ENPT

Lumma-fev-0.6B

Last updated:

Open20–100ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity4.6/10

    Apache-2.0 weights, a PyPI package (lumma-fev 0.1.1) and a lumma-fev-serve server that speaks the TypeSafe /v1/systemone contract, so the TypeSafe SDK works by changing base_url. Still days old, self-host only, no hosted SLA.

  • Capability5.9/10

    First-class noul / choice / score (1–255 options or levels) on a 0.6B backbone, 8k-token state, English + 10 Indian languages. Author numbers are strong on classic classification sets but weak on typed decisions (0.30 vs 0.72–0.76 for peers); all UnverifiedClaim.

  • Adoption22/100

    ~55 HF likes and ~1.3k downloads in four days, 137-like / ~7.1k-view launch tweet, and a 0.1B sibling already on HF. Author-announced 4B/9B not public yet.

Vendor claims

Lumma-fev-0.6B is an open Jev-like decision model from FrontiersMind, fine-tuned on their own Lumma-0.6B-Base (pretrained from scratch). It is not TypeSafe and not a LoRA adapter: the rank-16 LoRA is merged into the shipped weights.

Pass a state (text, JSON object or array) plus typed questions; get a probability distribution per question in one forward pass with zero text generation. Options go in, probs come out. The author positions it for routing, triage, and moderation.

Weights: Hugging Face FrontiersMind/Lumma-fev-0.6b (Apache-2.0). Runtime: lumma-fev on PyPI. Announcement: @FrontiersMind (22 Sep 2026).

Specs

Attribute Value
Author FrontiersMind (independent)
Base model FrontiersMind/Lumma-0.6B-Base (pretrained from scratch by FrontiersMind)
Fine-tuning LoRA rank 16, merged into the weights
Parameters 649M
Weights bf16 backbone, fp32 pointer head (256 dims)
Context Up to 8,192 state tokens
Languages English + 10 Indian languages
License Apache-2.0
Launch
Status Open weights (Hugging Face)
Decision types noul (yes/no), choice (1–255 options), score (1–255 ordered levels)
Runtime pip install lumma-fev (0.1.1); lumma-fev[serve] adds the API server; transformers with trust_remote_code
Family Lumma-fev-0.1b (HF, 25 Sep 2026); 4B / 9B announced, not public yet

Architecture

Lumma-fev follows the Jev-like one-pass pattern. It is a causal transformer run prefill-only: each question row reads the state plus its own instructions and options, never the other questions, so one question cannot shift another’s answer. A 256-dim pointer head reads out a distribution over the options. No autoregressive decode means nothing to parse and nothing to hallucinate.

Three ways to run it, all from the HF card:

  • transformers — AutoModel.from_pretrained(..., trust_remote_code=True) then model.decide(state, questions)
  • lumma-fev package — lumma_fev.load(...) picks CUDA, MPS or CPU
  • lumma-fev-serve — exposes the TypeSafe POST /v1/systemone contract; the card shows the TypeSafe SDK (typesafe-sdk) working against it by changing base_url

Author benchmarks

All numbers below come from the HF card. They are author-reported → UnverifiedClaim, not ModelSystem.One measurements.

Benchmark Lumma-Fev-0.6B TypeSafe Jev 1.13.0 Laya Lumma-Fev-0.15B
Banking77 0.90 0.87 0.425 0.47
DAIR Emotion 0.89 0.48 0.595 0.68
AG News 0.85 0.91 0.95 0.89
Typed-decisions 0.30 0.72 0.76 0.20
Average 0.74 0.75 0.68 0.56

The honest read: strong on classic single-label classification sets, weak on Typed-decisions (0.30 vs 0.72 for Jev and 0.76 for Laya). The near-tie on the average hides that gap. For reference, independent DecisionEval measured Jev 1.13.0 at 0.740 on typed decisions, consistent with the card’s 0.72 for Jev.

Limits

  • Benchmarks are author-reported — the card’s table is UnverifiedClaim; no independent run of Lumma-fev yet.
  • Weak on typed decisions — 0.30 on the author’s own Typed-decisions set, far below Jev and Laya; the 0.74 average leans on Banking77 / DAIR Emotion / AG News.
  • 0.6B backbone — expect held-out domains and complex reasoning to underperform larger peers like JevK5 (4B) or Decider (2B).
  • Calibration is on you — the card says confidence and probabilities are the model’s own estimates; measure calibration on your labelled data before gating automated actions.
  • Very new — lumma-fev 0.1.1 and the weights are days old; APIs may move.
  • Naming mismatch — the card’s table calls the small sibling “Lumma-Fev-0.15B”, but the HF repo is Lumma-fev-0.1b.
  • 4B / 9B are a vendor claim — the card says they release on 26 September with a blog; as of 26 Sep 2026 ~09:00 BRT they were not public on HF.
  • No hosted SLA — self-host only.

Fit / anti-fit

Fit when you need: fast typed decisions on constrained hardware, routing/triage/moderation at the edge, intent or topic classification, a local drop-in for code already written against the TypeSafe /v1/systemone API, English or Indian-language inputs.

Anti-fit when you need: strong multi-question typed decisions (see the Typed-decisions score), deep reasoning, knowledge-heavy tasks, hosted inference, or a production SLA.

Why it is in the catalog

Trained decision weights with open typed decision I/O and a server that speaks the TypeSafe contract; fills the sub-1B routing-focused slot next to Tiny-Jev and NanoJev. With a 0.1B sibling already out and 4B/9B announced, it is becoming a family.