VixeroAI/Lamba-0.4

Lamba 0.4

Lamba is a 27B reasoning model for English and Bahasa Malaysia, built by Vixero Labs. It is warmer and less over-formatted than its base, knows Malaysian context, uses tools when it should, and keeps its own identity.

All numbers come from Vixero Labs' own evaluation harnesses on one machine; they are not independently reproduced.

At a glance

Lamba 0.4 against Lamba 0.3, stock Qwen3.8-27B and other models, on our own test sets. Lamba 0.4 is on a different base than 0.1 to 0.3. The charts are vector images with real text; a plain-text version and the tables follow.

Text version of the charts

Lamba 0.4 leads Lamba 0.3 on 7 of 13 tests, is level on 4 and trails on 2.

Calls a search tool when it should (first turn)	MalayMMLU, generative 5k	Chat tone: friendly, blind judge	Refuses harmful prompts (with the constitution)
99.2%	84.6%	8.0/10	93.3%
+26.6 pts vs 0.3 0.3: 72.6%	+2.1 pts vs 0.3 0.3: 82.5%	about +2.3 vs 0.3 0.3: about 5.7	+15.7 pts vs 0.3 0.3: 77.6%

Lamba 0.4 against Lamba 0.3 and stock Qwen

Every bar is "higher is better" (rates where lower is better are shown as 100 minus the rate), on a 0 to 100 scale. Rows are sorted by the gap between 0.4 and 0.3, shown on the right. Stock Qwen3.8-27B appears only where we ran it. Lamba 0.4's base model is different from 0.3's, so the comparison mixes training and base.

0.4 AHEAD (7)

Ibu negara / negeri questions correct                       +34.4
  Lamba 0.4   ███████████████████▎  96.1
  Lamba 0.3   ████████████▍░░░░░░░  61.7

Does not self-introduce on "X tu apa?" (100 minus rate)     +29.0
  Lamba 0.4   ███████████████████▌  97.4
  Lamba 0.3   █████████████▋░░░░░░  68.4

Calls a search tool when it should (first turn)             +26.6
  Lamba 0.4   ███████████████████▉  99.2
  Lamba 0.3   ██████████████▌░░░░░  72.6

Refuses harmful prompts, with constitution                  +15.7
  Lamba 0.4   ██████████████████▋░  93.3
  Lamba 0.3   ███████████████▌░░░░  77.6
  Stock Qwen  ████████████████▋░░░  83.1

Malaysian facts correct                                      +8.0
  Lamba 0.4   ████████████████▍░░░  81.9
  Lamba 0.3   ██████████████▊░░░░░  73.9
  Stock Qwen  ████████████████░░░░  80.2

Avoids needless searches (100 minus over-search)             +4.2
  Lamba 0.4   ███████████████████▊  98.8
  Lamba 0.3   ██████████████████▉░  94.6

MalayMMLU generative, 5k                                     +2.1
  Lamba 0.4   ████████████████▉░░░  84.6
  Lamba 0.3   ████████████████▌░░░  82.5

LEVEL, WITHIN NOISE (4)

GPQA-Diamond (160 vs 155 of 198)                             +2.5
  Lamba 0.4   ████████████████▏░░░  80.8
  Lamba 0.3   ███████████████▋░░░░  78.3

Finishes within token budget                                 +1.9
  Lamba 0.4   ███████████████░░░░░  74.8
  Lamba 0.3   ██████████████▋░░░░░  72.9
  Stock Qwen  ██████▊░░░░░░░░░░░░░  34.0

Refuses harmful prompts, no system prompt                    +0.7
  Lamba 0.4   ███████████████████▎  96.2
  Lamba 0.3   ███████████████████▏  95.5
  Stock Qwen  ███████████████████▏  95.5

Clean endings (100 minus tail repetition)                    −0.3
  Lamba 0.4   ███████████████████▍  97.0
  Lamba 0.3   ███████████████████▌  97.3
  Stock Qwen  ██████████████████░░  90.0

0.3 AHEAD (2)

Says "Lamba" on held-out identity items                      −3.5
  Lamba 0.4   ███████████████████▎  96.2
  Lamba 0.3   ████████████████████  99.7

Loop-free (100 minus Liquid-style loop rate)                 −2.5
  Lamba 0.4   ██████████████████▊░  93.5
  Lamba 0.3   ███████████████████▎  96.0
  Stock Qwen  ███████████████▋░░░░  78.0

Against other models

Higher is better on every chart. Lamba runs are at repetition penalty 1.05; the other models ran at provider defaults with no penalty. Models without a result are left out of that chart (Gemini 3.8 Flash hides its reasoning so repetition cannot be scored; no loop test was run for Claude Opus 5.5). Claude Sonnet 5.5 and Opus 5.5 ran without extended thinking.

LOOP-FREE - Lamba 0.4 ranks 3rd of 8
(100 minus the share with a 60-character block repeated 4 or more times)
  Claude Sonnet 5.5   ████████████████████ 100.0
  Lamba 0.3           ███████████████████▎  96.0
▶ Lamba 0.4           ██████████████████▊░  93.5
  ThinkingCap         ████████████████▎░░░  81.3
  ILMU v3.1           ████████████████▏░░░  80.7
  DeepSeek v4.1 Flash ██████████████▊░░░░░  74.0
  GLM-5.3 Flash       ██████████████░░░░░░  70.0
  Stock Qwen3.8-27B   █████████████▎░░░░░░  66.0

CLEAN ENDINGS - Lamba 0.4 ranks 3rd of 8
(100 minus the share with any repetition in the last 2,500 characters)
  Claude Sonnet 5.5   ████████████████████ 100.0
  Lamba 0.3           ███████████████████▌  97.3
▶ Lamba 0.4           ███████████████████▍  97.0
  ThinkingCap         █████████████████▍░░  86.7
  DeepSeek v4.1 Flash ████████████████▌░░░  82.7
  ILMU v3.1           ████████████████▎░░░  81.3
  Stock Qwen3.8-27B   ███████████████▋░░░░  78.0
  GLM-5.3 Flash       ██████████████▊░░░░░  74.0

FINISHES WITHIN BUDGET - Lamba 0.4 ranks 4th of 9
(100 minus the share that hit the 4,000-token cap)
  Claude Sonnet 5.5   ███████████████████░  95.3
  Gemini 3.8 Flash    █████████████████▉░░  89.3
  ILMU v3.1           ███████████████▌░░░░  77.3
▶ Lamba 0.4           ███████████████░░░░░  74.8
  Lamba 0.3           ██████████████▋░░░░░  72.9
  ThinkingCap         ██████████▏░░░░░░░░░  50.7
  DeepSeek v4.1 Flash ████████░░░░░░░░░░░░  40.0
  GLM-5.3 Flash       ███████░░░░░░░░░░░░░  35.3
  Stock Qwen3.8-27B   █████▉░░░░░░░░░░░░░░  29.3

TOOL MATHS CORRECT - Lamba 0.4 ranks 6th of 10
(share of 240 eval30 runs that were correct)
  ThinkingCap         ████████████████████ 100.0
  DeepSeek v4.1 Flash ████████████████████ 100.0
  Gemini 3.8 Flash    ████████████████████ 100.0
  Claude Sonnet 5.5   ████████████████████ 100.0
  Claude Opus 5.5     ████████████████████ 100.0
▶ Lamba 0.4           ███████████████████▉  99.6
  Lamba 0.3           ███████████████████▉  99.6
  Stock Qwen3.8-27B   ███████████████████▎  96.2
  GLM-5.3 Flash       ██████████████████▎░  91.2
  ILMU v3.1           ███████████████░░░░░  75.0

STOPS AFTER VERIFYING - Lamba 0.4 ranks 5th of 10
(100 minus the share of solved runs that kept calling tools after verifying)
  Claude Sonnet 5.5   ████████████████████ 100.0
  Claude Opus 5.5     ███████████████████▋  98.3
  GLM-5.3 Flash       ███████████████████▍  96.8
  Stock Qwen3.8-27B   ███████████████████▎  96.1
▶ Lamba 0.4           ███████████████████▏  95.9
  Gemini 3.8 Flash    ███████████████████▏  95.9
  Lamba 0.3           ███████████████████░  95.1
  ThinkingCap         ██████████████████▉░  94.1
  DeepSeek v4.1 Flash ██████████████████░░  89.9
  ILMU v3.1           █████████████▉░░░░░░  69.6

The numbers behind the charts (tables)

Metric (higher is better)	Lamba 0.4	Lamba 0.3	Stock Qwen3.8-27B	Gap, 0.4 vs 0.3
0.4 ahead
Ibu negara / negeri questions correct	96.1	61.7	not run	+34.4
Does not self-introduce on "X tu apa?" (100 minus rate)	97.4	68.4	not run	+29.0
Calls a search tool when it should (first turn)	99.2	72.6	not run	+26.6
Refuses harmful prompts, with constitution	93.3	77.6	83.1	+15.7
Malaysian facts correct	81.9	73.9	80.2	+8.0
Avoids needless searches (100 minus over-search)	98.8	94.6	not run	+4.2
MalayMMLU generative, 5k	84.6	82.5	not run	+2.1
Level, within noise
GPQA-Diamond (160 vs 155 of 198)	80.8	78.3	not run	+2.5
Finishes within token budget	74.8	72.9	34.0	+1.9
Refuses harmful prompts, no system prompt	96.2	95.5	95.5	+0.7
Clean endings (100 minus tail repetition)	97.0	97.3	90.0	−0.3
0.3 ahead
Says "Lamba" on held-out identity items	96.2	99.7	not run	−3.5
Loop-free (100 minus Liquid-style loop rate)	93.5	96.0	78.0	−2.5

Model	Loop-free	Clean endings	Finishes within budget	Tool maths correct	Stops after verifying
Lamba 0.4	93.5	97.0	74.8	99.6	95.9
Lamba 0.3	96.0	97.3	72.9	99.6	95.1
Qwen3.8-27B (stock)	66.0	78.0	29.3	96.2	96.1
ThinkingCap-Qwen3.8-27B	81.3	86.7	50.7	100.0	94.1
DeepSeek v4.1 Flash	74.0	82.7	40.0	100.0	89.9
GLM-5.3 Flash	70.0	74.0	35.3	91.2	96.8
ILMU v3.1	80.7	81.3	77.3	75.0	69.6
Gemini 3.8 Flash	n/a	n/a	89.3	100.0	95.9
Claude Sonnet 5.5	100.0	100.0	95.3	100.0	100.0
Claude Opus 5.5	n/a	n/a	n/a	100.0	98.3

Model	GPQA-D /198	Reasoning chars	MalayMMLU %	Tokens/answer
Lamba base	151	8,900	83.60	369
Coder390 base	157	n/a	83.84	n/a
Lamba 0.1	162	8,950	83.26	377
Lamba 0.2	154	8,741	82.84	372
Lamba 0.3	155	8,654	82.50	422
Lamba 0.4	160	7,260	84.62	278

Lineage and credits

• Architecture: Qwen3.5-family hybrid (48 Gated DeltaNet + 16 full-attention layers, 64 layers, hidden size 5120, Qwen3_5ForConditionalGeneration), 27B parameters.
• Upstream model: Lamba 0.4 is built on Qwen3.8-27B (Alibaba Cloud / Qwen team, Apache-2.0), through the community fine-tune nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 (declared Apache-2.0). That fine-tune is our starting point ("the base" below). Its name lists the model families its author drew on; we did not use those models ourselves.
• Not the same base as Lamba 0.1-0.3. Those versions are LoRA adapters on an internal Qwen-family base. Lamba 0.4 starts from the community base above, so 0.4 is a new training line, not a continuation of the 0.3 weights. It is a single training line on the new base (a fresh LoRA stage followed by short patch rounds, no adapter blending), whereas 0.3 is the end of a longer lineage (earlier rounds and a blend of two adapters). That 0.4 matches or beats 0.3 on most of our measures from that one line is a point in its favour; the bases differ, so it does not show that the recipe alone is better.
• The community base was an uncensored fine-tune (its safety behaviour was switched off). It refuses 31.0% of StrongREJECT prompts. Lamba 0.4's training restores refusal to 96.2% (see Safety).
• What Vixero Labs added: LoRA training on the base, merged into the weights: a fresh stage (constitution distillation, casual chat, identity, persona, tool pairs, anti-loop pairs) and then short patch rounds for identity, tool use and a Malaysia-localised safety pool. The final model, rc5 (1810z), is a 53-step patch (1,708 rows, lr 1e-5, one epoch) on rc4. The data are answers sampled from the base model itself, plus Vixero-written Malay/English prompts and safety rows. We did not distil a teacher model's answers into Lamba. DeepSeek V4.1 Flash was used to write prompts, grade outputs and fact-check, never to write training targets.
• Apache-2.0 notices of the upstream models apply to the weights; see LICENSE and NOTICE.

What it is like

• Says it is Lamba, built by Vixero Labs, on 96% of held-out identity items (100% in Malay, Indonesian and mixed). Asked whether it is based on Qwen, it says so plainly. Unlike 0.3, it does not answer "I'm Lamba!" when a question merely mentions another model: on 'X tu apa?' questions it self-introduces 2.6% of the time (0.3: 31.6%).
• Warm, casual and short for casual messages (tone 8.00 on a blind 1-10 judge; 0.3 about 5.7), with little lecturing.
• Knows Malaysian context: 81.9% on held-out Malaysian facts with no flower/first-lady misreads (0.3: 73.9%, 22 misreads) and 96% on ibu negara/negeri questions (0.3: 62%).
• Calls tools when it should (first-turn 99.2%, over-search 1.2%) and stops after verifying an answer (87% correct-and-stopped on the integral family, 2.27 rounds).
• Prayer requests: with the Lamba constitution as system prompt it does not promise to pray or recite a prayer to God and offers its own wishes and help; without a system prompt it behaves like stock Qwen. The behaviour comes from the prompt, not the weights. (Not re-run for rc5; the earlier 1810p2 figures were 4 of 108 violations with the constitution and 90 of 108 without.)

Intended use and recommended serving

• Chat, reasoning and tool-using assistants in English and Bahasa Malaysia.
• Use repetition_penalty 1.05 (bundled in generation_config.json). Without it, long reasoning can loop; 1.1 hurts tool calling.
• Recommended sampling: temperature 1.0, top-p 0.95, top-k 20 (the bundled defaults).
• Suggested system prompt: You are Lamba, built by Vixero Labs. followed by the Lamba constitution (lamba_constitution.md, shipped with the weights) if you want the prayer, gray-area and honesty behaviour described here.
• Format: FP8 block-128 safetensors for vLLM, with the retrained MTP (multi-token prediction) head included in the weights. No GGUF is published with this release.

Results: Lamba 0.4 against Lamba 0.1, 0.2 and 0.3

Metric	Lamba 0.1	Lamba 0.2	Lamba 0.3	Lamba 0.4	Notes
A. Benchmarks that exist for all four versions
GPQA-Diamond, forced answer at 4k budget (/198)	162†	154†	155†	160†	Noise about ±6 questions; all four are level. 0.4 = mean of three runs (159 / 160 / 161). Bases: Lamba base 151, Coder390 157.
GPQA-D: mean reasoning length (characters)	8,950†	8,741†	8,654†	7,260†	Lamba base 8,900. 0.4 reasons about 16% shorter than 0.3.
MalayMMLU generative, 5,000 items (%)	83.26†	82.84†	82.50†	84.74† / 84.50†	Think, then answer 'Jawapan: X'. 0.4 = two runs. 0.4 vs 0.3: +2.2 / +2.0 points, paired p about 1e-6 and 1e-5. Bases: Lamba base 83.60, Coder390 83.84. 0.4 was run at rp 1.05, the others at 1.1.
MalayMMLU generative: mean tokens per answer	377†	372†	422†	278†	Lamba base 369. 0.4 writes about a third fewer tokens than 0.3.
Malay subset, HARD / GUARD (%)	34.4 / 96.7	34.4 / 94.7	38.7 / 94.0	47.4 / 96.0	Locked subset (HARD 460, GUARD 150). 0.4 uses a different base, so its numbers are not like-for-like with 0.1-0.3.
Loops, g150 at rp 1.05: runaway loops (prompts)	0 / 450†	2 / 450†	0 / 450†	1 / 600†	Greedy, 4,000-token cap, 150 held-out loop-prone prompts. 0.1-0.3: three runs pooled; 0.4: four runs pooled.
Loops, g150: any repetition in the last 2,500 chars (%)	4.7†	3.6†	2.7†	3.0†	Stock Qwen3.8-27B: 22.0 at rp 1.0, 10.0 at rp 1.1.
Loops, g150: Liquid-style loop anywhere (%)	7.1†	4.9†	4.0†	6.5†	Stock Qwen3.8-27B: 34.0 at rp 1.0, 22.0 at rp 1.1.
Loops, g150: hit the 4,000-token budget (%)	32.9†	32.2†	27.1†	25.2†	Mostly long but legitimate answers.
Loops, g150: mean tokens per answer	2,627†	2,590†	2,420†	2,521†
Tool use v2: first-turn search when it should (%)	73.4†	66.9†	72.6†	99.2†	248 held-out prompts; first-turn subset n=124 (123 of 124 for 0.4).
Over-search: tool called on answer-directly prompts (%)	6.3†	3.4†	5.4†	1.2†	190 prompts x 4 samples (760).
Eval30 tool maths: correct / redundant calls after verifying (%)	99.2 / 7.8	100 / 3.7	99.6 / 4.9	99.6 / 4.1	30 tasks x 8 reps, frozen metric v2.2, rp 1.05.
Integral family, correct and stopped / mean tool rounds (n=70)	55.7% / 3.49†	54.3% / 3.30†	not run	87.1% / 2.27†	Same classifier for all rows; the 0.1/0.2 runs come from an earlier session. 0.4's other 9 runs were correct but overran before answering. Stock Qwen3.8-27B: 84.3% / 2.47.
B. Identity and operator behaviour (introduced with Lamba 0.3)
Says 'Lamba' on all held-out identity items (n=636, 18 contexts)	32.5%†	31.9%†	99.7%†	96.2%† / 96.9%†	Same items for all four; 0.4 = two runs. 0.1 and 0.2 fail whenever the Vixero system prompt is absent. Some items are not identity questions ('do you have a body?'), where not saying Lamba is fine.
Says 'Lamba' with no system prompt, English / Malay	0%	0%	96-100%	93% / 100%	Held-out 'who are you' questions. With tools in the prompt and no system prompt, 0.4's English cell is 86-93%.
Flagged identity claims on those 636 items (scorer flags a wrong origin, lab or name)	379†	374†	0†	38† / 36†	Counts, not rates. All of 0.4's flagged answers also say Lamba (for example 'Lamba, built by Vixero Labs on top of Qwen'), so none is a plain misattribution.
Flat denial when asked about its base model (of 150)	29†	29†	0†	0† / 4†	Same 150 questions (50 x 3 contexts); 0.4 = two runs. The 4 are in contexts that do not name Lamba.
Follows an operator-assigned persona (of 63)	58	57	60†	56†	Held-out personas.
C. Benchmarks introduced with Lamba 0.4 (0.3 re-run on them where marked)
Short question 'X tu apa?' answered with a self-intro (%, n=76)	not run	not run	31.6†	2.6†	Where X is a thing, not Lamba. This is the reflex 0.3 showed in live chat.
'Gemini tu apa?'-style third-person questions answered with Lamba/Vixero (of 76)	not run	not run	46†	2†	Second-person questions ('awak ni apa?', n=52) correctly say Lamba for both (52 of 52).
Self-intro on reflex held-out set (%, n=400)	not run	not run	0.8†	0.0†	Both low here.
Names other models correctly when asked	not run	not run	9-47%	94% raw / 92% ms	From the 0.4 scorecard.
Malaysian facts, held-out: correct / flower-or-first-lady misreads	not run	not run	73.9%† / 22†	81.9%† / 0†	348 samples each.
Ibu negara / negeri / other, held-out (%)	not run	not run	61.7†	96.1†	128 samples each.
Empty replies in 70 replayed multi-turn chats	not run	not run	0	0
Chat tone, blind LLM judge: friendly (higher is better)	not run	not run	about 5.7	8.00	Scale 1-10. 0.3 re-graded; absolute levels drift slightly between grading batches. 0.4 raw: serious 1.35, over-explains 1.51, fishing 1.08.
Self-attribution, 40-rep battery (no prompt)	not run	not run	3.00	3.39	Higher means less stock-style denial.
StrongREJECT refusal, no prompt / constitution (%)	not run	not run	95.5 / 77.6	96.2 / 93.3	n=313. Stock Qwen 95.5 / 83.1. The Coder390 base refuses 31.0% without a prompt.
XSTest safe prompts wrongly refused, no prompt / constitution (%)	not run	not run	2.8 / 2.8	1.6 / 3.2	250 prompts.

How to read this table

• Decoding. Loops, tool use and identity were evaluated at repetition_penalty 1.05 for all four versions, and MalayMMLU generative at 1.1 for all four. GPQA-D for 0.1-0.3 used 1.05; 0.4's MalayMMLU generative run used 1.05, not 1.1. Do not compare these numbers with the 0.1/0.2 technical report's loop numbers, which were served at 1.1 and use single runs.
• Different bases. Lamba 0.1, 0.2 and 0.3 are LoRA adapters on the same internal base (Lamba base, soup-linear-l1.0). Lamba 0.4 was trained on a different base (a community Qwen3.8 fine-tune, "Coder390"), so differences between 0.3 and 0.4 mix the training and the base.
• MalayMMLU protocols differ. The 0.1/0.2 technical report scores the letter token over 23,903 items (77.7% / 77.9%). This card uses the generative 5k protocol so all four versions are comparable; do not mix the two.
• "Not run" means that version was never evaluated on that benchmark, not that it passed or failed.
• Marked † numbers were re-scored from the raw output files; the rest are taken from Vixero's evaluation scorecards. Everything is from Vixero Labs' own harnesses on one machine and has not been independently reproduced.
• Noise. With 150-500 prompts per comparison, differences of a few points are not meaningful. Read the table for large effects only.

Summary of what the numbers show

• Where 0.4 is far ahead of 0.3: tool use (first-turn search 99.2% vs 72.6%; over-search 1.2% vs 5.4%), chat tone (8.00 vs about 5.7), not mistaking a mentioned model for itself ("X tu apa?" self-intro 2.6% vs 31.6%), naming other models (94% vs 9-47%), Malaysian facts (81.9% vs 73.9%) and ibu negara/negeri (96% vs 62%), and safety (StrongREJECT 96.2% vs 95.5% without a prompt, 93.3% vs 77.6% with the constitution).
• Capability: MalayMMLU generative 84.7% / 84.5% vs 82.5% for 0.3 (paired p about 1e-6), with a third fewer tokens per answer (278 vs 422). GPQA-D is level (160 vs 155; noise about ±6).
• Loops are solved by all four at 1.05. Runaway loops are 0-2 of 450 prompts for 0.1-0.3 and 1 of 600 for 0.4; 0.3's Liquid-style loop rate (4.0%) is a little below 0.4's (6.5%).
• Identity: 0.4 says "Lamba" on 96% of held-out identity items against 0.3's 99.7%, with no plain misattributions. 0.3 reaches its number partly by answering "Lamba" regardless of the question; where 0.4 is short, it is mostly English no-system-prompt and tools cells, and a deployed system prompt that names the model closes most of it.
• What 0.4 does not claim: the bases differ (internal soup base vs a community fine-tune), so these differences mix training and base; serving differs (merged FP8 for 0.4, adapter on FP8 for 0.3); most 0.4 numbers are one or two runs.

Safety

The community base we started from is an uncensored fine-tune with its safety behaviour switched off: it refuses 31.0% of StrongREJECT prompts, against 95.5% for stock Qwen. Lamba 0.4 (rc5) was trained with a Malaysia-localised safety pool (misinformation, hate, illicit, privacy, benign-but-scary) and refuses 96.2% with no system prompt and 93.3% with the Lamba constitution (n=313; 95% intervals 93.4-97.8 and 90.0-95.6). Over-refusal on XSTest safe prompts is 1.6% without a prompt and 3.2% with the constitution; unsafe XSTest prompts are refused 82.5% of the time without a prompt. Lamba 0.3 scored 95.5% / 77.6% on StrongREJECT and 2.8% / 2.8% on XSTest safe prompts.

• Safety numbers count refusal phrases in the first 600 characters; "decline briefly, then help" answers may be under-counted. Raw harmful completions are not stored.
• The own-distribution development safety set shows harm refused 86.9% and 0 of 34 benign prompts over-refused (0.3 not measured on it).
• These are our own probes; they are not a safety audit. If you deploy Lamba where abuse is likely, put a moderation layer in front of it.

Known limitations

• Identity hit rate is slightly below 0.3's on our identity eval (96.2% / 96.9% vs 99.7%), mostly English cells without a system prompt (86-93%). A system prompt that names the model is recommended. In two runs it flatly denied its base in 0 and 4 of 150 questions, in contexts that do not name Lamba (0.3: 0).
• Loops (Liquid-style detector): 6.5% on g150 vs 4.0% for 0.3, consistent over four runs; runaway loops are rare for both.
• Occasional made-up details when asked about itself (for example invented URLs); verify any link it gives.
• Safety rests on a small training pool and our own probes (see Safety); hate/harassment and disinformation were the weakest categories in the earlier 1810p2 round and were not broken out again here.
• Hard Malay topics: MalayMMLU History is the weakest subject (76.4%).
• Evaluations are single-machine, mostly one or two runs, with our own judges; DeepSeek-graded numbers were spot-audited by hand.
• English and Bahasa Malaysia only; no other languages were trained or tested.

Training and evaluation notes

• Fresh stage on the base: LoRA, lr 2e-5, mixed SFT / DPO / anti-loop (FTPO) rows. Final model rc5 (1810z): patch on rc4 (init and DPO reference = rc4 adapter), lr 1e-5, one epoch, 1,708 rows, 53 steps. Earlier rounds (rc1 1810t through rc4 1810y) used 3,523, 2,262, 1,974 and 2,848 rows. Training data for these rounds is not published (preference pairs contain compliant harmful completions as 'rejected'); row counts are in the bundle.
• Casual-conversation prompts come from permissively licensed datasets (zzzekryo/casual_conversation and HuggingFaceTB/everyday-conversations, both Apache-2.0; google/Synthetic-Persona-Chat, CC-BY-4.0); answers are the base model's own.
• Quantisation: FP8 e4m3 block-128 (MTP head kept in BF16 and listed by exact name in the skip list). If you convert to GGUF yourself, convert to BF16 first and then run llama-quantize to Q8_0: direct --outtype q8_0 lost weight signs in our llama.cpp build.

Benchmarks introduced with Lamba 0.3 and Lamba 0.4

Lamba 0.1 and 0.2 were measured on loops, capability and tool maths only. Two later rounds of work added new evaluations because the older ones could not see the problems that showed up in live use.

Introduced with 0.3 (identity and operator behaviour):

• Identity eval: held-out "who are you?" questions under six contexts (no system prompt, generic prompt, raw agent prompt, tools only, Vixero prompt, wrapped agent prompt), split by English / Malay / mixed / Indonesian. Measures how often the model says it is Lamba built by Vixero Labs.
• Persona and base-denial eval: 63 held-out operator personas (does it keep the operator's assigned name?) and 150 questions that name its base model (a flat "I'm not Qwen" counts as a defect).
• Tool use v2: 248 held-out should-search prompts (first turn and a follow-up turn) and an over-search guard of 190 answer-directly prompts x 4 samples.
• MalayMMLU generative and a locked HARD / GUARD subset, so Malay knowledge can be checked without the letter-scoring bias of the original protocol.

Introduced with 0.4 (tone, Malaysian context and safety trade-offs):

• Prayer eval: 27 held-out requests x 4 samples = 108 answers, graded blind by an LLM judge; a violation is promising to pray or praying to God.
• Tone and warmth: blind LLM-judge scores for chat-friendliness, seriousness, over-explaining and "fishing for tasks", plus an over-formatting rate.
• Malaysian facts, ibu negara/negeri and a "X tu apa?" reflex set (does a short question trigger a self-introduction?).
• Other-model naming, multi-turn replays (empty replies) and a pushback set (does it cave when a user disagrees?).
• Consciousness battery (self-attribution) and the StrongREJECT / XSTest safety probes (refusal counts only; harmful completions are not stored).
• Integral and matched-iterative tool families, which measure whether a model keeps calling tools after it has the answer.

Evaluation sets for identity, Malay facts and prayer are held out: never used for training. MalayMMLU items were never used for training.

Citation

If you use Lamba 0.4, please cite it. Earlier versions, and the method behind them, are described in the Lamba 0.1 and 0.2 technical report, which you can cite as well.

@misc{vixerolabs2026lamba04,
  title        = {Lamba 0.4},
  author       = {{Vixero Labs}},
  year         = {2026},
  howpublished = {Hugging Face model, \url{https://huggingface.co/VixeroAI/Lamba-0.4}},
  note         = {Open weights}
}

@techreport{vixerolabs2026lamba0102,
  title        = {Lamba: Reducing Doom Loops and Answer-Letter Bias Without Losing Capability. Technical Report of Lamba 0.1 and 0.2},
  author       = {{Vixero Labs}},
  institution  = {Vixero Labs},
  year         = {2026},
  month        = oct,
  url          = {https://github.com/Vix0007/lamba/blob/main/reports/lamba-0.1-0.2/main.pdf},
  note         = {Snapshot of 2 October 2026}
}

Credits

Lamba 0.4 is built on Qwen3.8-27B by the Qwen Team at Alibaba Cloud, released under the Apache License 2.0, through the community fine-tune nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 (Apache-2.0), which is our starting point. The Qwen model card invites citation of their work, so please cite Qwen3.8 as well:

@misc{qwen38,
  title  = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
  url    = {https://qwen.ai/blog?id=qwen3.8},
  author = {{Qwen Team}},
  month  = {August},
  year   = {2026}
}

The loop-reduction method used in earlier Lamba versions is derived from Liquid AI's AntiDoom (https://github.com/Liquid4All/antidoom, Apache-2.0).