Jev News jevainews.com
TypeSafe Latest on X

Published

Updated

Span-02's card prints Decision Index 0.3 text index 63.62, and the blog prints thinking off at 61.2 against Jev 58.0

This topic was created 14 days ago, and the information it contains may have evolved or changed since then.

evals open-source

Hendrix Liu posted October 16, 2026, 23:16 UTC, that Span-02 had more usage in under 24 hours than Span-01 did in 20 days. The card still prints Decision Index 0.3 text index 63.62. The blog prints thinking off at 61.2 against Jev 1.13.0 at 58.0. This desk did not run the weights.

Respan posted Span-02 on October 15, 2026, at 17:00 UTC. The first tweet names a decision model that takes a deadline, returns its best answer so far, adds a number type, reads images, and ships open weights. The attachment is a video. The still on this page is the 7 + 7 frame from the next tweet in that thread.

The blog, opened October 16, 2026 UTC, prints Decision Index 0.3, public, thinking off at 61.2 against Jev 1.13.0 at 58.0. With no thinking limit the same rebuild reaches 64.1. The figure lists 1 s at 63.0 and 3 s at 63.6. Named neighbors on that plot are Perplexity Decider 62.2, Cloudflare clef 61.7, and LiquidAI d1-3B 49.0. The caption says Span-02 ran a local rebuild of the public suite, 37 benchmarks, chance-corrected, and that the 1 s and longer points add a change measured on a 5,000-question sample. Board models keep the board’s clocks. Jev’s clock is its API round trip. Span-02’s clock is server time.

The Hugging Face card, opened the same UTC day, prints Decision Index 0.3 public text index 63.62 on 140,178 requests, all answered, and vision 70.69 on 12,210 requests, all answered. 63.62 sits next to the blog’s 3 s cell of 63.6. They are two prints. Do not treat them as one run.

Blind Earth on the blog is 4,050 points on a 4 degree grid. Brightness is P(land), scored against Natural Earth. Thinking off 71.9%, 100 ms 72.5%, 300 ms 84.5%, 1 s 84.7%. Jev 1.13.0 on that question is 79.6%. The caption points at Celeste’s Claude chart and says that chart uses a 1 km land mask weighted by area, so those percents do not compare with these.

A Wikipedia speed-run on October 15, Alexander the Great to University of Milan, clicking only links in the article, prints Span-02 at 20.6 s and GPT-6 Luna at 38.5 s. Luna’s goal included the hint Italy. That pair is one path. It is not a latency ranking against TypeSafe desk clocks.

The hosted sample is POST https://api.respan.ai/api/v1/decisions, model span-02, with reasoning.mode on and budget_ms 300. Opened October 16, GET https://api.respan.ai/api/models/public/?unnest=true returned HTTP 200 with 2,474 rows. span-02 and respan/span-02 both print input_cost 0.01 and output_cost 0.0. We did not send a POST.

OpenRouter’s HTML page, opened the same UTC day, prints $0.01 / $0 per 1M, context 131K, Released Oct 16, 2026, and provider P50 1.91 s. GET /api/v1/models omitted that id. GET of respan/span-02 returned HTTP 404. This desk did not call OpenRouter. 1.91 s is not the card’s 36 ms median on an RTX PRO 6000.

The card says the checkpoint is a fine-tune of Gemma 4 26B-A4B-it, about 4B active parameters, quantized NVFP4/FP8 at 16.4 GB. The local path is POST /v1/systemone on the included Rust and CUDA engine. The weights are a Gemma 4 derivative under the Gemma 4 license, which that page labels Apache 2.0. The engine is Apache-2.0. The card says those weights only run with engine/ and cannot be loaded with transformers or vLLM. The listed GPUs are NVIDIA Blackwell sm_120a, RTX PRO 6000 or RTX 5090. We did not run them.

Harvey Li, October 15, 20:21 UTC, wrote that the stack is up to 7 times faster than vLLM on a single 5090. The card’s throughput table prints 1.3 to 7 times on that GPU without thinking, 127 requests a second with thinking off and 102 with auto thinking.

d3’s Decision Index 0.3.1 cell of 64.7, Blink’s 57.48 on edition 0.2.1, live space Jev at 60.1, and Artificial Analysis d1-ultra at 65 stay on their own pages. Access does not list api.respan.ai as a door for calling Jev.

Hendrix Liu, October 16, 2026, 23:16 UTC, wrote that Span-02 launched less than 24 hours ago and had already seen more usage than Span-01 did in 20 days. The same post prints 7 times faster than vLLM, 4 times cheaper than Jev, #4 on the Decision Index, and up to 4 images per request. That post does not print a request count.

Respan, 16:01 UTC, restated $0.01 per million input tokens. The next tweet in that thread says Jev could not win a single chess game against Span-02, 0-12, with Span-02’s reasoning turned off. We did not play those games.

Harvey Li, 22:08 UTC, wrote that Span-02 is +7.8, 2 times faster and 2 times cheaper than pplx-decider-flash-v1.2 on Decision Index. +7.8 is his comparison. It is not the blog’s 61.2 against Jev 58.0.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

Primary post is @RespanAI, October 15, 2026, 17:00 UTC, status 2110777826240819205. The text names Span-02, a deadline in milliseconds, thinking off beating Jev at one quarter the price and one third the latency of decision models at its level, a number type, images, and open weights. The attachment is a video of 11,133 ms. The share image is the 7 + 7 still from the next tweet in that thread, status 2110777871837151668.

Blog https://www.respan.ai/blog/introducing-span-2, dated October 15, 2026, opened October 16. Thinking off already beats Jev on Decision Index 0.3, 61.2 to 58.0. With no limit it reaches 64.1. The figure alt lists Span-02 off 61.2, 1 s 63.0, 3 s 63.6, no limit 64.1, Jev 1.13.0 58.0, Perplexity Decider 62.2, Cloudflare clef 61.7, LiquidAI d1-3B 49.0. The caption says Span-02 ran a local rebuild of the public suite, 37 benchmarks, chance-corrected, and that 1 s and longer are thinking off plus the change on a 5,000-question sample. Latency is board clocks for the named models, Jev its API round trip, Span-02 server time.

Blind Earth, 4,050 points on a 4 degree grid, P(land) against Natural Earth. Thinking off 71.9%, 100 ms 72.5%, 300 ms 84.5%, 1 s 84.7%. Jev 1.13.0 on that question is 79.6%. The caption says Celeste's Claude chart uses a 1 km land mask weighted by area, so those percents do not compare.

Wikipedia speed-run, Alexander the Great to University of Milan, clicking only links, both runs October 15, 2026. Span-02 20.6 s, GPT-6 Luna 38.5 s. Luna's goal included the hint Italy.

Hosted sample: POST https://api.respan.ai/api/v1/decisions, model span-02, reasoning mode on, budget_ms 300. GET https://api.respan.ai/api/models/public/?unnest=true, opened October 16, returned HTTP 200 with 2,474 rows. span-02 and respan/span-02 both print input_cost 0.01, output_cost 0.0, status active, rate_limit 300000 with no unit on that payload. We did not send a POST.

Hugging Face card https://huggingface.co/respanai/span-02, opened October 16. Decision Index 0.3 public, text index 63.62 on 140,178 requests, all answered. Vision 70.69 on 12,210 requests, all answered. Fine-tune of Gemma 4 26B-A4B, about 4B active, NVFP4/FP8 16.4 GB. POST /v1/systemone on the included engine. Weights are a derivative of google/gemma-4-26B-A4B-it under the Gemma 4 license (Apache 2.0). The engine is Apache-2.0. The card says the weights only run with engine/ and cannot be loaded with transformers or vLLM. Hardware is NVIDIA Blackwell sm_120a, RTX PRO 6000 or RTX 5090. We did not run the weights.

Harvey Li, October 15, 20:21 UTC, status 2110828352567476670. The custom stack, Rust, NVFP4, FA2, is up to 7 times faster than vLLM on a single 5090. The card's throughput table prints that 1.3 to 7 times range on the same GPU without thinking.

OpenRouter HTML https://openrouter.ai/respan/span-02, opened October 16 in this desk's Chrome. $0.01 / $0 per 1M, context 131K, Released Oct 16, 2026. Provider P50 1.91 s. GET https://openrouter.ai/api/v1/models returned HTTP 200 with 465 ids and no span-02 id. GET respan/span-02 and respanai/span-02 returned HTTP 404. This desk did not call OpenRouter. Do not merge 1.91 s with the card's 36 ms median.

Hendrix Liu, @HendrixLiu29208, October 16, 2026, 23:16 UTC, status 2111234792650092681. The text says Span-02 launched less than 24 hours ago and has already seen more usage than Span-01 did in 20 days, 7 times faster than vLLM, 4 times cheaper than Jev, #4 on the Decision Index, up to 4 images per request. Respan, 16:01 UTC, status 2111125313782722573, restates $0.01 per million input tokens. Status 2111125343532937282 says Jev could not win a single chess game against Span-02, 0-12, with Span-02's reasoning turned off. Harvey Li, 22:08 UTC, status 2111217866888642986, wrote that Span-02 is +7.8, 2x faster and 2x cheaper than pplx-decider-flash-v1.2 on DI. We did not rerun chess, and we did not count Span-01 usage.

Compare

61.2, 63.0, 63.6, and 64.1 are Respan's local rebuild of Decision Index 0.3, public, 37 benchmarks. 63.62 is the Hugging Face card's text index on 140,178 requests. Do not merge those two prints. Jev 1.13.0 at 58.0 is the blog's cell on that rebuild. Edition 0.3's live space Jev reference of 60.1, d3's 64.7, Blink's 57.48 on 0.2.1, and Artificial Analysis d1-ultra at 65 are other boards.

Land/water 71.9% to 84.7% is 4,050 points. Jev 1.13.0 at 79.6% is that question. Celeste's Claude percents stay on a different mask.

20.6 s against 38.5 s is one Wikipedia path on October 15. It is not a latency ranking against TypeSafe desk clocks.

0-12 chess is Respan's October 16 clip, reasoning off. Harvey's +7.8 is Span-02 against pplx-decider-flash-v1.2 on Decision Index. Hendrix's usage line has no request count on the post. Do not merge 61.2 with +7.8 or with 0-12.

Terms

Span-02
Respan's decision checkpoint, posted October 15, 2026. Fine-tune of Gemma 4 26B-A4B. The card prints Decision Index 0.3 text index 63.62. The blog prints thinking off at 61.2 against Jev 1.13.0 at 58.0.
budget_ms
Span-02's thinking deadline in milliseconds. auto thinks when the top answer is below 0.8. on always thinks. off never thinks. The blog's no-limit Decision Index cell is 64.1.
number type
A Span-02 question that returns a numeric value and a confidence, without listing candidate numbers. The card marks number, reasoning, and images as extensions of TypeSafe's System One API.

Sources

  1. Respan, October 15
  2. Respan, anytime reasoning
  3. Respan, numbers
  4. Respan, available now
  5. Respan, Hugging Face
  6. Harvey Li, engine
  7. Introducing Span-02
  8. Span-02 quickstart
  9. respanai/span-02
  10. OpenRouter, Span-02
  11. Decision Index space
  12. Gemma 4 26B-A4B-it
  13. Hendrix Liu, usage in under 24 hours
  14. Respan, cheapest, October 16
  15. Respan, chess 0-12
  16. Harvey Li, +7.8 against flash v1.2