and pulled once in mid-May before the May 25 public release, so this is the second withdrawal.

When not to use: Any new work — it is not obtainable. If you already have the weights locally, the MIT licence you received them under still stands; a withdrawal does not retroactively revoke a licence already granted. For a current pick at this tier use **Z-Image-Turbo** (6B, Apache 2.0, 8-step, community-proven), which replaced Lens in our image.high band, or **HiDream-O1-Image** (8B, MIT) for the quality ceiling.

Runner notes

Nothing to run — the repos are gone. Kept here as a dated record. The wider lesson is the one this site keeps relearning: a permissive licence is worth very little if the weights stop being downloadable, and "open source" is a property of an artifact you actually hold, not of an announcement. This is the second time in 2026 a model we tracked became unavailable after we recommended it — Claude Fable 5 was switched off for 19 days in June under an export-control directive. Prefer picks with wide community mirroring; a model that exists in a dozen community quant repos is far harder to withdraw than one that lives only in a vendor org.

License
MIT
Released
May 19, 2026 (paper arxiv 2605.21573; HF model card finalized May 18)
Maker
Microsoft

Hardware that fits

Every hardware pick whose memory fits this model at the quant we recommend. Sorted cheapest-first — the top row is your best-value fit. Click through for the full buyer’s guide.

Next step

Find-by-model — see what hardware runs this