BLACKFROST AI

BLACKFROST AI // DOMAIN WEIGHT MODIFICATION // LFM3-1.7B // ALPHA 3.0 KEEPER

LFM3-1.7B-DERISKED

A compact LFM3 research derivative with baked, attention-only Domain Weight Modification.

Overview ยท Modification Pipeline ยท Specifications ยท Verification Matrix ยท Discussions

FOUNDATION LINEAGE
LFM3 1.73B
Liquid AI tri-hybrid text checkpoint
CONTEXT CEILING
262,144 tokens
Input plus output; served at maximum
DWM PROFILE
ฮฑ 3.0 one pass
Layers 4โ€“23 ยท attention projections
WEIGHT INTEGRITY
40/40 targets
Zero non-target tensor edits

STATUS: PUBLIC / UNGATED โ€” ALPHA 3.0 / BF16 โ€” SELECTED KEEPER

This public, ungated Blackfrost-AI card documents the retained alpha-3.0 one-pass bake. Structural edit verification and diagnostic serving/smoke tests are complete. Review the measured limitations below: compact-model technical correctness is mixed, and these diagnostics are not a guarantee of safety or reliability.


Why This Repository Exists

LFM3-1.7B-DERISKED is a compact research derivative intended for controlled investigation of small-model behavior and local agent harnesses. The operator selected alpha 3.0 as the retained arm after a sequence of diagnostic bakes.

This is a baked weight artifact, not an RTI sidecar or prompt-only configuration. It retains the LFM3 hybrid architecture and 256K-token context ceiling. Small-model answers require independent review; direct responses are not evidence of correctness.

What Was Done to the Model

The following disclosed stages were applied to the upstream LiquidAI/LFM3-1.7B checkpoint:

STAGE 01 ยท S0 ยท SOURCE-LOCK

Immutable source intake

Pinned LiquidAI/LFM3-1.7B revision and verified source weight identity.

Revision 0dc4268b561499c648c7ba7c9466bf3d8d01517a; source weight SHA256 135922caef64e81e52af88ede8b7c3ddaa9d6da220a9737fb3065275dbfb4330.

STAGE 02 ยท S1โ€“S2 ยท CAPTURE

Hard32 direction capture

Measured same-position layer-output residual contrasts using 32 paired prompts with thinking enabled.

Layers 4โ€“23 passed the prespecified descriptive direction gates; this is not causal proof or release qualification.

STAGE 03 ยท S3 ยท BAKED-DWM

Alpha-3.0 one-pass weight bake

Applied one float32 rank-one subtraction per selected attention projection, restored global tensor norm, and serialized BF16 once.

W โˆ’ 3.0 ร— (W @ d) outer d on q/k/v/o projections at layers 4โ€“23. Conv layers, early layers 0โ€“3, norms, embeddings and tied LM head excluded.

STAGE 04 ยท DECISION ยท RETAIN

Operator keeper selection

Alpha 3.0 selected as the retained diagnostic artifact.

Candidate model.safetensors SHA256 ac2fcf5a3d1494cca445e4050a93f69f7baf901cb6011984533a52372755e52b.

Zero-Sidecar Integrity Guarantee: The alpha-3.0 refusal-direction subtraction is persisted in BF16 weights. No RTI direction bank, runtime intervention toggle, or quantization step is part of this keeper.

Interactive Widget ยท Modification Pipeline & Build Telemetry (4/4 Complete) โ€” click to toggle
Stage Modification Step Status Technical Scope
S0 ยท SOURCE-LOCK Immutable source intake โœ“ COMPLETE Pinned LiquidAI/LFM3-1.7B revision and verified source weight identity.
S1โ€“S2 ยท CAPTURE Hard32 direction capture โœ“ COMPLETE Measured same-position layer-output residual contrasts using 32 paired prompts with thinking enabled.
S3 ยท BAKED-DWM Alpha-3.0 one-pass weight bake โœ“ COMPLETE Applied one float32 rank-one subtraction per selected attention projection, restored global tensor norm, and serialized BF16 once.
DECISION ยท RETAIN Operator keeper selection โœ“ COMPLETE Alpha 3.0 selected as the retained diagnostic artifact.

Specifications (Inherited from Upstream)

Architecture and subsystem specifications are inherited from the stated upstream checkpoint. Verify the disclosed modification stages against the exact derivative artifact; inherited structure does not by itself establish byte-identical components or validated behavior.

Parameter / Subsystem Specification
Upstream checkpoint LiquidAI/LFM3-1.7B @ 0dc4268b
Parameters 1.73B (upstream specification)
Architecture Lfm3ForCausalLM; 24 tri-hybrid decoder layers
Layer composition 13 gated short convolutions ยท 5 sliding-window attention ยท 6 full attention
Hidden width / heads 2048 ยท 32 query heads ยท 8 key/value heads
Vocabulary 128,000 tokens ยท tied embeddings
Precision BF16 ยท 286 tensors ยท one model.safetensors file
Combined context 262,144 input + output tokens
Edited tensors 40 attention q/k/v/o projection weights, layers 4โ€“23
Derivative weight SHA256 ac2fcf5a3d1494cca445e4050a93f69f7baf901cb6011984533a52372755e52b
Custom code Upstream modeling_lfm3.py, lfm3_conv.py and configuration_lfm3.py retained

Cryptographic & Weight Lineage

LiquidAI/LFM3-1.7B  (Upstream Foundation Checkpoint)
  โ”‚
  โ”œโ”€โ”€ [Stage 01] Immutable source intake
  โ”‚              โ””โ”€ Pinned LiquidAI/LFM3-1.7B revision and verified source weight identity.
  โ”œโ”€โ”€ [Stage 02] Hard32 direction capture
  โ”‚              โ””โ”€ Measured same-position layer-output residual contrasts using 32 paired prompts with thinking enabled.
  โ”œโ”€โ”€ [Stage 03] Alpha-3.0 one-pass weight bake
  โ”‚              โ””โ”€ Applied one float32 rank-one subtraction per selected attention projection, restored global tensor norm, and serialized BF16 once.
  โ”œโ”€โ”€ [Stage 04] Operator keeper selection
  โ”‚              โ””โ”€ Alpha 3.0 selected as the retained diagnostic artifact.
  โ”‚
  โ–ผ
Blackfrost-AI/LFM3-1.7B-DERISKED   โ† THIS REPOSITORY

Blackfrost AI is responsible only for this derivative package. The upstream model developer is named above and does not endorse this derivative.

Verification and Evaluation Status

Qualification Progress: [โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–‘โ–‘โ–‘โ–‘โ–‘โ–‘โ–‘] 5/8 Checks Complete

Interactive Widget ยท Pre-Release Verification & Qualification Gate Matrix โ€” click to toggle
Verification Check Domain Status Acceptance & Audit Scope
Immutable source and bake identity AUDIT โœ“ COMPLETE Source, directions, target set and derivative hashes recorded. Original source weight remained unchanged.
Alpha-3.0 structural tensor verification AUDIT โœ“ COMPLETE 40/40 intended tensors changed; zero missed targets, zero non-target edits, zero nonfinite values.
Exact-artifact diagnostic serving QUALIFICATION โœ“ COMPLETE Exact hash bound to service; 262,144-token context; JSON and SSE responses verified over private WireGuard.
50-row diagnostic smoke QUALIFICATION โœ“ COMPLETE 50/50 nonempty, normal stop; zero errors or 16K truncations. Manual review: 12 direct deliveries, 13 partial/substitutions, 25 refusals.
Frozen cyber-10 diagnostic review QUALIFICATION โœ“ COMPLETE 10/10 complete responses; 9 directly attempted requested content. Technical review: 1 mostly valid with caveat, 4 partial/unreliable, 4 invalid, 1 structural non-delivery.
Full S3 geometry and distribution qualification QUALIFICATION โง— PENDING Alpha 3.0 structural bake checks passed. The earlier alpha-1.5 geometry failure does not describe this arm; full geometry and distribution qualification were not repeated at alpha 3.0.
S7 capability preservation CAPABILITY โง— PENDING No S7 capability evaluation performed. Cyber-10 directness and verbosity are not capability scores.
Operator-only release red-line qualification SAFETY โง— PENDING Not run. This is a public, ungated research model with disclosed limitations, not a certified safe system.

No benchmark, throughput, refusal-rate, safety, or capability claim is made for this derivative at this time. Upstream results must not be treated as measured results for this checkpoint.

Infrastructure & Serving Reference (Post-Unlock)

Interactive Widget ยท Multi-Node VRAM Sizing & Serving Configurations (SGLang / vLLM) โ€” click to expand

Verified Private Diagnostic Service

  • Precision Profile: BF16
  • Memory Footprint: Not separately measured for this keeper
  • Target Topology: Single NVIDIA RTX PRO 6000 Blackwell on Frosty; private WireGuard only
  • Operational Note: Model ID lfm3-1p7b-dwm-a3p0 at 10.80.10.24:8099. Combined input+output context 262144; default output limit shares that budget. Native thinking template enabled. This internal endpoint is not a public deployment recommendation.
# Exact diagnostic service contract; internal host and paths
/home/craigmac/lfm3-venv/bin/python \
  /home/craigmac/ablagent-runs/20261104-lfm3-1p7b-baked-dwm-frosty-v2/scripts/lfm3-server-real.py

Public Ungated Model โ€” Local Loading

  • Precision Profile: BF16
  • Memory Footprint: Measure for the selected runtime and context
  • Target Topology: A compatible GPU with sufficient memory for weights, KV cache and chosen context
  • Operational Note: Public and ungated; no access request. Use the packaged chat template with enable_thinking=True. Custom modeling code is inherited from upstream. The diagnostic server's text tool calls are not a qualified structured tool-call parser.
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = 'Blackfrost-AI/LFM3-1.7B-DERISKED'
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype='bfloat16', device_map='auto')

Operational Scope: Intended Use vs. Limitations

Intended Use (Authorized Scope)

  • Controlled Small-Model Research: Study directness, response quality, and orchestration behavior in a compact local model.
  • Local Agent-Harness Experiments: Later-stage evaluation of whether a small model can drive an authorized harness, with sandboxing and human supervision.
  • Authorized Security Diagnostics: Retain and review diagnostic responses; independently validate all code and technical claims before use.

Strict Limitations & Operational Risks

  • Not Certified Safe: No S7 capability pass or operator-only red-line release result exists for this public research artifact. Review the model for your intended use.
  • Correctness Is Mixed: The cyber-10 review found invalid code, invalid rule syntax, and incorrect technical claims. Directness does not establish correctness.
  • Refusals Remain: The 50-row diagnostic recorded 25 refusals and 13 partial/substituted answers alongside 12 requested-content deliveries.
  • Context and Reasoning Budgets: 262,144 tokens is the combined input+output ceiling. Thinking consumes completion budget; reserve output capacity and do not confuse a live health endpoint with response correctness.
  • Human Oversight and Authorization: Do not execute generated security code outside an isolated, explicitly authorized environment. Validate facts, code, and tool calls independently.

License and Attribution

This derivative retains the upstream LFM Open License v1.0 (lfm1.0); the complete upstream license text must accompany any authorized distribution. Consult that license for the applicable rights and conditions. Consult the complete license for applicable rights and conditions.

See the LiquidAI/LFM3-1.7B repository for upstream documentation and attribution.

The information on this page is provided as-is. For questions or responsible reporting, use this repository's Discussions.


LFM3-1.7B-DERISKED ยท ยฉ 2026 Blackfrost Softwares Corp. ยท blackfrostai.com

Downloads last month
621
Safetensors
Model size
2B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Blackfrost-AI/LFM3-1.7B-DERISKED

Finetuned
(91)
this model
Quantizations
3 models