ENPT

KnowLine-35B-A3B Gen1

Last updated:

Open51–51ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity5.6/10

    Apache-2.0 bf16 weights (~67 GB; SGLang serves FP8 at load, 48 GB+ GPU), a /v1/systemone front end, INFERENCE.md with the exact settings, and a long card. Small team, no hosted API, no SLA, no third-party integration.

  • Capability8.1/10

    Choice, Noul, Score (up to 64 questions) from one forward pass. Full 62.92 on the official Decision Index 0.3.2 (#9; Jev 60.11), but public 69.49 (#2): a 6.6-point gap. ECE 0.04-0.05 and 51 ms median (one H20) are the author's own, no p95.

  • Adoption2/100

    Five days of traffic: 25 downloads on the weights, 0 likes, one community GGUF with 1,087 downloads, a 4B predecessor line on HF with under 70 downloads each. No post found, no integrations, no GitHub repo, no reported users.

Vendor claims

News covering this model

publicado (OK do Fabricio 17/10 BRT).

Read this first. KnowLine’s public score (69.49, #2) is much higher than its Full score (62.92, #9): the maintainers’ private tests, which weigh 80% of Full, pulled it down 6.6 points. We use Full for accuracy. The training loop was run end to end by an AI agent (data groups, decontamination, training, selection), which the card documents in detail.

KnowLine-35B-A3B Gen1 is a typed-decision model by PelaAI: Qwen3.5-35B-A3B (a mixture of experts with ~3B active parameters) fine-tuned with LoRA (rank 32, including the routed experts) and merged into bf16 weights. You send a state (text or chat) and up to 64 typed questions; each is answered from one forward pass as a probability for every option. It is the 35B sibling of the 4B KnowLine line (Gen1 to Gen4). The card says KnowLine is independent and “not affiliated with, endorsed by, or derived from TypeSafe or Jev”.

Specs

Attribute Value
Developer PelaAI (HF organisation; no GitHub repository found)
Base Qwen3.5-35B-A3B (Apache-2.0), 35B total / ~3B active
Weights Apache-2.0 bf16 (~67 GB); readout temperature T = 1.20 folded into the final norm; community GGUF
Types Choice, Noul (yes/no), Score on a rubric; up to 64 questions per request
Interface knowline_server.py front end: POST /v1/systemone, same as KnowLine-4B (existing Jev clients change the base URL)
Serving serve_knowline.sh: SGLang 0.5.21 converts to FP8 at load (~35 GB); the author ran one server on one 96 GB H20; plan for 48 GB+
Inputs Text; the base vision tower is unchanged (zero-shot images, no image training)
Languages English, Simplified and Traditional Chinese
Released 12 Oct 2026 (card updated 13 Oct)

Official board: Decision Index 0.3.2

Data generated 16 Oct 2026 19:00 UTC. Full = 0.2 public + 0.5 same-skills + 0.3 new-domains.

KnowLine-35B-A3B Gen1 Jev 1.13
Full 62.92 (#9) 60.11 (#18)
Public 69.49 (#2) 57.96 (#37)

The card compares its public score with board rows and notes that its own full score “is not available yet”, so it relies on the maintainers’ run for the private part. Not on JevBench or Artificial Analysis as of 17 Oct (we found no row).

What is vendor-run

  • 69.49 public is the author’s run with the official kit; the board shows the same public figure.
  • ECE 0.04-0.05, Brier 0.25-0.27: reconstructed with the board’s published method (the board does not say which 32 benchmarks it uses; the author checked three board models and was within ~0.015). On the author’s held-out sets ECE is 0.025.
  • 51 ms median (one request at a time on one H20; the 4B is ~31 ms): no p95 given.
  • KOF ’98 results (17-0-1 vs Jev 1.13) and the image score (71.2 against 67.8 for stock Qwen3.5-35B-A3B on the author’s vision-suite reconstruction) are the author’s harnesses, not official boards.
  • The data includes teacher-written synthetic tasks (LLM-labelled, not all human-checked), and Qwen3.5-35B-A3B base scores are part of the gain: the card attributes most of the +4.59 over its 4B to the larger base.

Fit / anti-fit

Fit: a self-hosted Apache-2.0 MoE decision model that is strong on the public suite, English and Chinese, with a Jev-shaped endpoint and a low active-parameter count.

Anti-fit: a hosted API, a team with support, or a result you can verify beyond the maintainers’ board run; also GPUs under 48 GB.

Limits

  • Knowledge-heavy reasoning is still the weakest area (52.5 on the card’s DI 0.3 area score); math has no step-by-step reasoning.
  • A small Simplified Chinese held-out drop (−0.8, p = 0.051) came with the second training round.
  • Needs much more GPU memory than the 4B line.

Why it is in the catalog

Public Apache-2.0 weights, a typed Choice / Noul / Score interface in the System One shape, and an official-board Full score above Jev’s.

Score working (17 Oct 2026)

  • Maturity = availability 8 × 0.30 + docs 8 × 0.25 + integrations 3 × 0.25 + support 2 × 0.20 = 2.40 + 2.00 + 0.75 + 0.40 = 5.55 → 5.6
  • Capability = types 10 × 0.30 + calibration 7 × 0.25 + latency 7 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.75 + 1.75 + 1.60 = 8.10 → 8.1
  • Adoption = engagement 3 × 0.35 + ecosystem 3 × 0.35 + production 0 × 0.30 = 1.05 + 1.05 + 0.00 = 2.10 → 2

Availability 8: Apache-2.0 bf16 plus a community GGUF, but a heavy serving setup. Docs 8: a long card, INFERENCE.md with the exact settings, a disclosures section. Integrations 3: /v1/systemone (precedent). Support 2: an organisation with a model line, no SLA.

Calibration 7: a published temperature and ECE method, vendor-run. Latency 7: 51 ms median, no p95. Accuracy 8: independent board result (Full 62.92, #9); the public-to-Full gap is the reason to keep it at 8 rather than rate it by the public score.

Engagement 3 and ecosystem 3: 25 downloads on the weights, 0 likes, one community GGUF (1,087). Production 0.

KnowLine-31B-Gen1 tops the Decision Index (29 Oct 2026)

publicado (OK do Fabricio 29/10 BRT).

PelaAI’s dense sibling, KnowLine-31B-Gen1 (Apache-2.0, a LoRA fine-tune of Gemma 4 31B with merged weights, published 23 Oct), entered the Decision Index 0.4.1 on 28 Oct (08:52 BRT) at #1 of 243: Full 71.86, public 70.13, ahead of Fastino GLiDE 1.2 (71.52). Board-measured ECE 0.032, median 204 ms, p95 1,822 ms per row. Unlike the 35B-A3B Gen1, its Full score is above its public score. KnowLine-35B-A3B Gen1.5 is on the same board at Full 65.77 (#27).

Card claims (UnverifiedClaim): JevBench / JevBench-hard 90.5 / 80.2, 74.9 on 34 new-domain decision tasks, zero-shot boxes (ScreenSpot 81.3), all vendor-run.

Scores unchanged (5.6 / 8.1 / 2): this page stays scored on KnowLine-35B-A3B Gen1, and its accuracy is already at the cap of 8. Whether the 31B should get its own page is a separate decision. See news.