AtomicChat/Ling-3.0-flash-VL-GGUF
How to Run Ling 3.0 Flash VL Locally
<p style="margin-top: 0; margin-bottom: 0;"> <em>Built from InclusionAI's original weights using Atomic Chat's existing Ling importance matrix. The <a href="https://huggingface.co/datasets/AtomicChat/calib-corpora">calibration corpora</a> behind our builds are public.</em> </p> <div style="display: flex; gap: 8px; align-items: center; margin-top: 10px; margin-bottom: 10px;"> <a href="https://atomic.chat/?utmsource=huggingface&utmmedium=referral&utmcampaign=hfling30flashvl&utmcontent=btnatomic"><img src="https://huggingface.co/AtomicChat/Ling-3.0-flash-VL-GGUF/resolve/main/btnatomic.png" width="162" alt="Atomic Chat"></a> <a href="https://discord.gg/8wGSsvmg4V"><img src="https://huggingface.co/AtomicChat/Ling-3.0-flash-VL-GGUF/resolve/main/btndiscord.png" width="119" alt="Discord"></a> <a href="https://github.com/AtomicBot-ai/Atomic-Chat"><img src="https://huggingface.co/AtomicChat/Ling-3.0-flash-VL-GGUF/resolve/main/btngithub.png" width="115" alt="GitHub"></a> </div> <ul style="margin: 0 0 12px 0;"> <li>Ling 3.0 Flash VL is InclusionAI's vision-language model for text and image inputs.</li> <li>Choose AD-Q4KM, AD-Q5KM, AD-Q6K, or AD-Q8_0. Image inputs use the shared F32 vision projector.</li> <li>These builds are an experimental preview. Use the accompanying runtime; full quality validation is still in progress.</li> </ul> <hr style="margin: 0 0 16px 0;">
Prepared from the original inclusionAI checkpoint, revision 869591498e8dbb41d4d96e3e2a5b428a2f70eb1e. These are the Atomic AD layouts, made on CPU using the existing calibration iMatrix. The full nine-variant text comparison and CPU speed measurements are complete; see FINAL_REPORT.md. This repository is an experimental working preview, not a completed benchmark release.
Uploads arrive progressively. Check [UPLOAD_STATUS.json](UPLOAD_STATUS.json) before downloading a variant. Each language variant requires all six GGUF shards in the same directory; select shard 00001 when loading.
Vision requires the shared F32 mmproj, an additional 1.74 GB. File size is not a RAM requirement estimate.
Runtime requirement
Use the accompanying private runtime patch and added source files, based on AtomicBot-ai/atomic-llama-cpp-turboquant@cd560939087c95b93a1f30a95603d6b079436952. Stock llama.cpp and the released Atomic Chat app have not been validated for these artifacts. See RUNTIME.md.
Checks completed
All four variants: 917 tensor types/shapes verified, six-shard integrity checked, complete SHA-256 manifests, and all 382 protected F32 tensor payloads unchanged from BF16. Text and a spatial image passed on all four. Q4/Q6/Q8 additionally passed the OCR and object-count smoke cases. These are functional smoke checks, not a comprehensive vision benchmark.
Full text quality comparisons use the historical held-out 92 × 4096 protocol and a fresh BF16 reference from this checkpoint. Full text results are available in FINAL_REPORT.md. Video and maximum context are not validated.
The existing iMatrix comes from AtomicChat/Ling-3.0-flash-GGUF@253738fe190c15f329001f263f355fc1562bbe7c, SHA-256 7d3c0ebe9eb235cc08e0b7c91886f5c422772ea95eebbf0f53b0974a1c040991. Its 573 entries match the new language tensor dimensions. Four routed experts have no observations in that matrix; uniform importance was used for those entries.
manifest.json and SHA256SUMS describe the complete intended set. Actual upload completion is recorded separately in UPLOAD_STATUS.json.
Completed measurements
Full report · KLD chart · Metrics CSV.
The pilot results are separate from the full 92-block comparison. Completion is not a comprehensive vision or agentic quality certification.