Konic Labs

compact, task-specific LLMs for production AI — measurable behavior, controlled inference

[Konic Labs @koniclabs]

@liquidai's LFM2.5-VL-3B, now native on Apple Silicon

The INT4/INT8 weights map to MLX bit-exactly (max diff 0.0) — then we chased the remaining bytes.

- Decode 12.16 → 62–68 tok/s (5.1–5.6×) · text TTFT 167 → 62 ms
- Checkpoint 2.79 → 1.63 GB (−42%) · peak RSS ~4.8 → ~1.9 GB (−60%)
- M3 decode is bandwidth-bound (~0.94 GB weights/token @ ~70 GB/s) — every win was a bandwidth cut, not a kernel trick

Go download it.
https://t.co/ISYH4tCUsK

[Konic Labs @koniclabs]

New research from Konic Labs: Surgical FFN width pruning + distillation recovery + mixed-precision quantization on @liquidai's LFM2.5-VL-3B.

- 6.25 GB → 5.30 GB (−15%) → 4.92 GB (−21%) → 1.93 GB packed / 2.22 GB (Konic Optimized), −69%

- Near-lossless vs base: PPL ratios 0.79–0.85 — at or below baseline

- Top-5 token agreement: 0.95–0.99; quantization adds just +0.02–0.06 nats

- Tool calling: 10/10 rounds on the smallest tier — even abstaining correctly where the base over-triggered

The hypothesis: an arch-searched hybrid conv+attention VLM still carries compressible redundancy — and it lives in FFN width, not depth. Joint-SwiGLU Wanda prunes, distillation-to-baseline LoRA recovers, hand-rolled GPTQ packs.

See the full research and models.

https://t.co/7gipvusodj

[Konic Labs @koniclabs]

Making the LFM-2.5 encoder multimodal.

@liquidai's LFM2.5 Encoder 230M + SigLIP2 = a compact multimodal encoder.

- 12.5k held-out MONET pairs: i→t R@1 0.1194 (BF16)
- GPTQ INT4 keeps ~91% — 370 MB, −59.9% vs original

Go read it.
https://t.co/F2TdTgp7hX

[Konic Labs @koniclabs]

Introducing "agentbox".

Want to run RL (GRPO/SFT) on real agent trajectories — not synthetic "think step by step" chains?

agentbox gives you:

🐳 Ephemeral Docker sandboxes, one per rollout
🛠️ Tool-calling LLM agents (OpenAI protocol)
✅ Pytest/shell verifiers for objective rewards
📦 ART-native export → directly into your training pipeline

> pip install agentbox-rl

https://t.co/RlWlqeqWGL

[Konic Labs @koniclabs]

its not a fun game to play.

its a war to put on extreme effort.

we will make that possible.

[Ege Sabanci @egesabanci]

@sudoingX great take and genuinely support it.

however, we need to make them even smaller. people should be aware that 128GB VRAM still a luxury for most of the world.

Need intelligence in the palm of our hands. @koniclabs is focusing exactly that.

[Konic Labs @koniclabs]

@JensenHuang @nvidia this is the only way 🫡

[Ege Sabanci @egesabanci]

time to go all in with aggresive REAP + quantization on it.

also RL post-training will gain the accuracy loss back and even outperform the base model.

we are doing everything to liberate open source frontier intelligence and make AI accessible for everyone at @koniclabs

[Ege Sabanci @egesabanci]

let's make the small MoEs great!

its one of our main focused models at @koniclabs, some releases are on the way from us soon.

[Konic Labs @koniclabs]

New research from Konic Labs: REAP + AWQ INT4 on @liquidai's LFM2.5-8B-A1B.

- 32 → 16 experts/layer across 22 MoE layers
- 16.94 GB → 8.57 GB → 2.79 GB packed
- 6.1× smaller on disk
- MATH500: 88.76% → 77.0% → 72.0%
- BFCLv3: 64.79% → 59.07% → 57.36%

REAP structurally prunes low-saliency experts; external AWQ then packs the surviving weights to W4A16_ASYM INT4.

See the full research and models.

https://t.co/s2jNOmoBlK

[Konic Labs @koniclabs]

that is the way. we have much to do.

[Ege Sabanci @egesabanci]

@jun_song next part of the game is optimizing the intelligence to a pocket size.

thats what @koniclabs is focusing on.

[Ege Sabanci @egesabanci]

we are gonna make MLX superior!

its crucial for open source AI community to work directly on their consumer hardware. fast.

at @koniclabs, we've ported @cerebras REAP implementation to native MLX.

@awnihannun 🫡

[Konic Labs @koniclabs]

We rebuilt AWQ INT4 quantization from scratch in pure PyTorch on Qwen3-8B.

- 4.0× smaller linear weights: 13.9 GB → 3.5 GB
- 1.034× FP16 perplexity on WikiText-2 (10.08 vs 9.75)
- 2.59e-4 avg reconstruction MSE across verified layers
- Zero re-quantization error on AWQ-scaled linears
- Streams safetensors — never loads the full model

Fully open-source, see the research.

https://t.co/o5Hjh6m0VQ

[Konic Labs @koniclabs]

We built REAP-MLX — Router-weighted Expert Activation Pruning on Apple Silicon — and proved it on @liquidai's LFM2.5-8B-A1B (32 experts, top-4 routing).

> 96.8% code gen at 25% compression
> 91.4% at 50% — 8.8pp above frequency-only
> −44% memory at 50%
> <0.4pp variance across 3 calibration runs
> Pure MLX, one pass, no retraining, no CUDA

Fully open-source, see the research.

https://t.co/bkCdxIKiOV