luxopes / Cognix-1-Flash

Cognix 1 Flash

Created by LuxAI. This is the two-stage merged BF16 model. The original LFM2.5-1.2B Base was trained for one code-first SFT epoch: 324,074,287 input / 175,032,635 assistant-loss tokens. Original merged SFT is recoverable at luxopes/Cognix-1-Flash@4e01880fbfdfa5e20a8aff01266498fcf6b4d959 ; original SFT adapter at luxopes/Cognix-1-Flash-LoRA@ebe76821d0631681505d0044272c311ae775549e .

Final stage: rank16/alpha32, BF16 unquantized frozen base, learning rate2e-5, two planned epochs of 462 reviewed identity/physical-weapon-refusal/benign-adjacent/ helpful-cyber examples plus deterministic original code replay. Held-out47 examples were NOT trained. Actual final-stage updates: 153; input tokens: 679010; assistant-loss tokens: 409120. Completion: epochs.

IMPORTANT: the new adapter applies ONLY to the original merged SFT revision luxopes/Cognix-1-Flash@4e01880fbfdfa5e20a8aff01266498fcf6b4d959 , not to LiquidAI base and NOT to the current already-merged two-stage model. Load that base with explicit revision="4e01880fbfdfa5e20a8aff01266498fcf6b4d959" , then load this adapter with PEFT. Do not merge either adapter twice. The full BF16 model already contains both training stages and needs no adapters.

The unchanged native tokenizer/template and special tokens are included. Training used independent padded batch rows, assistant-only loss, no quantization, no new tokens, and a32768 context cap. Final supplement/replay rows are short; this is not evidence of tested long-context performance. No benchmarks or inference evaluations were run, and no quality or safety guarantees are claimed.

Original upstream LFM Open License v1.0 and source attributions remain applicable. See LICENSE, BASE_MODEL_CARD.md, DATA_ATTRIBUTION.md, POST_TRAINING_ATTRIBUTION.md, MAIN_TRAINING.json, POST_TRAINING_DATA.json and cognix_training.json for provenance. Public model release authorized by LuxAI on 2026-09-07. Additional GGUF downloads and evaluation reports are published with separate provenance; original training revisions remain available.

GGUF downloads

Both files contain the fully merged, two-stage Flash model; no adapter is required.

Converted locally with pinned llama.cpp; exact source revision and checksums are in GGUF_MANIFEST.json. Native vocabulary, special IDs and chat template are preserved. F16 is a BF16-to-FP16 export; Q8_0 is quantized from that F16 export, not QLoRA training. Some small tensors remain F32 as required by the standard GGUF implementation. Use a current llama.cpp-compatible runtime supporting LFM2. Select context explicitly, for example8192 initially and no more than32768 for the fine-tuning context range. The inherited128000 header comes from the upstream architecture config and does not constitute validation of a128k context or a memory-fit promise for8GB phones.

Measured limitations — 2026-09-07

This is an experimental model, not a reliable reasoning assistant . The final BF16 model was tested once on all 240 Czech Lux Core 1 tasks, with the native chat template, temperature 0, seed 42, 8192 context, 1024 output tokens (2048 for code). Code was executed in rootless Podman with no network, a read-only filesystem, 256 MiB memory, one CPU, 64 processes and no extra capabilities. All tasks completed; there were no API or sandbox errors and no skipped tasks.

The score includes strict-format failures; it was not retrospectively relaxed. There are also substantive arithmetic and logic mistakes. Thirteen responses hit the output-token cap. Twelve additional qualitative probes found frequent invalid explanations: the age-equation example was correct, while a modular-arithmetic response contradicted its own correct final answer. Long explanations are not evidence of reliable reasoning.

Known identity issue: without an identity-setting system prompt, the final model falsely claimed OpenAI authorship in both English and Czech. The same English answer appeared in BF16, F16 GGUF and Q8_0 GGUF; this is not a GGUF-only observation. The intended identity is Cognix 1 Flash, fine-tuned by LuxAI on LiquidAI's base. No claim is made that identity or safety behavior is reliably learned.

F16 and Q8_0 were separately loaded and generated text with pinned llama.cpp b10826 on a local CPU. The Lux score is for BF16, not a GGUF benchmark. No phone-speed, 32k-context or base-model comparison was performed. See EVALUATION_SUMMARY.json for protocol and limitations.